summerAIO commited on
Commit
6848dc2
·
verified ·
1 Parent(s): 55bb8b3

Initial private upload of Sun*+ (provisional eval 6/31 chunks; models final, results to be overwritten)

Browse files

Package from /workspace/sunstar-plus/hf_release. Models, inference code, and station metadata are complete. README and COMPARISON.md are clearly marked as provisional (test eval 6/31 chunks: 2026-01–02 and early June). Final nationwide metrics will overwrite the docs after eval finishes.

Files changed (50) hide show
  1. COMPARISON.md +290 -0
  2. README.md +183 -0
  3. data/splits.json +308 -0
  4. data/stations_nationwide.csv +0 -0
  5. infer_example.py +31 -0
  6. models/config.json +135 -0
  7. models/features_A.json +1 -0
  8. models/features_B.json +1 -0
  9. models/p1_A_humidity_pct_s0.txt.gz +3 -0
  10. models/p1_A_humidity_pct_s1.txt.gz +3 -0
  11. models/p1_A_precip_mm_s0.txt.gz +3 -0
  12. models/p1_A_precip_mm_s1.txt.gz +3 -0
  13. models/p1_A_temp_c_s0.txt.gz +3 -0
  14. models/p1_A_temp_c_s1.txt.gz +3 -0
  15. models/p1_A_wind_speed_ms_s0.txt.gz +3 -0
  16. models/p1_A_wind_speed_ms_s1.txt.gz +3 -0
  17. models/p1_A_wind_u_ms_s0.txt.gz +3 -0
  18. models/p1_A_wind_u_ms_s1.txt.gz +3 -0
  19. models/p1_A_wind_v_ms_s0.txt.gz +3 -0
  20. models/p1_A_wind_v_ms_s1.txt.gz +3 -0
  21. models/p1_B_humidity_pct_s0.txt.gz +3 -0
  22. models/p1_B_humidity_pct_s1.txt.gz +3 -0
  23. models/p1_B_precip_mm_s0.txt.gz +3 -0
  24. models/p1_B_precip_mm_s1.txt.gz +3 -0
  25. models/p1_B_temp_c_s0.txt.gz +3 -0
  26. models/p1_B_temp_c_s1.txt.gz +3 -0
  27. models/p1_B_wind_speed_ms_s0.txt.gz +3 -0
  28. models/p1_B_wind_speed_ms_s1.txt.gz +3 -0
  29. models/p1_B_wind_u_ms_s0.txt.gz +3 -0
  30. models/p1_B_wind_u_ms_s1.txt.gz +3 -0
  31. models/p1_B_wind_v_ms_s0.txt.gz +3 -0
  32. models/p1_B_wind_v_ms_s1.txt.gz +3 -0
  33. models/sigma_A_humidity_pct.txt.gz +3 -0
  34. models/sigma_A_temp_c.txt.gz +3 -0
  35. models/sigma_A_wind_u_ms.txt.gz +3 -0
  36. models/sigma_A_wind_v_ms.txt.gz +3 -0
  37. models/sigma_B_humidity_pct.txt.gz +3 -0
  38. models/sigma_B_temp_c.txt.gz +3 -0
  39. models/sigma_B_wind_u_ms.txt.gz +3 -0
  40. models/sigma_B_wind_v_ms.txt.gz +3 -0
  41. requirements.txt +9 -0
  42. results/metrics_nationwide_and_okayama.csv +0 -0
  43. results/v2_official_okayama.csv +73 -0
  44. selftest.py +23 -0
  45. sunstar_plus/__init__.py +6 -0
  46. sunstar_plus/era5io.py +43 -0
  47. sunstar_plus/features.py +161 -0
  48. sunstar_plus/model.py +90 -0
  49. sunstar_plus/obsfeats.py +54 -0
  50. sunstar_plus/terrain.py +69 -0
COMPARISON.md ADDED
@@ -0,0 +1,290 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ > **暫定 (評価チャンク 6/31 のみ: 2026-01〜02 と 06 上旬)。最終結果は評価完了後に自動で上書きされます。**
2
+
3
+ # Sun*+ vs Sun* V2 — 評価結果 (テスト期間 2026-01-01〜2026-10-31 UTC, アメダス実測)
4
+
5
+ RMSE は小さいほど良い。改善率 = 1 − Sun*+/V2。V2 は **公開済み V2 モデルを同一の特徴行列(全国版パイプライン)で全国地点に適用**したもの (V2 は岡山の 22 地点でのみ学習)。ガイダンスは ERA5 *再解析* (完全予報)。
6
+
7
+ ## 1. 全国評価 (主)
8
+
9
+ 評価地点: 学習に使った地点の層化サンプル 160 + **学習・チューニングに一切使っていない全国 held-out 113 地点** (岡山県を除く全都道府県/地方から層化抽出)。
10
+
11
+ ### 1-1. 全国 全評価地点 (273)
12
+
13
+ #### リード0 (観測なしダウンスケーリング)
14
+
15
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
16
+ |---|---|---|---|---|---|---|---|---|
17
+ | 気温 [°C] | 0 | 268693 | 2.074 | 1.922 | **1.293** | +32.7% | 1.085 | 0.667 |
18
+ | 風速 [m/s] | 0 | 265916 | 2.242 | 1.721 | **1.260** | +26.7% | — | — |
19
+ | 風 u [m/s] | 0 | 265916 | 2.347 | 1.743 | **1.396** | +19.9% | 0.965 | 0.726 |
20
+ | 風 v [m/s] | 0 | 265916 | 2.115 | 1.772 | **1.379** | +22.2% | 0.993 | 0.719 |
21
+ | 相対湿度 [%] | 0 | 232549 | 12.144 | 12.398 | **9.433** | +23.9% | 7.332 | 5.110 |
22
+ | 降水量 [mm/h] | 0 | 391954 | 0.548 | 0.548 | **0.512** | +6.5% | — | — |
23
+
24
+ #### リード別 (観測同化; 初期時刻までの AMeDAS 観測 + 近傍局)
25
+
26
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
27
+ |---|---|---|---|---|---|---|---|---|
28
+ | 気温 [°C] | 1 | 268639 | 2.074 | 1.094 | **0.771** | +29.6% | 0.583 | 0.398 |
29
+ | 風速 [m/s] | 1 | 263238 | 2.247 | 1.296 | **0.939** | +27.6% | — | — |
30
+ | 風 u [m/s] | 1 | 263238 | 2.353 | 1.439 | **1.097** | +23.8% | 0.771 | 0.569 |
31
+ | 風 v [m/s] | 1 | 263238 | 2.120 | 1.448 | **1.104** | +23.7% | 0.784 | 0.572 |
32
+ | 相対湿度 [%] | 1 | 232504 | 12.145 | 7.245 | **5.587** | +22.9% | 4.013 | 2.869 |
33
+ | 降水量 [mm/h] | 1 | 391892 | 0.548 | 0.518 | **0.456** | +12.0% | — | — |
34
+ | 気温 [°C] | 3 | 268591 | 2.074 | 1.364 | **1.031** | +24.4% | 0.743 | 0.537 |
35
+ | 風速 [m/s] | 3 | 263022 | 2.248 | 1.443 | **1.091** | +24.4% | — | — |
36
+ | 風 u [m/s] | 3 | 263022 | 2.353 | 1.575 | **1.265** | +19.7% | 0.854 | 0.657 |
37
+ | 風 v [m/s] | 3 | 263022 | 2.119 | 1.612 | **1.265** | +21.5% | 0.883 | 0.660 |
38
+ | 相対湿度 [%] | 3 | 232461 | 12.145 | 9.095 | **7.410** | +18.5% | 5.113 | 3.893 |
39
+ | 降水量 [mm/h] | 3 | 391833 | 0.548 | 0.536 | **0.507** | +5.5% | — | — |
40
+ | 気温 [°C] | 6 | 268577 | 2.074 | 1.534 | **1.169** | +23.8% | 0.844 | 0.610 |
41
+ | 風速 [m/s] | 6 | 262861 | 2.247 | 1.512 | **1.154** | +23.7% | — | — |
42
+ | 風 u [m/s] | 6 | 262861 | 2.351 | 1.629 | **1.331** | +18.3% | 0.889 | 0.695 |
43
+ | 風 v [m/s] | 6 | 262861 | 2.119 | 1.681 | **1.330** | +20.9% | 0.925 | 0.696 |
44
+ | 相対湿度 [%] | 6 | 232447 | 12.144 | 10.246 | **8.257** | +19.4% | 5.847 | 4.410 |
45
+ | 降水量 [mm/h] | 6 | 391790 | 0.548 | 0.540 | **0.516** | +4.5% | — | — |
46
+ | 気温 [°C] | 12 | 268563 | 2.074 | 1.630 | **1.237** | +24.1% | 0.900 | 0.644 |
47
+ | 風速 [m/s] | 12 | 262739 | 2.246 | 1.543 | **1.186** | +23.2% | — | — |
48
+ | 風 u [m/s] | 12 | 262739 | 2.349 | 1.655 | **1.360** | +17.8% | 0.904 | 0.712 |
49
+ | 風 v [m/s] | 12 | 262739 | 2.118 | 1.710 | **1.360** | +20.4% | 0.942 | 0.713 |
50
+ | 相対湿度 [%] | 12 | 232429 | 12.145 | 10.738 | **8.679** | +19.2% | 6.172 | 4.668 |
51
+ | 降水量 [mm/h] | 12 | 391724 | 0.548 | 0.541 | **0.515** | +4.9% | — | — |
52
+ | 気温 [°C] | 24 | 268551 | 2.074 | 1.618 | **1.247** | +23.0% | 0.894 | 0.649 |
53
+ | 風速 [m/s] | 24 | 262850 | 2.244 | 1.555 | **1.198** | +23.0% | — | — |
54
+ | 風 u [m/s] | 24 | 262850 | 2.350 | 1.662 | **1.372** | +17.4% | 0.908 | 0.717 |
55
+ | 風 v [m/s] | 24 | 262850 | 2.118 | 1.711 | **1.371** | +19.8% | 0.941 | 0.718 |
56
+ | 相対湿度 [%] | 24 | 232404 | 12.146 | 10.762 | **8.821** | +18.0% | 6.187 | 4.746 |
57
+ | 降水量 [mm/h] | 24 | 391647 | 0.548 | 0.541 | **0.514** | +5.1% | — | — |
58
+
59
+ ### 1-2. 全国 held-out 地点のみ (113) — 空間汎化
60
+
61
+ #### リード0 (観測なしダウンスケーリング)
62
+
63
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
64
+ |---|---|---|---|---|---|---|---|---|
65
+ | 気温 [°C] | 0 | 110841 | 2.065 | 1.960 | **1.480** | +24.5% | 1.069 | 0.768 |
66
+ | 風速 [m/s] | 0 | 109558 | 2.205 | 1.630 | **1.459** | +10.5% | — | — |
67
+ | 風 u [m/s] | 0 | 109558 | 2.378 | 1.704 | **1.548** | +9.2% | 0.909 | 0.813 |
68
+ | 風 v [m/s] | 0 | 109558 | 2.025 | 1.734 | **1.557** | +10.2% | 0.939 | 0.817 |
69
+ | 相対湿度 [%] | 0 | 99157 | 12.061 | 12.754 | **9.995** | +21.6% | 7.482 | 5.504 |
70
+ | 降水量 [mm/h] | 0 | 162587 | 0.528 | 0.520 | **0.487** | +6.3% | — | — |
71
+
72
+ #### リード別 (観測同化; 初期時刻までの AMeDAS 観測 + 近傍局)
73
+
74
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
75
+ |---|---|---|---|---|---|---|---|---|
76
+ | 気温 [°C] | 1 | 110819 | 2.065 | 1.079 | **0.817** | +24.3% | 0.566 | 0.423 |
77
+ | 風速 [m/s] | 1 | 108435 | 2.211 | 1.237 | **0.957** | +22.7% | — | — |
78
+ | 風 u [m/s] | 1 | 108435 | 2.384 | 1.400 | **1.104** | +21.1% | 0.736 | 0.573 |
79
+ | 風 v [m/s] | 1 | 108435 | 2.029 | 1.415 | **1.134** | +19.9% | 0.749 | 0.588 |
80
+ | 相対湿度 [%] | 1 | 99136 | 12.062 | 7.218 | **5.454** | +24.4% | 3.984 | 2.823 |
81
+ | 降水量 [mm/h] | 1 | 162566 | 0.528 | 0.491 | **0.439** | +10.7% | — | — |
82
+ | 気温 [°C] | 3 | 110803 | 2.065 | 1.354 | **1.092** | +19.3% | 0.724 | 0.571 |
83
+ | 風速 [m/s] | 3 | 108362 | 2.211 | 1.375 | **1.124** | +18.3% | — | — |
84
+ | 風 u [m/s] | 3 | 108362 | 2.385 | 1.528 | **1.285** | +15.9% | 0.810 | 0.669 |
85
+ | 風 v [m/s] | 3 | 108362 | 2.029 | 1.578 | **1.320** | +16.3% | 0.841 | 0.689 |
86
+ | 相対湿度 [%] | 3 | 99118 | 12.061 | 9.070 | **7.338** | +19.1% | 5.081 | 3.879 |
87
+ | 降水量 [mm/h] | 3 | 162548 | 0.528 | 0.509 | **0.484** | +5.0% | — | — |
88
+ | 気温 [°C] | 6 | 110802 | 2.065 | 1.532 | **1.251** | +18.4% | 0.824 | 0.654 |
89
+ | 風速 [m/s] | 6 | 108313 | 2.211 | 1.436 | **1.200** | +16.5% | — | — |
90
+ | 風 u [m/s] | 6 | 108313 | 2.383 | 1.576 | **1.362** | +13.6% | 0.840 | 0.713 |
91
+ | 風 v [m/s] | 6 | 108313 | 2.028 | 1.647 | **1.402** | +14.9% | 0.881 | 0.734 |
92
+ | 相対湿度 [%] | 6 | 99114 | 12.061 | 10.294 | **8.268** | +19.7% | 5.834 | 4.446 |
93
+ | 降水量 [mm/h] | 6 | 162537 | 0.528 | 0.513 | **0.490** | +4.4% | — | — |
94
+ | 気温 [°C] | 12 | 110802 | 2.065 | 1.631 | **1.329** | +18.5% | 0.878 | 0.692 |
95
+ | 風速 [m/s] | 12 | 108283 | 2.210 | 1.463 | **1.239** | +15.4% | — | — |
96
+ | 風 u [m/s] | 12 | 108283 | 2.382 | 1.601 | **1.398** | +12.7% | 0.853 | 0.732 |
97
+ | 風 v [m/s] | 12 | 108283 | 2.027 | 1.675 | **1.444** | +13.8% | 0.897 | 0.756 |
98
+ | 相対湿度 [%] | 12 | 99108 | 12.061 | 10.824 | **8.714** | +19.5% | 6.177 | 4.717 |
99
+ | 降水量 [mm/h] | 12 | 162525 | 0.528 | 0.514 | **0.489** | +4.9% | — | — |
100
+ | 気温 [°C] | 24 | 110805 | 2.065 | 1.618 | **1.335** | +17.5% | 0.877 | 0.696 |
101
+ | 風速 [m/s] | 24 | 108317 | 2.209 | 1.472 | **1.251** | +15.0% | — | — |
102
+ | 風 u [m/s] | 24 | 108317 | 2.383 | 1.606 | **1.406** | +12.5% | 0.856 | 0.736 |
103
+ | 風 v [m/s] | 24 | 108317 | 2.028 | 1.675 | **1.454** | +13.2% | 0.894 | 0.759 |
104
+ | 相対湿度 [%] | 24 | 99099 | 12.061 | 10.846 | **8.845** | +18.4% | 6.199 | 4.781 |
105
+ | 降水量 [mm/h] | 24 | 162513 | 0.528 | 0.513 | **0.490** | +4.6% | — | — |
106
+
107
+ ### 1-3. 学習地点サンプル
108
+
109
+ #### リード0 (観測なしダウンスケーリング)
110
+
111
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
112
+ |---|---|---|---|---|---|---|---|---|
113
+ | 気温 [°C] | 0 | 157852 | 2.081 | 1.895 | **1.144** | +39.6% | 1.097 | 0.597 |
114
+ | 風速 [m/s] | 0 | 156358 | 2.268 | 1.782 | **1.100** | +38.2% | — | — |
115
+ | 風 u [m/s] | 0 | 156358 | 2.325 | 1.769 | **1.279** | +27.7% | 1.004 | 0.666 |
116
+ | 風 v [m/s] | 0 | 156358 | 2.176 | 1.799 | **1.239** | +31.1% | 1.031 | 0.650 |
117
+ | 相対湿度 [%] | 0 | 133392 | 12.206 | 12.127 | **8.993** | +25.8% | 7.221 | 4.818 |
118
+ | 降水量 [mm/h] | 0 | 229367 | 0.562 | 0.567 | **0.529** | +6.6% | — | — |
119
+
120
+ #### リード別 (観測同化; 初期時刻までの AMeDAS 観測 + 近傍局)
121
+
122
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
123
+ |---|---|---|---|---|---|---|---|---|
124
+ | 気温 [°C] | 1 | 157820 | 2.081 | 1.105 | **0.737** | +33.3% | 0.595 | 0.381 |
125
+ | 風速 [m/s] | 1 | 154803 | 2.273 | 1.336 | **0.926** | +30.7% | — | — |
126
+ | 風 u [m/s] | 1 | 154803 | 2.331 | 1.466 | **1.091** | +25.6% | 0.796 | 0.565 |
127
+ | 風 v [m/s] | 1 | 154803 | 2.181 | 1.470 | **1.083** | +26.3% | 0.809 | 0.561 |
128
+ | 相対湿度 [%] | 1 | 133368 | 12.207 | 7.265 | **5.684** | +21.8% | 4.034 | 2.903 |
129
+ | 降水量 [mm/h] | 1 | 229326 | 0.562 | 0.537 | **0.468** | +12.8% | — | — |
130
+ | 気温 [°C] | 3 | 157788 | 2.081 | 1.371 | **0.985** | +28.1% | 0.756 | 0.513 |
131
+ | 風速 [m/s] | 3 | 154660 | 2.273 | 1.489 | **1.068** | +28.3% | — | — |
132
+ | 風 u [m/s] | 3 | 154660 | 2.330 | 1.608 | **1.250** | +22.2% | 0.885 | 0.649 |
133
+ | 風 v [m/s] | 3 | 154660 | 2.181 | 1.635 | **1.225** | +25.1% | 0.912 | 0.640 |
134
+ | 相対湿度 [%] | 3 | 133343 | 12.207 | 9.113 | **7.464** | +18.1% | 5.136 | 3.903 |
135
+ | 降水量 [mm/h] | 3 | 229285 | 0.562 | 0.554 | **0.522** | +5.7% | — | — |
136
+ | 気温 [°C] | 6 | 157775 | 2.081 | 1.535 | **1.107** | +27.9% | 0.859 | 0.579 |
137
+ | 風速 [m/s] | 6 | 154548 | 2.272 | 1.563 | **1.121** | +28.2% | — | — |
138
+ | 風 u [m/s] | 6 | 154548 | 2.328 | 1.665 | **1.309** | +21.4% | 0.923 | 0.682 |
139
+ | 風 v [m/s] | 6 | 154548 | 2.180 | 1.705 | **1.276** | +25.2% | 0.956 | 0.669 |
140
+ | 相対湿度 [%] | 6 | 133333 | 12.206 | 10.210 | **8.250** | +19.2% | 5.857 | 4.384 |
141
+ | 降水量 [mm/h] | 6 | 229253 | 0.562 | 0.559 | **0.533** | +4.6% | — | — |
142
+ | 気温 [°C] | 12 | 157761 | 2.081 | 1.629 | **1.167** | +28.4% | 0.915 | 0.610 |
143
+ | 風速 [m/s] | 12 | 154456 | 2.271 | 1.597 | **1.147** | +28.2% | — | — |
144
+ | 風 u [m/s] | 12 | 154456 | 2.326 | 1.692 | **1.333** | +21.2% | 0.941 | 0.697 |
145
+ | 風 v [m/s] | 12 | 154456 | 2.179 | 1.734 | **1.299** | +25.1% | 0.974 | 0.683 |
146
+ | 相対湿度 [%] | 12 | 133321 | 12.207 | 10.673 | **8.653** | +18.9% | 6.168 | 4.632 |
147
+ | 降水量 [mm/h] | 12 | 229199 | 0.562 | 0.560 | **0.532** | +4.8% | — | — |
148
+ | 気温 [°C] | 24 | 157746 | 2.081 | 1.619 | **1.180** | +27.1% | 0.907 | 0.617 |
149
+ | 風速 [m/s] | 24 | 154533 | 2.268 | 1.610 | **1.158** | +28.1% | — | — |
150
+ | 風 u [m/s] | 24 | 154533 | 2.326 | 1.699 | **1.347** | +20.7% | 0.945 | 0.704 |
151
+ | 風 v [m/s] | 24 | 154533 | 2.178 | 1.736 | **1.310** | +24.5% | 0.973 | 0.689 |
152
+ | 相対湿度 [%] | 24 | 133305 | 12.208 | 10.700 | **8.803** | +17.7% | 6.178 | 4.719 |
153
+ | 降水量 [mm/h] | 24 | 229134 | 0.562 | 0.560 | **0.530** | +5.4% | — | — |
154
+
155
+ ### 勝敗サマリ (RMSE, 変数×リード)
156
+
157
+ | 地点群 | Sun*+ が良い / 全指標 |
158
+ |---|---|
159
+ | nat_all | 36 / 36 |
160
+ | nat_holdout | 36 / 36 |
161
+ | nat_seen | 36 / 36 |
162
+ | okayama_seen | 36 / 36 |
163
+ | okayama_holdout | 29 / 36 |
164
+
165
+ ### 1-4. 地方別 (全国 全評価地点, 改善率 = 1 − Sun*+/V2 の RMSE)
166
+
167
+ #### リード 0h — RMSE (V2 → Sun*+) と 改善率
168
+
169
+ | 地方 | 地点数 | 気温 [°C] | 風速 [m/s] | 風 u [m/s] | 風 v [m/s] | 相対湿度 [%] | 降水量 [mm/h] |
170
+ |---|---|---|---|---|---|---|---|
171
+ | 北海道 | 52 | 2.10→**1.62** (+23.0%) | 1.80→**1.32** (+26.5%) | 1.76→**1.43** (+18.8%) | 1.90→**1.42** (+25.5%) | 12.15→**9.63** (+20.7%) | 0.32→**0.30** (+6.5%) |
172
+ | 東北 | 48 | 1.48→**1.09** (+26.2%) | 1.85→**1.38** (+25.4%) | 1.95→**1.52** (+22.1%) | 1.55→**1.27** (+17.7%) | 11.79→**9.52** (+19.2%) | 0.46→**0.43** (+5.5%) |
173
+ | 関東 | 29 | 1.79→**1.19** (+33.7%) | 1.45→**1.11** (+23.5%) | 1.56→**1.31** (+16.1%) | 1.51→**1.29** (+14.4%) | 15.42→**10.16** (+34.1%) | 0.47→**0.45** (+3.9%) |
174
+ | 甲信 | 19 | 3.07→**1.27** (+58.7%) | 2.19→**1.29** (+41.1%) | 1.80→**1.47** (+18.6%) | 2.23→**1.59** (+28.4%) | 15.46→**9.42** (+39.1%) | 0.58→**0.53** (+7.8%) |
175
+ | 北陸 | 8 | 1.73→**1.04** (+40.0%) | 1.71→**1.15** (+32.7%) | 1.76→**1.42** (+19.2%) | 1.78→**1.37** (+22.9%) | 11.11→**8.74** (+21.4%) | 0.65→**0.60** (+8.4%) |
176
+ | 東海 | 23 | 2.15→**1.27** (+40.9%) | 1.56→**1.29** (+17.1%) | 1.90→**1.42** (+25.0%) | 1.42→**1.27** (+10.4%) | 11.96→**9.38** (+21.6%) | 0.31→**0.29** (+8.2%) |
177
+ | 近畿 | 19 | 1.57→**1.02** (+34.8%) | 1.29→**1.33** (-3.6%) | 1.46→**1.29** (+11.9%) | 1.41→**1.24** (+11.6%) | 10.39→**8.84** (+14.9%) | 0.30→**0.29** (+5.6%) |
178
+ | 中国 | 24 | 1.65→**1.12** (+31.9%) | 1.53→**1.20** (+21.5%) | 1.77→**1.41** (+20.4%) | 1.76→**1.41** (+19.8%) | 10.87→**8.09** (+25.6%) | 0.44→**0.40** (+8.9%) |
179
+ | 四国 | 13 | 1.68→**1.49** (+11.4%) | 1.48→**1.19** (+19.9%) | 1.70→**1.38** (+18.9%) | 2.10→**1.80** (+14.4%) | 12.92→**9.86** (+23.7%) | 0.20→**0.20** (+4.0%) |
180
+ | 九州 | 31 | 1.86→**1.25** (+32.6%) | 1.61→**1.04** (+35.1%) | 1.58→**1.22** (+22.7%) | 1.66→**1.22** (+26.8%) | 11.21→**9.44** (+15.8%) | 0.90→**0.83** (+8.4%) |
181
+ | 沖縄 | 7 | 1.16→**0.81** (+30.4%) | 2.66→**1.40** (+47.2%) | 1.61→**1.34** (+16.8%) | 3.03→**1.62** (+46.5%) | 6.01→**4.57** (+23.9%) | 1.50→**1.45** (+3.5%) |
182
+
183
+ #### リード 6h — RMSE (V2 → Sun*+) と 改善率
184
+
185
+ | 地方 | 地点数 | 気温 [°C] | 風速 [m/s] | 風 u [m/s] | 風 v [m/s] | 相対湿度 [%] | 降水量 [mm/h] |
186
+ |---|---|---|---|---|---|---|---|
187
+ | 北海道 | 52 | 1.77→**1.43** (+19.6%) | 1.58→**1.22** (+23.0%) | 1.66→**1.38** (+16.4%) | 1.78→**1.41** (+20.7%) | 10.25→**8.22** (+19.7%) | 0.32→**0.30** (+3.9%) |
188
+ | 東北 | 48 | 1.23→**1.03** (+16.0%) | 1.61→**1.23** (+23.8%) | 1.81→**1.42** (+21.6%) | 1.49→**1.25** (+16.3%) | 9.68→**8.32** (+14.1%) | 0.45→**0.43** (+4.0%) |
189
+ | 関東 | 29 | 1.48→**1.14** (+23.4%) | 1.33→**1.07** (+19.4%) | 1.50→**1.29** (+13.4%) | 1.48→**1.24** (+16.1%) | 11.92→**8.78** (+26.4%) | 0.46→**0.45** (+3.5%) |
190
+ | 甲信 | 19 | 2.14→**1.20** (+43.9%) | 1.84→**1.21** (+34.3%) | 1.67→**1.32** (+21.0%) | 2.14→**1.58** (+25.9%) | 11.67→**8.82** (+24.4%) | 0.56→**0.53** (+6.3%) |
191
+ | 北陸 | 8 | 1.36→**0.96** (+29.2%) | 1.50→**1.17** (+21.6%) | 1.72→**1.41** (+17.7%) | 1.70→**1.40** (+18.1%) | 9.93→**8.33** (+16.1%) | 0.63→**0.59** (+5.8%) |
192
+ | 東海 | 23 | 1.59→**1.12** (+29.1%) | 1.37→**1.06** (+22.0%) | 1.70→**1.24** (+26.9%) | 1.37→**1.15** (+16.4%) | 10.03→**8.07** (+19.6%) | 0.30→**0.29** (+4.0%) |
193
+ | 近畿 | 19 | 1.27→**0.96** (+24.7%) | 1.22→**1.11** (+8.8%) | 1.37→**1.21** (+12.1%) | 1.36→**1.18** (+13.0%) | 8.81→**7.67** (+12.9%) | 0.30→**0.28** (+5.3%) |
194
+ | 中国 | 24 | 1.36→**1.06** (+21.9%) | 1.44→**1.19** (+17.0%) | 1.68→**1.40** (+16.3%) | 1.70→**1.40** (+17.4%) | 9.53→**7.75** (+18.6%) | 0.44→**0.41** (+6.2%) |
195
+ | 四国 | 13 | 1.42→**1.25** (+11.9%) | 1.34→**1.10** (+18.1%) | 1.60→**1.35** (+15.5%) | 1.96→**1.54** (+21.5%) | 10.97→**8.55** (+22.1%) | 0.20→**0.20** (+2.5%) |
196
+ | 九州 | 31 | 1.47→**1.11** (+24.6%) | 1.42→**1.02** (+28.4%) | 1.47→**1.18** (+19.6%) | 1.58→**1.20** (+24.2%) | 9.76→**8.13** (+16.7%) | 0.89→**0.84** (+5.8%) |
197
+ | 沖縄 | 7 | 1.02→**0.76** (+25.0%) | 2.09→**1.21** (+42.0%) | 1.46→**1.31** (+10.2%) | 2.68→**1.44** (+46.3%) | 5.43→**4.42** (+18.6%) | 1.49→**1.46** (+2.2%) |
198
+
199
+ ## 2. 岡山 V2 プロトコル (副: V2 自身の評価条件。V2 は **公開済みの V2 公式評価値** `runs/v2_*_test_metrics.csv`)
200
+
201
+ 同一期間・同一地点(学習22地点 / 未学習3地点: 新見・和気・玉野)・同一指標。
202
+
203
+ ### 2-1. 学習22地点
204
+
205
+ | 変数 | リード[h] | n | ERA5補間 | V2(公式) | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
206
+ |---|---|---|---|---|---|---|---|---|
207
+ | 気温 [°C] | 0 | 18711 | 1.484 | 0.845 | **0.921** | -9.0% | 0.456 | 0.496 |
208
+ | 風速 [m/s] | 0 | 18383 | 1.591 | 0.699 | **0.795** | -13.6% | — | — |
209
+ | 風 u [m/s] | 0 | 18383 | 1.648 | 0.873 | **0.920** | -5.3% | 0.458 | 0.483 |
210
+ | 風 v [m/s] | 0 | 18383 | 1.704 | 0.950 | **0.999** | -5.2% | 0.493 | 0.516 |
211
+ | 相対湿度 [%] | 0 | 17272 | 10.038 | 5.856 | **6.696** | -14.3% | 2.992 | 3.410 |
212
+ | 降水量 [mm/h] | 0 | 31673 | 0.790 | 0.720 | **0.274** | +61.9% | — | — |
213
+ | 気温 [°C] | 1 | 18707 | 1.484 | 0.614 | **0.624** | -1.6% | 0.323 | 0.330 |
214
+ | 風速 [m/s] | 1 | 18096 | 1.594 | 0.650 | **0.711** | -9.4% | — | — |
215
+ | 風 u [m/s] | 1 | 18096 | 1.652 | 0.822 | **0.840** | -2.1% | 0.431 | 0.443 |
216
+ | 風 v [m/s] | 1 | 18096 | 1.708 | 0.889 | **0.936** | -5.3% | 0.461 | 0.485 |
217
+ | 相対湿度 [%] | 1 | 17269 | 10.038 | 4.105 | **4.356** | -6.1% | 2.025 | 2.129 |
218
+ | 降水量 [mm/h] | 1 | 31670 | 0.790 | 0.696 | **0.248** | +64.4% | — | — |
219
+ | 気温 [°C] | 3 | 18703 | 1.484 | 0.750 | **0.805** | -7.4% | 0.398 | 0.430 |
220
+ | 風速 [m/s] | 3 | 18076 | 1.594 | 0.692 | **0.780** | -12.8% | — | — |
221
+ | 風 u [m/s] | 3 | 18076 | 1.652 | 0.868 | **0.907** | -4.5% | 0.455 | 0.479 |
222
+ | 風 v [m/s] | 3 | 18076 | 1.707 | 0.948 | **1.001** | -5.6% | 0.492 | 0.519 |
223
+ | 相対湿度 [%] | 3 | 17265 | 10.037 | 5.124 | **5.791** | -13.0% | 2.540 | 2.850 |
224
+ | 降水量 [mm/h] | 3 | 31667 | 0.790 | 0.720 | **0.277** | +61.6% | — | — |
225
+ | 気温 [°C] | 6 | 18702 | 1.484 | 0.813 | **0.898** | -10.5% | 0.436 | 0.484 |
226
+ | 風速 [m/s] | 6 | 18075 | 1.593 | 0.704 | **0.803** | -14.0% | — | — |
227
+ | 風 u [m/s] | 6 | 18075 | 1.649 | 0.880 | **0.935** | -6.2% | 0.462 | 0.493 |
228
+ | 風 v [m/s] | 6 | 18075 | 1.707 | 0.962 | **1.016** | -5.7% | 0.499 | 0.528 |
229
+ | 相対湿度 [%] | 6 | 17264 | 10.038 | 5.543 | **6.395** | -15.4% | 2.787 | 3.196 |
230
+ | 降水量 [mm/h] | 6 | 31666 | 0.790 | 0.721 | **0.276** | +61.7% | — | — |
231
+ | 気温 [°C] | 12 | 18702 | 1.484 | 0.845 | **0.956** | -13.2% | 0.455 | 0.516 |
232
+ | 風速 [m/s] | 12 | 18068 | 1.591 | 0.708 | **0.813** | -14.8% | — | — |
233
+ | 風 u [m/s] | 12 | 18068 | 1.647 | 0.885 | **0.947** | -7.0% | 0.465 | 0.500 |
234
+ | 風 v [m/s] | 12 | 18068 | 1.705 | 0.964 | **1.021** | -5.9% | 0.501 | 0.532 |
235
+ | 相対湿度 [%] | 12 | 17264 | 10.040 | 5.732 | **6.648** | -16.0% | 2.916 | 3.374 |
236
+ | 降水量 [mm/h] | 12 | 31666 | 0.790 | 0.722 | **0.277** | +61.6% | — | — |
237
+ | 気温 [°C] | 24 | 18702 | 1.484 | 0.856 | **0.976** | -14.0% | 0.462 | 0.529 |
238
+ | 風速 [m/s] | 24 | 18071 | 1.590 | 0.712 | **0.817** | -14.8% | — | — |
239
+ | 風 u [m/s] | 24 | 18071 | 1.649 | 0.889 | **0.951** | -7.0% | 0.467 | 0.501 |
240
+ | 風 v [m/s] | 24 | 18071 | 1.704 | 0.970 | **1.033** | -6.4% | 0.504 | 0.537 |
241
+ | 相対湿度 [%] | 24 | 17264 | 10.040 | 5.823 | **6.744** | -15.8% | 2.976 | 3.444 |
242
+ | 降水量 [mm/h] | 24 | 31666 | 0.790 | 0.721 | **0.279** | +61.3% | — | — |
243
+
244
+ ### 2-2. 未学習3地点(新見・和気・玉野)
245
+
246
+ | 変数 | リード[h] | n | ERA5補間 | V2(公式) | **Sun*+** | 改善率 | V2 CRPS | Sun*+ CRPS |
247
+ |---|---|---|---|---|---|---|---|---|
248
+ | 気温 [°C] | 0 | 4319 | 1.598 | 1.334 | **1.205** | +9.7% | 0.727 | 0.668 |
249
+ | 風速 [m/s] | 0 | 4267 | 1.841 | 0.942 | **1.115** | -18.4% | — | — |
250
+ | 風 u [m/s] | 0 | 4267 | 1.873 | 1.152 | **1.142** | +0.9% | 0.616 | 0.610 |
251
+ | 風 v [m/s] | 0 | 4267 | 1.758 | 1.068 | **0.990** | +7.3% | 0.559 | 0.543 |
252
+ | 相対湿度 [%] | 0 | 4319 | 9.837 | 7.934 | **8.519** | -7.4% | 4.308 | 4.682 |
253
+ | 降水量 [mm/h] | 0 | 4318 | 0.664 | 0.613 | **0.129** | +79.0% | — | — |
254
+ | 気温 [°C] | 1 | 4318 | 1.598 | 0.792 | **0.693** | +12.5% | 0.430 | 0.376 |
255
+ | 風速 [m/s] | 1 | 4218 | 1.845 | 0.803 | **0.803** | -0.0% | — | — |
256
+ | 風 u [m/s] | 1 | 4218 | 1.877 | 1.002 | **0.903** | +9.9% | 0.534 | 0.485 |
257
+ | 風 v [m/s] | 1 | 4218 | 1.762 | 0.923 | **0.870** | +5.7% | 0.482 | 0.451 |
258
+ | 相対湿度 [%] | 1 | 4318 | 9.837 | 4.760 | **4.764** | -0.1% | 2.485 | 2.493 |
259
+ | 降水量 [mm/h] | 1 | 4317 | 0.664 | 0.582 | **0.119** | +79.5% | — | — |
260
+ | 気温 [°C] | 3 | 4318 | 1.598 | 0.985 | **0.929** | +5.7% | 0.537 | 0.507 |
261
+ | 風速 [m/s] | 3 | 4219 | 1.846 | 0.878 | **0.907** | -3.3% | — | — |
262
+ | 風 u [m/s] | 3 | 4219 | 1.877 | 1.091 | **1.030** | +5.5% | 0.582 | 0.552 |
263
+ | 風 v [m/s] | 3 | 4219 | 1.762 | 0.997 | **0.937** | +6.1% | 0.521 | 0.489 |
264
+ | 相対湿度 [%] | 3 | 4318 | 9.837 | 6.069 | **6.636** | -9.3% | 3.196 | 3.527 |
265
+ | 降水量 [mm/h] | 3 | 4316 | 0.664 | 0.611 | **0.141** | +77.0% | — | — |
266
+ | 気温 [°C] | 6 | 4318 | 1.598 | 1.112 | **1.063** | +4.4% | 0.609 | 0.583 |
267
+ | 風速 [m/s] | 6 | 4216 | 1.846 | 0.910 | **0.954** | -4.8% | — | — |
268
+ | 風 u [m/s] | 6 | 4216 | 1.876 | 1.118 | **1.076** | +3.8% | 0.598 | 0.578 |
269
+ | 風 v [m/s] | 6 | 4216 | 1.761 | 1.021 | **0.957** | +6.3% | 0.535 | 0.504 |
270
+ | 相対湿度 [%] | 6 | 4318 | 9.838 | 6.878 | **7.580** | -10.2% | 3.678 | 4.106 |
271
+ | 降水量 [mm/h] | 6 | 4316 | 0.664 | 0.615 | **0.132** | +78.5% | — | — |
272
+ | 気温 [°C] | 12 | 4318 | 1.598 | 1.168 | **1.128** | +3.5% | 0.640 | 0.618 |
273
+ | 風速 [m/s] | 12 | 4214 | 1.845 | 0.925 | **0.969** | -4.8% | — | — |
274
+ | 風 u [m/s] | 12 | 4214 | 1.876 | 1.128 | **1.081** | +4.1% | 0.603 | 0.583 |
275
+ | 風 v [m/s] | 12 | 4214 | 1.760 | 1.031 | **0.966** | +6.3% | 0.540 | 0.511 |
276
+ | 相対湿度 [%] | 12 | 4318 | 9.837 | 7.173 | **7.997** | -11.5% | 3.877 | 4.386 |
277
+ | 降水量 [mm/h] | 12 | 4316 | 0.664 | 0.613 | **0.129** | +79.0% | — | — |
278
+ | 気温 [°C] | 24 | 4318 | 1.598 | 1.133 | **1.141** | -0.7% | 0.620 | 0.622 |
279
+ | 風速 [m/s] | 24 | 4217 | 1.844 | 0.935 | **0.966** | -3.3% | — | — |
280
+ | 風 u [m/s] | 24 | 4217 | 1.878 | 1.126 | **1.082** | +3.9% | 0.600 | 0.582 |
281
+ | 風 v [m/s] | 24 | 4217 | 1.760 | 1.030 | **0.970** | +5.8% | 0.538 | 0.511 |
282
+ | 相対湿度 [%] | 24 | 4318 | 9.838 | 7.214 | **8.103** | -12.3% | 3.895 | 4.427 |
283
+ | 降水量 [mm/h] | 24 | 4316 | 0.664 | 0.612 | **0.132** | +78.5% | — | — |
284
+
285
+ ### 勝敗 (岡山 V2 プロトコル, RMSE)
286
+
287
+ | 地点群 | Sun*+ が良い / 全指標 |
288
+ |---|---|
289
+ | okayama_holdout | 23 / 36 |
290
+ | okayama_seen | 6 / 36 |
README.md ADDED
@@ -0,0 +1,183 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: ja
3
+ library_name: lightgbm
4
+ tags: [weather, downscaling, nowcasting, lightgbm, amedas, era5, japan]
5
+ ---
6
+ # Sun*+ — 全国版 気象ダウンスケーリング & 短期予報 (LightGBM)
7
+
8
+ Sun*+ は [Sun* V2](https://huggingface.co/summerAIO/sunstar-v2)(岡山県のみ)を **日本全国のアメダス 1,285 地点** に拡張した後継モデルです。
9
+ ERA5 (0.25°) などの格子ガイダンスを **地点 (アメダス) の実測へ補正** し、さらに初期時刻までの観測と近傍局から **リード 1/3/6/12/24 時間** の予測を行います。
10
+ 気温・風 (u/v/風速)・相対湿度・降水量の 6 変数、σ (予測のばらつき; 気温/風u/風v/湿度) も出力できます。
11
+
12
+ > ⚠️ **暫定結果 (テスト評価 6/31 チャンクのみ: 2026-01〜02 と 06 上旬)**。下表の数字は途中経過です。最終結果は評価完了後にこのリポジトリへ上書きします。ライセンス表記は公開前に決定します。
13
+
14
+ ## 何が V2 から変わったか
15
+ | | Sun* V2 | **Sun\*+** |
16
+ |---|---|---|
17
+ | 学習地点 | 岡山 22 地点 | **全国 1,285 地点 (うち評価用に 113 地点 + 岡山3地点を学習から除外)** |
18
+ | 特徴量 | 123 (ERA5 + 地形等) | V2 の 123 を完全に含み +15 (緯度経度・地方・周辺海陸比 20/50/150 km・DEM−地点標高差・沿岸×日射/風・谷/冷気溜まり指標など) |
19
+ | 観測同化 | AMeDAS 観測特徴 | 同左 + **近傍局の残差特徴 (14, 25%はドロップアウト)** |
20
+ | 学習データ | 岡山のみ | **2,233,500 行 (2023-01〜2025-12 のサンプル)**, 検証 = 2025-07〜12 |
21
+ | 降水 | — | Tweedie (p=1.3) で mm/h を直接回帰 |
22
+
23
+ ## 評価プロトコル (V2 と同一)
24
+ - テスト期間 **2026-01-01〜2026-10-31 (UTC)**, アメダス時別実測。学習は 2025-12-31 まで。
25
+ - **全国 (主)**: 学習地点の層化サンプル 160 + **学習・チューニングに一切使っていない全国 held-out 113 地点** (岡山以外の全地方から層化抽出)。
26
+ - **V2 も同じ全国地点・同じ特徴行列で評価** (公開済み V2 モデルを全国パイプラインに適用; V2 は岡山でしか学習していない)。
27
+ - **岡山 (副)**: V2 公式評価 (学習 22 地点 / 未学習 3 地点: 新見・和気・玉野) と同一地点・同一期間。
28
+ - ガイダンスは ERA5 **再解析** (完全予報)。実運用の NWP 予報では誤差が増えるため **絶対精度は楽観的** です。
29
+
30
+ ## 結果 (RMSE, 小さいほど良い) — 詳細は `COMPARISON.md`
31
+
32
+ > ⚠️ **暫定 (評価チャンク 6/31 のみ)**。最終結果で上書きされるまで、これらの数字を確定値として使わないでください。
33
+ ### 全国・全評価地点 (273) — 主
34
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 |
35
+ |---|---|---|---|---|---|---|
36
+ | 気温 [°C] | 0 | 268,693 | 2.074 | 1.922 | **1.293** | +32.7% |
37
+ | 風速 [m/s] | 0 | 265,916 | 2.242 | 1.721 | **1.260** | +26.7% |
38
+ | 風 u [m/s] | 0 | 265,916 | 2.347 | 1.743 | **1.396** | +19.9% |
39
+ | 風 v [m/s] | 0 | 265,916 | 2.115 | 1.772 | **1.379** | +22.2% |
40
+ | 相対湿度 [%] | 0 | 232,549 | 12.144 | 12.398 | **9.433** | +23.9% |
41
+ | 降水量 [mm/h] | 0 | 391,954 | 0.548 | 0.548 | **0.512** | +6.5% |
42
+ | 気温 [°C] | 1 | 268,639 | 2.074 | 1.094 | **0.771** | +29.6% |
43
+ | 風速 [m/s] | 1 | 263,238 | 2.247 | 1.296 | **0.939** | +27.6% |
44
+ | 風 u [m/s] | 1 | 263,238 | 2.353 | 1.439 | **1.097** | +23.8% |
45
+ | 風 v [m/s] | 1 | 263,238 | 2.120 | 1.448 | **1.104** | +23.7% |
46
+ | 相対湿度 [%] | 1 | 232,504 | 12.145 | 7.245 | **5.587** | +22.9% |
47
+ | 降水量 [mm/h] | 1 | 391,892 | 0.548 | 0.518 | **0.456** | +12.0% |
48
+ | 気温 [°C] | 6 | 268,577 | 2.074 | 1.534 | **1.169** | +23.8% |
49
+ | 風速 [m/s] | 6 | 262,861 | 2.247 | 1.512 | **1.154** | +23.7% |
50
+ | 風 u [m/s] | 6 | 262,861 | 2.351 | 1.629 | **1.331** | +18.3% |
51
+ | 風 v [m/s] | 6 | 262,861 | 2.119 | 1.681 | **1.330** | +20.9% |
52
+ | 相対湿度 [%] | 6 | 232,447 | 12.144 | 10.246 | **8.257** | +19.4% |
53
+ | 降水量 [mm/h] | 6 | 391,790 | 0.548 | 0.540 | **0.516** | +4.5% |
54
+ | 気温 [°C] | 24 | 268,551 | 2.074 | 1.618 | **1.247** | +23.0% |
55
+ | 風速 [m/s] | 24 | 262,850 | 2.244 | 1.555 | **1.198** | +23.0% |
56
+ | 風 u [m/s] | 24 | 262,850 | 2.350 | 1.662 | **1.372** | +17.4% |
57
+ | 風 v [m/s] | 24 | 262,850 | 2.118 | 1.711 | **1.371** | +19.8% |
58
+ | 相対湿度 [%] | 24 | 232,404 | 12.146 | 10.762 | **8.821** | +18.0% |
59
+ | 降水量 [mm/h] | 24 | 391,647 | 0.548 | 0.541 | **0.514** | +5.1% |
60
+
61
+ ### 全国・held-out 地点のみ (113; 空間汎化)
62
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 |
63
+ |---|---|---|---|---|---|---|
64
+ | 気温 [°C] | 0 | 110,841 | 2.065 | 1.960 | **1.480** | +24.5% |
65
+ | 風速 [m/s] | 0 | 109,558 | 2.205 | 1.630 | **1.459** | +10.5% |
66
+ | 風 u [m/s] | 0 | 109,558 | 2.378 | 1.704 | **1.548** | +9.2% |
67
+ | 風 v [m/s] | 0 | 109,558 | 2.025 | 1.734 | **1.557** | +10.2% |
68
+ | 相対湿度 [%] | 0 | 99,157 | 12.061 | 12.754 | **9.995** | +21.6% |
69
+ | 降水量 [mm/h] | 0 | 162,587 | 0.528 | 0.520 | **0.487** | +6.3% |
70
+ | 気温 [°C] | 1 | 110,819 | 2.065 | 1.079 | **0.817** | +24.3% |
71
+ | 風速 [m/s] | 1 | 108,435 | 2.211 | 1.237 | **0.957** | +22.7% |
72
+ | 風 u [m/s] | 1 | 108,435 | 2.384 | 1.400 | **1.104** | +21.1% |
73
+ | 風 v [m/s] | 1 | 108,435 | 2.029 | 1.415 | **1.134** | +19.9% |
74
+ | 相対湿度 [%] | 1 | 99,136 | 12.062 | 7.218 | **5.454** | +24.4% |
75
+ | 降水量 [mm/h] | 1 | 162,566 | 0.528 | 0.491 | **0.439** | +10.7% |
76
+ | 気温 [°C] | 6 | 110,802 | 2.065 | 1.532 | **1.251** | +18.4% |
77
+ | 風速 [m/s] | 6 | 108,313 | 2.211 | 1.436 | **1.200** | +16.5% |
78
+ | 風 u [m/s] | 6 | 108,313 | 2.383 | 1.576 | **1.362** | +13.6% |
79
+ | 風 v [m/s] | 6 | 108,313 | 2.028 | 1.647 | **1.402** | +14.9% |
80
+ | 相対湿度 [%] | 6 | 99,114 | 12.061 | 10.294 | **8.268** | +19.7% |
81
+ | 降水量 [mm/h] | 6 | 162,537 | 0.528 | 0.513 | **0.490** | +4.4% |
82
+ | 気温 [°C] | 24 | 110,805 | 2.065 | 1.618 | **1.335** | +17.5% |
83
+ | 風速 [m/s] | 24 | 108,317 | 2.209 | 1.472 | **1.251** | +15.0% |
84
+ | 風 u [m/s] | 24 | 108,317 | 2.383 | 1.606 | **1.406** | +12.5% |
85
+ | 風 v [m/s] | 24 | 108,317 | 2.028 | 1.675 | **1.454** | +13.2% |
86
+ | 相対湿度 [%] | 24 | 99,099 | 12.061 | 10.846 | **8.845** | +18.4% |
87
+ | 降水量 [mm/h] | 24 | 162,513 | 0.528 | 0.513 | **0.490** | +4.6% |
88
+
89
+ ### 勝敗 (RMSE; 変数×リード)
90
+ | 地点群 | Sun*+ が良い / 全指標 |
91
+ |---|---|
92
+ | 全国 全評価地点 | 36 / 36 |
93
+ | 全国 held-out | 36 / 36 |
94
+ | 全国 学習地点サンプル | 36 / 36 |
95
+ | 岡山 学習22地点 (V2 公式値と比較) | 6 / 36 |
96
+ | 岡山 未学習3地点 (V2 公式値と比較) | 23 / 36 |
97
+
98
+ ### 岡山 V2 プロトコル (副) — 学習 22 地点 (V2 は公式評価値)
99
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 |
100
+ |---|---|---|---|---|---|---|
101
+ | 気温 [°C] | 0 | 18,711 | 1.665 | 0.845 | **0.921** | -9.0% |
102
+ | 風速 [m/s] | 0 | 18,383 | 1.861 | 0.699 | **0.795** | -13.6% |
103
+ | 風 u [m/s] | 0 | 18,383 | 2.002 | 0.873 | **0.920** | -5.3% |
104
+ | 風 v [m/s] | 0 | 18,383 | 1.732 | 0.950 | **0.999** | -5.2% |
105
+ | 相対湿度 [%] | 0 | 17,272 | 11.091 | 5.856 | **6.696** | -14.3% |
106
+ | 降水量 [mm/h] | 0 | 31,673 | 0.305 | 0.720 | **0.274** | +61.9% |
107
+ | 気温 [°C] | 1 | 18,707 | 1.665 | 0.614 | **0.624** | -1.6% |
108
+ | 風速 [m/s] | 1 | 18,096 | 1.867 | 0.650 | **0.711** | -9.4% |
109
+ | 風 u [m/s] | 1 | 18,096 | 2.009 | 0.822 | **0.840** | -2.1% |
110
+ | 風 v [m/s] | 1 | 18,096 | 1.737 | 0.889 | **0.936** | -5.3% |
111
+ | 相対湿度 [%] | 1 | 17,269 | 11.091 | 4.105 | **4.356** | -6.1% |
112
+ | 降水量 [mm/h] | 1 | 31,670 | 0.305 | 0.696 | **0.248** | +64.4% |
113
+ | 気温 [°C] | 6 | 18,702 | 1.665 | 0.813 | **0.898** | -10.5% |
114
+ | 風速 [m/s] | 6 | 18,075 | 1.865 | 0.704 | **0.803** | -14.0% |
115
+ | 風 u [m/s] | 6 | 18,075 | 2.005 | 0.880 | **0.935** | -6.2% |
116
+ | 風 v [m/s] | 6 | 18,075 | 1.735 | 0.962 | **1.016** | -5.7% |
117
+ | 相対湿度 [%] | 6 | 17,264 | 11.094 | 5.543 | **6.395** | -15.4% |
118
+ | 降水量 [mm/h] | 6 | 31,666 | 0.305 | 0.721 | **0.276** | +61.7% |
119
+ | 気温 [°C] | 24 | 18,702 | 1.665 | 0.856 | **0.976** | -14.0% |
120
+ | 風速 [m/s] | 24 | 18,071 | 1.861 | 0.712 | **0.817** | -14.8% |
121
+ | 風 u [m/s] | 24 | 18,071 | 2.005 | 0.889 | **0.951** | -7.0% |
122
+ | 風 v [m/s] | 24 | 18,071 | 1.731 | 0.970 | **1.033** | -6.4% |
123
+ | 相対湿度 [%] | 24 | 17,264 | 11.093 | 5.823 | **6.744** | -15.8% |
124
+ | 降水量 [mm/h] | 24 | 31,666 | 0.305 | 0.721 | **0.279** | +61.3% |
125
+
126
+ ### 岡山 V2 プロトコル (副) — 未学習 3 地点 (新見・和気・玉野)
127
+ | 変数 | リード[h] | n | ERA5補間 | V2 | **Sun*+** | 改善率 |
128
+ |---|---|---|---|---|---|---|
129
+ | 気温 [°C] | 0 | 4,319 | 1.850 | 1.334 | **1.205** | +9.7% |
130
+ | 風速 [m/s] | 0 | 4,267 | 2.310 | 0.942 | **1.115** | -18.4% |
131
+ | 風 u [m/s] | 0 | 4,267 | 2.330 | 1.152 | **1.142** | +0.9% |
132
+ | 風 v [m/s] | 0 | 4,267 | 1.787 | 1.068 | **0.990** | +7.3% |
133
+ | 相対湿度 [%] | 0 | 4,319 | 10.908 | 7.934 | **8.519** | -7.4% |
134
+ | 降水量 [mm/h] | 0 | 4,318 | 0.142 | 0.613 | **0.129** | +79.0% |
135
+ | 気温 [°C] | 1 | 4,318 | 1.850 | 0.792 | **0.693** | +12.5% |
136
+ | 風速 [m/s] | 1 | 4,218 | 2.317 | 0.803 | **0.803** | -0.0% |
137
+ | 風 u [m/s] | 1 | 4,218 | 2.335 | 1.002 | **0.903** | +9.9% |
138
+ | 風 v [m/s] | 1 | 4,218 | 1.793 | 0.923 | **0.870** | +5.7% |
139
+ | 相対湿度 [%] | 1 | 4,318 | 10.909 | 4.760 | **4.764** | -0.1% |
140
+ | 降水量 [mm/h] | 1 | 4,317 | 0.142 | 0.582 | **0.119** | +79.5% |
141
+ | 気温 [°C] | 6 | 4,318 | 1.850 | 1.112 | **1.063** | +4.4% |
142
+ | 風速 [m/s] | 6 | 4,216 | 2.319 | 0.910 | **0.954** | -4.8% |
143
+ | 風 u [m/s] | 6 | 4,216 | 2.335 | 1.118 | **1.076** | +3.8% |
144
+ | 風 v [m/s] | 6 | 4,216 | 1.790 | 1.021 | **0.957** | +6.3% |
145
+ | 相対湿度 [%] | 6 | 4,318 | 10.907 | 6.878 | **7.580** | -10.2% |
146
+ | 降水量 [mm/h] | 6 | 4,316 | 0.142 | 0.615 | **0.132** | +78.5% |
147
+ | 気温 [°C] | 24 | 4,318 | 1.850 | 1.133 | **1.141** | -0.7% |
148
+ | 風速 [m/s] | 24 | 4,217 | 2.318 | 0.935 | **0.966** | -3.3% |
149
+ | 風 u [m/s] | 24 | 4,217 | 2.339 | 1.126 | **1.082** | +3.9% |
150
+ | 風 v [m/s] | 24 | 4,217 | 1.790 | 1.030 | **0.970** | +5.8% |
151
+ | 相対湿度 [%] | 24 | 4,318 | 10.907 | 7.214 | **8.103** | -12.3% |
152
+ | 降水量 [mm/h] | 24 | 4,316 | 0.142 | 0.612 | **0.132** | +78.5% |
153
+
154
+ ## 使い方
155
+ ```bash
156
+ pip install -r requirements.txt
157
+ python infer_example.py --station 0658 --start 2026-10-01 --end 2026-10-03 # 新見; ARCO-ERA5 を自動取得 (数分)
158
+ ```
159
+ ```python
160
+ from sunstar_plus import SunstarPlus, era5io
161
+ import pandas as pd
162
+ st = pd.read_csv("data/stations_nationwide.csv", dtype={"block_no": str})
163
+ times, fields, grid = era5io.fetch_arco("2026-10-01", "2026-10-04") # 前後24h余白込み
164
+ m = SunstarPlus(".")
165
+ pred = m.downscale(times, fields, grid, st[st.block_no == "0658"].reset_index(drop=True)) # {var: (mean, sigma)}
166
+ # 観測あり予報: m.forecast(times, fields, grid, stations, obs, lead=6) # obs: (T,S,7) アメダス実測
167
+ ```
168
+ - `ERA5` 以外のガイダンス (NWP 等) を使う場合は ERA5 互換の変数・単位 (tp=mm/h, 気温 °C 等; `sunstar_plus/features.py` 参照) に揃えてください。**学習は ERA5 のみ**なので、NWP への適用は未検証です。
169
+ - `selftest.py`: 本パッケージの予測が評価パイプラインの予測と一致することを確認する検証 (リード0, 最大差 ~1e-6)。
170
+
171
+ ## 学習・データ
172
+ - 観測: 気象庁「過去の気象データ」(obsdl) のアメダス時別値 2023-01〜2026-10 (気温・風向風速→u/v・相対湿度・降水量)。**出典: 気象庁**。
173
+ - ガイダンス: ARCO-ERA5 (Copernicus / ECMWF ERA5 再解析) 日本域。地形: AWS Terrain Tiles (Terrarium) z12, ESA WorldCover, global-land-mask。
174
+ - モデル: 変数ごとの LightGBM 残差モデル (ERA5 双線形補間に対する残差; 降水・風速は直接回帰), lr 0.06, 127 leaves。(1) 〜2025-06 で学習し 2025-07〜12 で早期終了 (3000 ラウンド上限; 気温・風・湿度は上限に達し未収束 → さらに改善余地あり) → (2) 〜2025-12 の全期間で 1.15×最適ラウンドで再学習した **2 シード平均** が公開モデル。
175
+ - Task A (リード0, 観測なし) と Task B (観測同化, リードを特徴量として 1/3/6/12/24 を共通モデル化)。
176
+ - σ: |誤差| を LightGBM で予測し係数 k を CRPS で較正 (2025-07〜12)。未学習地点向けに検証期間の held-out/学習地点 RMSE 比で拡大 (`heldout=True`)。σ は 降水・風速には未提供。
177
+
178
+ ## 制約・注意
179
+ - ERA5 再解析 = 完全予報。実 NWP 予報では劣化する。リード 1〜24h の「予報」は、ERA5 が未来の天気を知っている設定で、観測同化の効果を測るもの。
180
+ - V2 は岡山専用のため、全国評価では不利 (領域外)。**公平な比較は岡山プロトコル (副)** を参照。岡山では V2 が優る指標が残る場合がある (上表)。
181
+ - 岡山の V2 公式値は V2 自身のパイプライン (V2 版地形特徴) で算出。Sun*+ は全国版地形特徴を使うため、同じ V2 モデルでも本パイプラインで評価すると僅かに悪化する (地形特徴の差)。
182
+ - 地点は 1,285 (南鳥島を除く)。観測欠測・品質 (8/5/4 のみ採用) による欠損がある。
183
+ - 降水の RMSE 改善は小さい (降水は元々予測困難; 確率的な量)。
data/splits.json ADDED
@@ -0,0 +1,308 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "hold_v2": [
3
+ "0658",
4
+ "0664",
5
+ "0670"
6
+ ],
7
+ "okayama_seen": [
8
+ "0655",
9
+ "0656",
10
+ "47756",
11
+ "0660",
12
+ "0661",
13
+ "0663",
14
+ "0665",
15
+ "0666",
16
+ "47768",
17
+ "0668",
18
+ "0669",
19
+ "0919",
20
+ "0929",
21
+ "0930",
22
+ "1140",
23
+ "1328",
24
+ "1330",
25
+ "1371",
26
+ "1531",
27
+ "1584",
28
+ "1585",
29
+ "1618"
30
+ ],
31
+ "hold_nat": [
32
+ "0032",
33
+ "0038",
34
+ "0041",
35
+ "0051",
36
+ "0065",
37
+ "0075",
38
+ "0086",
39
+ "0091",
40
+ "0101",
41
+ "0125",
42
+ "0137",
43
+ "0172",
44
+ "0185",
45
+ "0192",
46
+ "0227",
47
+ "0279",
48
+ "0297",
49
+ "0299",
50
+ "0307",
51
+ "0322",
52
+ "0326",
53
+ "0354",
54
+ "0356",
55
+ "0362",
56
+ "0382",
57
+ "0387",
58
+ "0394",
59
+ "0451",
60
+ "0457",
61
+ "0460",
62
+ "0483",
63
+ "0499",
64
+ "0511",
65
+ "0570",
66
+ "0587",
67
+ "0593",
68
+ "0602",
69
+ "0608",
70
+ "0642",
71
+ "0688",
72
+ "0690",
73
+ "0694",
74
+ "0707",
75
+ "0724",
76
+ "0726",
77
+ "0755",
78
+ "0790",
79
+ "0806",
80
+ "0832",
81
+ "0842",
82
+ "0857",
83
+ "0863",
84
+ "0878",
85
+ "0888",
86
+ "0893",
87
+ "0939",
88
+ "0945",
89
+ "0958",
90
+ "0962",
91
+ "0968",
92
+ "1001",
93
+ "1015",
94
+ "1026",
95
+ "1029",
96
+ "1045",
97
+ "1047",
98
+ "1048",
99
+ "1066",
100
+ "1111",
101
+ "1112",
102
+ "1132",
103
+ "1161",
104
+ "1166",
105
+ "1168",
106
+ "1183",
107
+ "1189",
108
+ "1199",
109
+ "1204",
110
+ "1226",
111
+ "1229",
112
+ "1234",
113
+ "1236",
114
+ "1285",
115
+ "1297",
116
+ "1304",
117
+ "1305",
118
+ "1315",
119
+ "1319",
120
+ "1335",
121
+ "1360",
122
+ "1367",
123
+ "1376",
124
+ "1383",
125
+ "1424",
126
+ "1425",
127
+ "1456",
128
+ "1457",
129
+ "1460",
130
+ "1479",
131
+ "1508",
132
+ "1518",
133
+ "1526",
134
+ "1530",
135
+ "1533",
136
+ "1534",
137
+ "1550",
138
+ "1552",
139
+ "1554",
140
+ "1562",
141
+ "1579",
142
+ "1586",
143
+ "1616",
144
+ "1646"
145
+ ],
146
+ "nat_seen_sample": [
147
+ "0002",
148
+ "0005",
149
+ "0006",
150
+ "0010",
151
+ "0024",
152
+ "0030",
153
+ "0047",
154
+ "0049",
155
+ "0074",
156
+ "0093",
157
+ "0113",
158
+ "0132",
159
+ "0152",
160
+ "0153",
161
+ "0162",
162
+ "0169",
163
+ "0175",
164
+ "0195",
165
+ "0206",
166
+ "0221",
167
+ "0247",
168
+ "0277",
169
+ "0288",
170
+ "0315",
171
+ "0329",
172
+ "0339",
173
+ "0348",
174
+ "0355",
175
+ "0359",
176
+ "0364",
177
+ "0366",
178
+ "0375",
179
+ "0383",
180
+ "0398",
181
+ "0417",
182
+ "0424",
183
+ "0445",
184
+ "0456",
185
+ "0462",
186
+ "0468",
187
+ "0495",
188
+ "0540",
189
+ "0544",
190
+ "0555",
191
+ "0566",
192
+ "0571",
193
+ "0589",
194
+ "0600",
195
+ "0619",
196
+ "0635",
197
+ "0645",
198
+ "0679",
199
+ "0686",
200
+ "0696",
201
+ "0729",
202
+ "0730",
203
+ "0737",
204
+ "0742",
205
+ "0765",
206
+ "0769",
207
+ "0771",
208
+ "0778",
209
+ "0795",
210
+ "0830",
211
+ "0881",
212
+ "0890",
213
+ "0897",
214
+ "0925",
215
+ "0971",
216
+ "0979",
217
+ "0986",
218
+ "1004",
219
+ "1013",
220
+ "1039",
221
+ "1041",
222
+ "1042",
223
+ "1056",
224
+ "1060",
225
+ "1061",
226
+ "1081",
227
+ "1093",
228
+ "1125",
229
+ "1126",
230
+ "1130",
231
+ "1131",
232
+ "1170",
233
+ "1172",
234
+ "1180",
235
+ "1194",
236
+ "1203",
237
+ "1214",
238
+ "1219",
239
+ "1237",
240
+ "1240",
241
+ "1244",
242
+ "1245",
243
+ "1249",
244
+ "1261",
245
+ "1265",
246
+ "1266",
247
+ "1269",
248
+ "1279",
249
+ "1283",
250
+ "1289",
251
+ "1291",
252
+ "1296",
253
+ "1299",
254
+ "1301",
255
+ "1307",
256
+ "1312",
257
+ "1316",
258
+ "1324",
259
+ "1325",
260
+ "1332",
261
+ "1348",
262
+ "1354",
263
+ "1365",
264
+ "1368",
265
+ "1374",
266
+ "1386",
267
+ "1391",
268
+ "1407",
269
+ "1410",
270
+ "1458",
271
+ "1463",
272
+ "1469",
273
+ "1470",
274
+ "1473",
275
+ "1475",
276
+ "1477",
277
+ "1489",
278
+ "1493",
279
+ "1512",
280
+ "1522",
281
+ "1523",
282
+ "1524",
283
+ "1529",
284
+ "1547",
285
+ "1559",
286
+ "1563",
287
+ "1564",
288
+ "1571",
289
+ "1572",
290
+ "1573",
291
+ "1574",
292
+ "1575",
293
+ "1576",
294
+ "1577",
295
+ "1583",
296
+ "1602",
297
+ "1604",
298
+ "1609",
299
+ "1613",
300
+ "1614",
301
+ "1617",
302
+ "1626",
303
+ "1638",
304
+ "1645",
305
+ "1651",
306
+ "1653"
307
+ ]
308
+ }
data/stations_nationwide.csv ADDED
The diff for this file is too large to render. See raw diff
 
infer_example.py ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Sun*+ 推論例 (ローカル/オフライン実行用)。
2
+ python infer_example.py --station 0658 --start 2026-10-01 --end 2026-10-03 # 観測なしダウンスケーリング (ARCO-ERA5 を自動取得)
3
+ python infer_example.py --station 0658 --start 2026-10-01 --end 2026-10-03 --nc era5_dir --static static.nc # 取得済み NetCDF を使う
4
+ 予測対象時刻の前後24時間の余白を自動的に含めて取得します。出力は CSV (時刻 UTC)。"""
5
+ import argparse, pandas as pd, numpy as np
6
+ from sunstar_plus import SunstarPlus, era5io
7
+
8
+ ap = argparse.ArgumentParser()
9
+ ap.add_argument("--station", required=True, help="stations_nationwide.csv の block_no (例 0658=新見)。カンマ区切り可")
10
+ ap.add_argument("--start", required=True); ap.add_argument("--end", required=True)
11
+ ap.add_argument("--nc"); ap.add_argument("--static"); ap.add_argument("--out", default="sunstar_plus_pred.csv")
12
+ ap.add_argument("--known", action="store_true", help="学習に使った地点なら指定 (σ の未知地点拡大を外す)")
13
+ a = ap.parse_args()
14
+ st = pd.read_csv("data/stations_nationwide.csv", dtype={"block_no": str}); ids = a.station.split(",")
15
+ sel = st[st.block_no.isin(ids)].reset_index(drop=True); assert len(sel), "block_no が見つかりません"
16
+ t0, t1 = pd.Timestamp(a.start) - pd.Timedelta(hours=24), pd.Timestamp(a.end) + pd.Timedelta(hours=24)
17
+ if a.nc:
18
+ import glob; times, fields, grid = era5io.load_nc(glob.glob(a.nc + "/20??-??.nc"), a.static); keep = (times >= t0) & (times <= t1)
19
+ times = times[keep]; fields = {k: v[keep] for k, v in fields.items()}
20
+ else: times, fields, grid = era5io.fetch_arco(t0, t1)
21
+ m = SunstarPlus(".")
22
+ out = m.downscale(times, fields, grid, sel, sigma=True, heldout=not a.known)
23
+ rows = []
24
+ for j, r in sel.iterrows():
25
+ d = pd.DataFrame({"time_utc": times, "block_no": r.block_no})
26
+ for v, (mean, sg) in out.items():
27
+ d[v] = mean[:, j]
28
+ if sg is not None: d[v + "_sigma"] = sg[:, j]
29
+ rows.append(d)
30
+ res = pd.concat(rows); res = res[(res.time_utc >= pd.Timestamp(a.start)) & (res.time_utc <= pd.Timestamp(a.end))]
31
+ res.to_csv(a.out, index=False); print(res.head(), "->", a.out)
models/config.json ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "models": {
3
+ "A": {
4
+ "temp_c": [
5
+ "p1_A_temp_c_s0.txt.gz",
6
+ "p1_A_temp_c_s1.txt.gz"
7
+ ],
8
+ "wind_u_ms": [
9
+ "p1_A_wind_u_ms_s0.txt.gz",
10
+ "p1_A_wind_u_ms_s1.txt.gz"
11
+ ],
12
+ "wind_v_ms": [
13
+ "p1_A_wind_v_ms_s0.txt.gz",
14
+ "p1_A_wind_v_ms_s1.txt.gz"
15
+ ],
16
+ "humidity_pct": [
17
+ "p1_A_humidity_pct_s0.txt.gz",
18
+ "p1_A_humidity_pct_s1.txt.gz"
19
+ ],
20
+ "wind_speed_ms": [
21
+ "p1_A_wind_speed_ms_s0.txt.gz",
22
+ "p1_A_wind_speed_ms_s1.txt.gz"
23
+ ],
24
+ "precip_mm": [
25
+ "p1_A_precip_mm_s0.txt.gz",
26
+ "p1_A_precip_mm_s1.txt.gz"
27
+ ]
28
+ },
29
+ "B": {
30
+ "temp_c": [
31
+ "p1_B_temp_c_s0.txt.gz",
32
+ "p1_B_temp_c_s1.txt.gz"
33
+ ],
34
+ "wind_u_ms": [
35
+ "p1_B_wind_u_ms_s0.txt.gz",
36
+ "p1_B_wind_u_ms_s1.txt.gz"
37
+ ],
38
+ "wind_v_ms": [
39
+ "p1_B_wind_v_ms_s0.txt.gz",
40
+ "p1_B_wind_v_ms_s1.txt.gz"
41
+ ],
42
+ "humidity_pct": [
43
+ "p1_B_humidity_pct_s0.txt.gz",
44
+ "p1_B_humidity_pct_s1.txt.gz"
45
+ ],
46
+ "wind_speed_ms": [
47
+ "p1_B_wind_speed_ms_s0.txt.gz",
48
+ "p1_B_wind_speed_ms_s1.txt.gz"
49
+ ],
50
+ "precip_mm": [
51
+ "p1_B_precip_mm_s0.txt.gz",
52
+ "p1_B_precip_mm_s1.txt.gz"
53
+ ]
54
+ }
55
+ },
56
+ "sigma": {
57
+ "A": {
58
+ "temp_c": {
59
+ "file": "sigma_A_temp_c.txt.gz",
60
+ "k": 0.9500000000000001,
61
+ "ratio": {
62
+ "0": 1.1751646832790135
63
+ }
64
+ },
65
+ "wind_u_ms": {
66
+ "file": "sigma_A_wind_u_ms.txt.gz",
67
+ "k": 1.0,
68
+ "ratio": {
69
+ "0": 1.1988303694521603
70
+ }
71
+ },
72
+ "wind_v_ms": {
73
+ "file": "sigma_A_wind_v_ms.txt.gz",
74
+ "k": 1.0,
75
+ "ratio": {
76
+ "0": 1.2928762775998992
77
+ }
78
+ },
79
+ "humidity_pct": {
80
+ "file": "sigma_A_humidity_pct.txt.gz",
81
+ "k": 1.0,
82
+ "ratio": {
83
+ "0": 1.0990371628553905
84
+ }
85
+ }
86
+ },
87
+ "B": {
88
+ "temp_c": {
89
+ "file": "sigma_B_temp_c.txt.gz",
90
+ "k": 0.9500000000000001,
91
+ "ratio": {
92
+ "1": 1.0061438854986164,
93
+ "3": 1.0174361461923962,
94
+ "6": 1.0289456111639248,
95
+ "12": 1.0354925287045398,
96
+ "24": 1.0265270390667012
97
+ }
98
+ },
99
+ "wind_u_ms": {
100
+ "file": "sigma_B_wind_u_ms.txt.gz",
101
+ "k": 1.0,
102
+ "ratio": {
103
+ "1": 1.0191275642002626,
104
+ "3": 1.03452439399401,
105
+ "6": 1.0521289627650094,
106
+ "12": 1.0561387120928374,
107
+ "24": 1.0522187630848694
108
+ }
109
+ },
110
+ "wind_v_ms": {
111
+ "file": "sigma_B_wind_v_ms.txt.gz",
112
+ "k": 0.9500000000000001,
113
+ "ratio": {
114
+ "1": 1.0607771732563274,
115
+ "3": 1.0918507004665554,
116
+ "6": 1.1180886920358573,
117
+ "12": 1.1281368142183115,
118
+ "24": 1.1264799049911653
119
+ }
120
+ },
121
+ "humidity_pct": {
122
+ "file": "sigma_B_humidity_pct.txt.gz",
123
+ "k": 0.9500000000000001,
124
+ "ratio": {
125
+ "1": 0.9591965464460005,
126
+ "3": 0.9769487514773961,
127
+ "6": 0.9879567700056188,
128
+ "12": 0.9909772191854498,
129
+ "24": 0.9883133523207693
130
+ }
131
+ }
132
+ }
133
+ },
134
+ "tag": "p1"
135
+ }
models/features_A.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"fnames": ["t2m", "d2m", "u10", "v10", "u100", "v100", "spd10", "spd100", "rh", "dd", "sp", "msl", "tp", "ssrd", "fdir", "tcc", "lcc", "mcc", "hcc", "tcwv", "cape", "skt", "skt_t2", "gust", "shear", "t2m_lapse", "wdir_s", "wdir_c", "t2m_d-6", "t2m_d-3", "t2m_d-1", "t2m_d+1", "t2m_d+3", "t2m_d+6", "skt_d-3", "skt_d-1", "skt_d+1", "skt_d+3", "u10_d-3", "u10_d-1", "u10_d+1", "u10_d+3", "v10_d-3", "v10_d-1", "v10_d+1", "v10_d+3", "spd10_d-3", "spd10_d-1", "spd10_d+1", "spd10_d+3", "tcc_d-3", "tcc_d-1", "tcc_d+1", "tcc_d+3", "rh_d-6", "rh_d-3", "rh_d-1", "rh_d+1", "rh_d+3", "rh_d+6", "ssrd_d-3", "ssrd_d-1", "ssrd_d+1", "ssrd_d+3", "tp_d-3", "tp_d-1", "tp_d+1", "tp_d+3", "t2m_mean24", "t2m_prev24", "tp_sum_p3", "tp_sum_f3", "tp_sum_p12", "tcc_mean6", "ssrd_mean6", "t2m_max24", "t2m_min24", "tp_nbmax", "tp_nbmin", "tp_wmean", "tp_wmax", "cape_wmax", "t2m_nbspread", "t2m_gy", "t2m_gx", "tcc_wmean", "elev", "dz", "orog_era", "lsm", "sdor", "hv", "lv", "terr_elevation_m", "terr_slope_deg", "terr_aspect_sin", "terr_aspect_cos", "terr_tpi_500m", "terr_tpi_2km", "terr_tpi_10km", "terr_relief_1km", "terr_relief_5km", "terr_coast_dist_km", "terr_land_mask", "terr_lc_tree_frac", "terr_lc_crop_frac", "terr_lc_built_frac", "terr_lc_water_frac", "terr_lc_grass_frac", "terr_lc_shrub_frac", "elev_mean_r10", "elev_mean_r40", "hr_s", "hr_c", "doy_s", "doy_c", "hr_utc", "sun_elev", "down_slope_flow", "flow_x_tpi2k", "flow_x_tpi10k", "night_x_tpi2k", "night_x_dz", "x_lat", "x_lon", "x_region", "x_t_elev_max_r10", "x_t_elev_min_r40", "x_t_sea_frac_20km", "x_t_sea_frac_50km", "x_t_sea_frac_150km", "x_dem_minus_stn", "x_coast_x_sun", "x_coast_x_spd", "x_t2m_dem", "x_valley_night", "x_cold_pool", "x_abs_lat_x_doy", "regimeB"]}
models/features_B.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"fnames": ["t2m", "d2m", "u10", "v10", "u100", "v100", "spd10", "spd100", "rh", "dd", "sp", "msl", "tp", "ssrd", "fdir", "tcc", "lcc", "mcc", "hcc", "tcwv", "cape", "skt", "skt_t2", "gust", "shear", "t2m_lapse", "wdir_s", "wdir_c", "t2m_d-6", "t2m_d-3", "t2m_d-1", "t2m_d+1", "t2m_d+3", "t2m_d+6", "skt_d-3", "skt_d-1", "skt_d+1", "skt_d+3", "u10_d-3", "u10_d-1", "u10_d+1", "u10_d+3", "v10_d-3", "v10_d-1", "v10_d+1", "v10_d+3", "spd10_d-3", "spd10_d-1", "spd10_d+1", "spd10_d+3", "tcc_d-3", "tcc_d-1", "tcc_d+1", "tcc_d+3", "rh_d-6", "rh_d-3", "rh_d-1", "rh_d+1", "rh_d+3", "rh_d+6", "ssrd_d-3", "ssrd_d-1", "ssrd_d+1", "ssrd_d+3", "tp_d-3", "tp_d-1", "tp_d+1", "tp_d+3", "t2m_mean24", "t2m_prev24", "tp_sum_p3", "tp_sum_f3", "tp_sum_p12", "tcc_mean6", "ssrd_mean6", "t2m_max24", "t2m_min24", "tp_nbmax", "tp_nbmin", "tp_wmean", "tp_wmax", "cape_wmax", "t2m_nbspread", "t2m_gy", "t2m_gx", "tcc_wmean", "elev", "dz", "orog_era", "lsm", "sdor", "hv", "lv", "terr_elevation_m", "terr_slope_deg", "terr_aspect_sin", "terr_aspect_cos", "terr_tpi_500m", "terr_tpi_2km", "terr_tpi_10km", "terr_relief_1km", "terr_relief_5km", "terr_coast_dist_km", "terr_land_mask", "terr_lc_tree_frac", "terr_lc_crop_frac", "terr_lc_built_frac", "terr_lc_water_frac", "terr_lc_grass_frac", "terr_lc_shrub_frac", "elev_mean_r10", "elev_mean_r40", "hr_s", "hr_c", "doy_s", "doy_c", "hr_utc", "sun_elev", "down_slope_flow", "flow_x_tpi2k", "flow_x_tpi10k", "night_x_tpi2k", "night_x_dz", "x_lat", "x_lon", "x_region", "x_t_elev_max_r10", "x_t_elev_min_r40", "x_t_sea_frac_20km", "x_t_sea_frac_50km", "x_t_sea_frac_150km", "x_dem_minus_stn", "x_coast_x_sun", "x_coast_x_spd", "x_t2m_dem", "x_valley_night", "x_cold_pool", "x_abs_lat_x_doy", "regimeB", "o_T0", "o_T_l1", "o_T_l2", "o_T_l3", "o_T_l6", "o_T_l12", "o_T_l24", "o_U0", "o_U_l1", "o_U_l2", "o_U_l3", "o_U_l6", "o_U_l12", "o_U_l24", "o_V0", "o_V_l1", "o_V_l2", "o_V_l3", "o_V_l6", "o_V_l12", "o_V_l24", "o_RH0", "o_RH_l1", "o_RH_l2", "o_RH_l3", "o_RH_l6", "o_RH_l12", "o_RH_l24", "o_SPD0", "o_SPD_l1", "o_SPD_l2", "o_SPD_l3", "o_SPD_l6", "o_SPD_l12", "o_SPD_l24", "o_P0", "o_P_l1", "o_P_l2", "o_P_l3", "o_P_l6", "o_P_l12", "o_P_l24", "r_T0", "r_T_l1", "r_T_l2", "r_T_l3", "r_T_l6", "r_T_l12", "r_T_l24", "r_T_m3", "r_T_m6", "r_T_m24", "r_T_m72", "r_T_m168", "r_T_sd24", "r_U0", "r_U_l1", "r_U_l2", "r_U_l3", "r_U_l6", "r_U_l12", "r_U_l24", "r_U_m3", "r_U_m6", "r_U_m24", "r_U_m72", "r_U_m168", "r_U_sd24", "r_V0", "r_V_l1", "r_V_l2", "r_V_l3", "r_V_l6", "r_V_l12", "r_V_l24", "r_V_m3", "r_V_m6", "r_V_m24", "r_V_m72", "r_V_m168", "r_V_sd24", "r_RH0", "r_RH_l1", "r_RH_l2", "r_RH_l3", "r_RH_l6", "r_RH_l12", "r_RH_l24", "r_RH_m3", "r_RH_m6", "r_RH_m24", "r_RH_m72", "r_RH_m168", "r_RH_sd24", "r_SPD0", "r_SPD_l1", "r_SPD_l2", "r_SPD_l3", "r_SPD_l6", "r_SPD_l12", "r_SPD_l24", "r_SPD_m3", "r_SPD_m6", "r_SPD_m24", "r_SPD_m72", "r_SPD_m168", "r_SPD_sd24", "o_T_trend3", "o_P_sum3", "o_P_sum24", "db_t2m", "db_u10", "db_v10", "db_rh", "db_spd10", "db_d2m", "lead", "nb_T0", "nb_U0", "nb_V0", "nb_RH0", "nb_SPD0", "nb_T_m6", "nb_U_m6", "nb_V_m6", "nb_RH_m6", "nb_SPD_m6", "nb_P0", "nb_P_m3", "nb_Pmax3", "nb_w"]}
models/p1_A_humidity_pct_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:461c57c56ae83f73fcbde74fbd88bc71d02003b506878725452d7c34f0eea1eb
3
+ size 16456907
models/p1_A_humidity_pct_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:214c19496e67188bfbb2f0eb1f0d6dd02c8479b860205f68974b7c950972ee50
3
+ size 16453113
models/p1_A_precip_mm_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b052bd9f0bee53e373e762a2105c9897f2adc8fc8561a6606369e30152bca8b
3
+ size 2577245
models/p1_A_precip_mm_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08b15d80f6d07e2cab99805ca076b08dd2a98eee56666ad35c2cf95a34ecea5a
3
+ size 2577572
models/p1_A_temp_c_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e924f4303871915a228841f0284ac9622eea3968dff6a3c191e4e9eda1b660e
3
+ size 16614774
models/p1_A_temp_c_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:54a433607af6d8de314a47ed1aee267e43b8243e0565a08f2b50edbca7069d5b
3
+ size 16611892
models/p1_A_wind_speed_ms_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:03bfd88c8d138f0addae26e6445f17bae6e5eb33762c3165c21125a7ea70da7e
3
+ size 16535455
models/p1_A_wind_speed_ms_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ed56ea936e49a07ee1cd76fd52839b28eaf4a0fb64705568a24c24c6bd6c09c
3
+ size 16534370
models/p1_A_wind_u_ms_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:707e42cabbd2708fb8d1be32d72156868d38f440e8dd6df6a4bfcec965e9896c
3
+ size 16493101
models/p1_A_wind_u_ms_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f127666edec565c53681cd8ba2cc909d8a32f907daa074c377745639d0afadcb
3
+ size 16499279
models/p1_A_wind_v_ms_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6f68e8f84dce47adc549f0872bed5dda8a7e56532244fd4d3380cb9fc2ac80c
3
+ size 16506838
models/p1_A_wind_v_ms_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5fcbb7c8112340744f7b2c0f4fa0eb8d95316fa12d5ac3c1634813980f4912a9
3
+ size 16511974
models/p1_B_humidity_pct_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd6aeab88df9827a2a41f93488e248a9fd339651d635bac702c69b6fb357efad
3
+ size 16644830
models/p1_B_humidity_pct_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:647fa19536748cb5d47b979164d5dc8b04a32e0bb96b2e831eea8d75d4e91de5
3
+ size 16641932
models/p1_B_precip_mm_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d233ead794977bd167434aac1163b02f51b5dab268708fd138bf99d4fd338a37
3
+ size 2692014
models/p1_B_precip_mm_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c05d3b54df9ec2774eb29c4ad48b19c817da9258c24fdbb039dffa83537ed1e
3
+ size 2692276
models/p1_B_temp_c_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:affb3b979f65f85379d7f653b961743284489f6dcf4987484df81a32bc96acdf
3
+ size 16734192
models/p1_B_temp_c_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:36180a5e2c61f498795ec9eccee9235e56f8ea6ce3b0f0a5f14e11deb023b0f8
3
+ size 16732053
models/p1_B_wind_speed_ms_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d4becd19b2761c62c8c00ce8ca26636e5c6c6703ac501d015ba3576d1987d58
3
+ size 16692176
models/p1_B_wind_speed_ms_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64b0ccf4836f29b8559bc427b0decb04b35be5286a3664264815052e91841222
3
+ size 16690068
models/p1_B_wind_u_ms_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16ee588a1a2c41d46c678d778ae035f3dde15701cd417f758e6dadeb83f3fbf2
3
+ size 16643262
models/p1_B_wind_u_ms_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:46bfb3245fe86a38a5afdacdfee1a3ad19d2f79da67a0988a8d477be9cde24a6
3
+ size 16635356
models/p1_B_wind_v_ms_s0.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f16ee43151246975f18574a7391a21bf9fdd634f4c80b076e538def84077cac
3
+ size 16612969
models/p1_B_wind_v_ms_s1.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8aeef1c2d5d90da43b9c2a1be5902514af8934029716dec456264a379058c2a0
3
+ size 16608834
models/sigma_A_humidity_pct.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61acb0dd59dbbd1266ad78eb2f5043ba47a290f5759bb8a30420eba26d14ece6
3
+ size 167581
models/sigma_A_temp_c.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1b7844561b6a9e7524ca35c5b7db5bf3f82d7e4b55751e569bd0fa4f52ae165
3
+ size 169312
models/sigma_A_wind_u_ms.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5d8c757d9c7bf631f552c18773ec6ce2f268cae900f1332079d55a95c1a045b
3
+ size 166721
models/sigma_A_wind_v_ms.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4f22c6de82b3557e60a0d304a08fc9d5eba55af07e4cbbd6c56dafec75a366b9
3
+ size 167001
models/sigma_B_humidity_pct.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7dc25855100360faf3b21fe4b1d12c2586c5882c3540ab2d8781988cde911372
3
+ size 172278
models/sigma_B_temp_c.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f26dab2affb8b9c5ed8ccad10da37214b206c73e4d868b908f8f508f40e3a9e2
3
+ size 173371
models/sigma_B_wind_u_ms.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b59b11fcf5162d2627d9346d8c7eb204200f4f7c8ecdc3a1d6aff1195d4309c7
3
+ size 171913
models/sigma_B_wind_v_ms.txt.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c246d55fdfef4259b84e637b08953f6fd1cd9243806020e081b3f03ca44e3c2
3
+ size 172615
requirements.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ numpy
2
+ pandas
3
+ scipy
4
+ lightgbm>=4.0
5
+ xarray
6
+ zarr
7
+ gcsfs
8
+ netCDF4
9
+ dask
results/metrics_nationwide_and_okayama.csv ADDED
The diff for this file is too large to render. See raw diff
 
results/v2_official_okayama.csv ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ group,lead_h,var,v2o_rmse
2
+ okayama_seen,0,temp_c,0.8446679392900487
3
+ okayama_seen,0,wind_u_ms,0.8731989444628082
4
+ okayama_seen,0,wind_v_ms,0.9497237512653542
5
+ okayama_seen,0,humidity_pct,5.855606201527987
6
+ okayama_seen,0,precip_mm,0.7201785029914338
7
+ okayama_seen,0,wind_speed_ms,0.6994221758470566
8
+ okayama_holdout,0,temp_c,1.3339056851792048
9
+ okayama_holdout,0,wind_u_ms,1.1521146101963686
10
+ okayama_holdout,0,wind_v_ms,1.06770412230384
11
+ okayama_holdout,0,humidity_pct,7.934065891180142
12
+ okayama_holdout,0,precip_mm,0.6131773889560278
13
+ okayama_holdout,0,wind_speed_ms,0.9418489324454394
14
+ okayama_seen,1,temp_c,0.6142638162488309
15
+ okayama_seen,1,wind_u_ms,0.8224830389576552
16
+ okayama_seen,1,wind_v_ms,0.8890610056762683
17
+ okayama_seen,1,humidity_pct,4.104506872804167
18
+ okayama_seen,1,precip_mm,0.6960890718650667
19
+ okayama_seen,1,wind_speed_ms,0.6502970820487856
20
+ okayama_seen,3,temp_c,0.7496734092102529
21
+ okayama_seen,3,wind_u_ms,0.8680952943164121
22
+ okayama_seen,3,wind_v_ms,0.9482316156885284
23
+ okayama_seen,3,humidity_pct,5.123878089604943
24
+ okayama_seen,3,precip_mm,0.7204841952388997
25
+ okayama_seen,3,wind_speed_ms,0.6916967402119256
26
+ okayama_seen,6,temp_c,0.8130027825761901
27
+ okayama_seen,6,wind_u_ms,0.8802205899732691
28
+ okayama_seen,6,wind_v_ms,0.9618554610639676
29
+ okayama_seen,6,humidity_pct,5.542997069256404
30
+ okayama_seen,6,precip_mm,0.7212372206057139
31
+ okayama_seen,6,wind_speed_ms,0.7042079753743455
32
+ okayama_seen,12,temp_c,0.8447050112785129
33
+ okayama_seen,12,wind_u_ms,0.885150430595629
34
+ okayama_seen,12,wind_v_ms,0.9639535749350343
35
+ okayama_seen,12,humidity_pct,5.732130312723504
36
+ okayama_seen,12,precip_mm,0.7216604724697135
37
+ okayama_seen,12,wind_speed_ms,0.7081275920621516
38
+ okayama_seen,24,temp_c,0.8558055030609867
39
+ okayama_seen,24,wind_u_ms,0.8886927766777
40
+ okayama_seen,24,wind_v_ms,0.9704600118456583
41
+ okayama_seen,24,humidity_pct,5.8228404327218275
42
+ okayama_seen,24,precip_mm,0.721480595019244
43
+ okayama_seen,24,wind_speed_ms,0.7119255228613133
44
+ okayama_holdout,1,temp_c,0.7918285362729199
45
+ okayama_holdout,1,wind_u_ms,1.0020327901045583
46
+ okayama_holdout,1,wind_v_ms,0.9225801800558296
47
+ okayama_holdout,1,humidity_pct,4.760351150264549
48
+ okayama_holdout,1,precip_mm,0.5818717161704787
49
+ okayama_holdout,1,wind_speed_ms,0.8031471094377639
50
+ okayama_holdout,3,temp_c,0.9853017624428246
51
+ okayama_holdout,3,wind_u_ms,1.0905539683761392
52
+ okayama_holdout,3,wind_v_ms,0.9970375289848942
53
+ okayama_holdout,3,humidity_pct,6.0690299514620705
54
+ okayama_holdout,3,precip_mm,0.610611447814481
55
+ okayama_holdout,3,wind_speed_ms,0.8783581996909686
56
+ okayama_holdout,6,temp_c,1.1121035950490492
57
+ okayama_holdout,6,wind_u_ms,1.1180595130808928
58
+ okayama_holdout,6,wind_v_ms,1.021238849949731
59
+ okayama_holdout,6,humidity_pct,6.878327311092592
60
+ okayama_holdout,6,precip_mm,0.614794598157702
61
+ okayama_holdout,6,wind_speed_ms,0.9095938948474201
62
+ okayama_holdout,12,temp_c,1.1683604049858658
63
+ okayama_holdout,12,wind_u_ms,1.12801792478262
64
+ okayama_holdout,12,wind_v_ms,1.0309160354301237
65
+ okayama_holdout,12,humidity_pct,7.173230875078692
66
+ okayama_holdout,12,precip_mm,0.6125185790812437
67
+ okayama_holdout,12,wind_speed_ms,0.9246225628995648
68
+ okayama_holdout,24,temp_c,1.1326607924782732
69
+ okayama_holdout,24,wind_u_ms,1.1260415056603073
70
+ okayama_holdout,24,wind_v_ms,1.0296770066925827
71
+ okayama_holdout,24,humidity_pct,7.214382330649627
72
+ okayama_holdout,24,precip_mm,0.6121249195329943
73
+ okayama_holdout,24,wind_speed_ms,0.9347908648148685
selftest.py ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """自己検証: 本パッケージ (SunstarPlus.downscale) の予測が、評価パイプライン (eval_stream.py) の予測 (lead 0) と一致するか。
2
+ python selftest.py <評価chunk.npz> <chunk 開始時刻> <ERA5月別ncディレクトリ> <static.nc> (開発環境用; 配布物には評価データは含まれない)"""
3
+ import sys, glob, numpy as np, pandas as pd
4
+ from sunstar_plus import SunstarPlus, era5io
5
+ chunk, t_start, era_dir, static = sys.argv[1:5]
6
+ z = np.load(chunk); t = z["L0_t"]; s = z["L0_s"]; sp = z["L0_sp"]
7
+ st = pd.read_csv("data/stations_nationwide.csv", dtype={"block_no": str}).reset_index(drop=True)
8
+ T0 = pd.Timestamp("2023-01-01"); times_abs = T0 + pd.to_timedelta(t, "h")
9
+ w0 = pd.Timestamp(t_start) + pd.Timedelta(hours=48); w1 = w0 + pd.Timedelta(hours=72)
10
+ stn = np.unique(s)[::max(1, len(np.unique(s)) // 12)][:12]
11
+ sel = st.iloc[stn].reset_index(drop=True)
12
+ times, fields, grid = era5io.load_nc(sorted(glob.glob(era_dir + "/20??-??.nc"))[-1:] if False else [f for f in glob.glob(era_dir + "/20??-??.nc") if f[-10:-3] in (pd.Timestamp(t_start).strftime("%Y-%m"), (pd.Timestamp(t_start) - pd.Timedelta(days=5)).strftime("%Y-%m"))], static)
13
+ keep = (times >= w0 - pd.Timedelta(hours=24)) & (times <= w1 + pd.Timedelta(hours=24)); times = times[keep]; fields = {k: v[keep] for k, v in fields.items()}
14
+ m = SunstarPlus("."); out = m.downscale(times, fields, grid, sel, sigma=False)
15
+ VARS = ["temp_c", "wind_u_ms", "wind_v_ms", "humidity_pct", "wind_speed_ms", "precip_mm"]
16
+ for k, v in enumerate(VARS):
17
+ d = []
18
+ for j, si in enumerate(stn):
19
+ mk = (s == si) & (times_abs >= w0) & (times_abs <= w1)
20
+ if not mk.any(): continue
21
+ idx = np.searchsorted(times, times_abs[mk]); d.append(np.abs(out[v][0][idx, j] - sp[mk, k]))
22
+ d = np.concatenate(d); d = d[np.isfinite(d)]
23
+ print(f"{v:14s} n={len(d):5d} max|diff|={d.max():.4g} mean|diff|={d.mean():.4g}")
sunstar_plus/__init__.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ """Sun*+ : 全国版 地点天気 ダウンスケーリング/観測同化モデル (LightGBM)"""
2
+ from .features import Grid, StaticSet, load_stations, build_features
3
+ from .model import SunstarPlus, VARS, LABEL
4
+ from .terrain import compute_terrain
5
+ from . import era5io
6
+ __version__ = "1.0"
sunstar_plus/era5io.py ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ERA5 の読み込みヘルパ。ARCO-ERA5 (gs://gcp-public-data-arco-era5, 匿名) から日本域を取得 / 取得済み NetCDF を Sun*+ 入力形式に変換。"""
2
+ import numpy as np, pandas as pd, xarray as xr
3
+ from .features import Grid
4
+
5
+ ERA_RAW = ["2m_temperature", "2m_dewpoint_temperature", "10m_u_component_of_wind", "10m_v_component_of_wind", "100m_u_component_of_wind", "100m_v_component_of_wind",
6
+ "surface_solar_radiation_downwards", "total_sky_direct_solar_radiation_at_surface", "total_precipitation", "total_cloud_cover", "low_cloud_cover",
7
+ "medium_cloud_cover", "high_cloud_cover", "mean_sea_level_pressure", "surface_pressure", "total_column_water_vapour",
8
+ "convective_available_potential_energy", "instantaneous_10m_wind_gust", "skin_temperature"]
9
+ STATIC = ["geopotential_at_surface", "land_sea_mask", "standard_deviation_of_orography", "high_vegetation_cover", "low_vegetation_cover"]
10
+ LAT = slice(46.5, 23.5); LON = slice(122.0, 146.5)
11
+ ZARR = "gs://gcp-public-data-arco-era5/ar/full_37-1h-0p25deg-chunk-1.zarr-v3"
12
+
13
+ def fetch_arco(t0, t1, lat=LAT, lon=LON):
14
+ """[t0, t1] (UTC, 1時間刻み) の日本域 ERA5 を取得し (times, fields, Grid) を返す。数分かかる (全球1時間チャンクを読むため)。"""
15
+ ds = xr.open_zarr(ZARR, chunks=None, storage_options={"token": "anon", "timeout": 60})
16
+ times = pd.date_range(t0, t1, freq="1h")
17
+ sub = xr.Dataset({v: ds[v].sel(time=times, latitude=lat, longitude=lon) for v in ERA_RAW}).chunk({"time": 1}).compute()
18
+ st = ds[STATIC]
19
+ if "time" in st.dims: st = st.sel(time="2024-01-01T00:00", drop=True)
20
+ st = st.sel(latitude=lat, longitude=lon).compute()
21
+ return convert(sub, st)
22
+
23
+ def convert(sub, static):
24
+ """ERA5 生単位の Dataset -> (times DatetimeIndex, fields dict[(T,H,W)], Grid)"""
25
+ f = {}
26
+ f["2m_temperature_c"] = sub["2m_temperature"].values - 273.15; f["2m_dewpoint_temperature_c"] = sub["2m_dewpoint_temperature"].values - 273.15
27
+ f["skin_temperature_c"] = sub["skin_temperature"].values - 273.15
28
+ f["total_precipitation"] = sub["total_precipitation"].values * 1000.0
29
+ f["mean_sea_level_pressure_hpa"] = sub["mean_sea_level_pressure"].values / 100.0; f["surface_pressure_hpa"] = sub["surface_pressure"].values / 100.0
30
+ for v in ("10m_u_component_of_wind", "10m_v_component_of_wind", "100m_u_component_of_wind", "100m_v_component_of_wind", "surface_solar_radiation_downwards",
31
+ "total_sky_direct_solar_radiation_at_surface", "total_cloud_cover", "low_cloud_cover", "medium_cloud_cover", "high_cloud_cover",
32
+ "total_column_water_vapour", "convective_available_potential_energy", "instantaneous_10m_wind_gust"):
33
+ f[v] = sub[v].values
34
+ f = {k: v.astype(np.float32) for k, v in f.items()}
35
+ g = Grid.from_arrays(static.latitude.values, static.longitude.values, static["geopotential_at_surface"].values / 9.80665, static["land_sea_mask"].values,
36
+ static["standard_deviation_of_orography"].values, static["high_vegetation_cover"].values, static["low_vegetation_cover"].values)
37
+ return pd.DatetimeIndex(sub.time.values), f, g
38
+
39
+ def load_nc(paths, static_nc):
40
+ """取得済み月別 NetCDF (ARCO 生変数名) を読み込み (times, fields, Grid)"""
41
+ parts = [xr.open_dataset(p).load() for p in sorted(paths)]
42
+ sub = xr.concat(parts, dim="time")
43
+ return convert(sub, xr.open_dataset(static_nc).load())
sunstar_plus/features.py ADDED
@@ -0,0 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Sun*+ ライブラリ: 全国版 特徴量生成 (Sun* V2 の特徴量定義を完全に含み, 全国向け特徴を追加)。
2
+ 出力の特徴列は先頭が V2 と同名同順 (V2 モデルを同じ特徴行列で評価可能)、その後ろに Sun*+ 追加特徴が続く。"""
3
+ from pathlib import Path
4
+ import numpy as np, pandas as pd, xarray as xr
5
+ from scipy.ndimage import uniform_filter
6
+
7
+ HOLD_V2 = ["0658", "0664", "0670"]
8
+ REGIME_B_START = "2024-10-01"
9
+ VARS6 = ["temp_c", "wind_u_ms", "wind_v_ms", "humidity_pct", "wind_speed_ms", "precip_mm"]
10
+ TERRAIN_VARS = ["elevation_m", "slope_deg", "aspect_sin", "aspect_cos", "tpi_500m", "tpi_2km", "tpi_10km", "relief_1km", "relief_5km",
11
+ "coast_dist_km", "land_mask", "lc_tree_frac", "lc_crop_frac", "lc_built_frac", "lc_water_frac", "lc_grass_frac", "lc_shrub_frac"]
12
+ ERA_VARS = ["2m_temperature_c", "2m_dewpoint_temperature_c", "10m_u_component_of_wind", "10m_v_component_of_wind",
13
+ "100m_u_component_of_wind", "100m_v_component_of_wind", "surface_pressure_hpa", "total_precipitation",
14
+ "surface_solar_radiation_downwards", "total_cloud_cover", "low_cloud_cover", "medium_cloud_cover", "high_cloud_cover",
15
+ "total_column_water_vapour", "convective_available_potential_energy", "mean_sea_level_pressure_hpa",
16
+ "skin_temperature_c", "instantaneous_10m_wind_gust", "total_sky_direct_solar_radiation_at_surface"]
17
+ REGION = {**{p: 0 for p in range(11, 25)}, **{p: 1 for p in range(31, 37)}, **{p: 2 for p in range(40, 47)}, 48: 3, 49: 3, 54: 3,
18
+ 55: 4, 56: 4, 57: 4, 50: 5, 51: 5, 52: 5, 53: 5, **{p: 6 for p in range(60, 66)}, **{p: 7 for p in (66, 67, 68, 69, 81)},
19
+ **{p: 8 for p in range(71, 75)}, **{p: 9 for p in range(82, 89)}, 91: 10}
20
+
21
+ def rh_from(t_c, td_c):
22
+ a, b = 17.625, 243.04
23
+ return np.clip(100 * np.exp(a * td_c / (b + td_c)) / np.exp(a * t_c / (b + t_c)), 0, 100)
24
+
25
+ # ---------------------------------------------------------------- stations
26
+ def load_stations(csv_path):
27
+ """data/stations_nationwide.csv (地点メタ + 地形特徴) を読む"""
28
+ st = pd.read_csv(csv_path, dtype={"block_no": str})
29
+ st["region"] = st["region"].astype(int)
30
+ return st
31
+
32
+ # ---------------------------------------------------------------- ERA5 (Japan grid)
33
+ class Grid:
34
+ """ERA5 0.25° 格子の静的場。Grid(static_nc) もしくは Grid.from_arrays(lat, lon, orog_m, lsm, sdor, hv, lv)"""
35
+ def __init__(self, static_nc=None):
36
+ if static_nc is None: return
37
+ st = xr.open_dataset(static_nc).load()
38
+ self.lat_c = st.latitude.values.astype(np.float64); self.lon_c = st.longitude.values.astype(np.float64)
39
+ self.orog = (st["geopotential_at_surface"].values / 9.80665).astype(np.float32)
40
+ self.lsm = st["land_sea_mask"].values.astype(np.float32); self.sdor = st["standard_deviation_of_orography"].values.astype(np.float32)
41
+ self.hv = st["high_vegetation_cover"].values.astype(np.float32); self.lv = st["low_vegetation_cover"].values.astype(np.float32)
42
+ @classmethod
43
+ def from_arrays(cls, lat, lon, orog_m, lsm, sdor, hv, lv):
44
+ g = cls(); g.lat_c = np.asarray(lat, np.float64); g.lon_c = np.asarray(lon, np.float64)
45
+ g.orog, g.lsm, g.sdor, g.hv, g.lv = [np.asarray(x, np.float32) for x in (orog_m, lsm, sdor, hv, lv)]
46
+ return g
47
+ def points(self, lat, lon, elev):
48
+ lat = np.asarray(lat, np.float64); lon = np.asarray(lon, np.float64)
49
+ fy = (self.lat_c[0] - lat) / 0.25; fx = (lon - self.lon_c[0]) / 0.25
50
+ iy0 = np.clip(np.floor(fy).astype(int), 1, len(self.lat_c) - 3); ix0 = np.clip(np.floor(fx).astype(int), 1, len(self.lon_c) - 3)
51
+ dy = np.clip(fy - iy0, 0, 1); dx = np.clip(fx - ix0, 0, 1)
52
+ return dict(iy0=iy0, ix0=ix0, dy=dy, dx=dx, lat=lat, lon=lon, elev=np.asarray(elev, np.float64))
53
+
54
+ # ---------------------------------------------------------------- feature helpers (same as V2)
55
+ def bil(f, p):
56
+ iy, ix, dy, dx = p["iy0"], p["ix0"], p["dy"], p["dx"]
57
+ g = lambda y, x: f[..., y, x]
58
+ return (g(iy, ix) * (1 - dy) * (1 - dx) + g(iy, ix + 1) * (1 - dy) * dx + g(iy + 1, ix) * dy * (1 - dx) + g(iy + 1, ix + 1) * dy * dx)
59
+
60
+ def shift(a, k):
61
+ out = np.full_like(a, np.nan)
62
+ if k > 0: out[:-k] = a[k:]
63
+ elif k < 0: out[-k:] = a[:k]
64
+ else: out[:] = a
65
+ return out
66
+
67
+ def roll_mean(a, lo, hi):
68
+ acc = np.zeros_like(a, dtype=np.float32); n = np.zeros_like(a, dtype=np.float32)
69
+ for k in range(lo, hi + 1):
70
+ s = shift(a, k); m = np.isfinite(s); acc += np.where(m, s, 0); n += m
71
+ return acc / np.maximum(n, 1)
72
+
73
+ def solar_elev(times, lat, lon):
74
+ doy = times.dayofyear.values.astype(np.float64)[:, None]; hr = (times.hour.values + 0.0)[:, None]
75
+ g = 2 * np.pi / 365.0 * (doy - 1 + (hr - 12) / 24)
76
+ decl = 0.006918 - 0.399912 * np.cos(g) + 0.070257 * np.sin(g) - 0.006758 * np.cos(2 * g) + 0.000907 * np.sin(2 * g)
77
+ eqt = 229.18 * (0.000075 + 0.001868 * np.cos(g) - 0.032077 * np.sin(g) - 0.014615 * np.cos(2 * g) - 0.040849 * np.sin(2 * g))
78
+ tst = hr * 60 + eqt + 4 * lon[None]; ha = np.deg2rad(tst / 4 - 180)
79
+ la = np.deg2rad(lat[None])
80
+ sin_e = np.sin(la) * np.sin(decl) + np.cos(la) * np.cos(decl) * np.cos(ha)
81
+ return np.rad2deg(np.arcsin(np.clip(sin_e, -1, 1)))
82
+
83
+ class StaticSet:
84
+ """per-station static arrays derived from the station table + ERA5 static grid"""
85
+ def __init__(self, st, grid):
86
+ self.st = st; self.grid = grid
87
+ self.p = grid.points(st.lat.values, st.lon.values, st.elevation_m.values)
88
+ self.terr = {v: st["t_" + v].values.astype(np.float32) for v in TERRAIN_VARS}
89
+ self.r10 = st.t_elev_mean_r10.values.astype(np.float32); self.r40 = st.t_elev_mean_r40.values.astype(np.float32)
90
+
91
+ def build_features(times, E, ss, sl=None):
92
+ """times (T,), E (T,H,W) dict covering times, ss StaticSet. returns F (T,S,nf), names, base dict (T,S).
93
+ Contexts use +-24 h shifts: caller must pad by >=24 h and crop."""
94
+ p = ss.p; g = ss.grid; S = len(p["lat"]); T = len(times)
95
+ names, cols = [], []
96
+ def add(n, a):
97
+ names.append(n); cols.append(np.asarray(a, np.float32).reshape(T, S) if np.ndim(a) == 2 else np.broadcast_to(np.asarray(a, np.float32)[None], (T, S)))
98
+ I = {k: bil(v, p) for k, v in E.items()}
99
+ t2, d2, u10, v10 = I["2m_temperature_c"], I["2m_dewpoint_temperature_c"], I["10m_u_component_of_wind"], I["10m_v_component_of_wind"]
100
+ u100, v100 = I["100m_u_component_of_wind"], I["100m_v_component_of_wind"]
101
+ tp = np.maximum(I["total_precipitation"], 0); ssrd = I["surface_solar_radiation_downwards"]
102
+ rh = rh_from(t2, d2); spd10 = np.hypot(u10, v10); spd100 = np.hypot(u100, v100)
103
+ elev = p["elev"]; orog_i = bil(g.orog, p); dz = (elev - orog_i).astype(np.float32)
104
+ base = dict(t2m=t2, u10=u10, v10=v10, rh=rh, tp=tp, spd10=spd10, d2m=d2)
105
+ for n, a in [("t2m", t2), ("d2m", d2), ("u10", u10), ("v10", v10), ("u100", u100), ("v100", v100), ("spd10", spd10), ("spd100", spd100),
106
+ ("rh", rh), ("dd", t2 - d2), ("sp", I["surface_pressure_hpa"]), ("msl", I["mean_sea_level_pressure_hpa"]), ("tp", tp), ("ssrd", ssrd / 3600.0),
107
+ ("fdir", I["total_sky_direct_solar_radiation_at_surface"] / 3600.0), ("tcc", I["total_cloud_cover"]), ("lcc", I["low_cloud_cover"]),
108
+ ("mcc", I["medium_cloud_cover"]), ("hcc", I["high_cloud_cover"]), ("tcwv", I["total_column_water_vapour"]),
109
+ ("cape", np.sqrt(np.maximum(I["convective_available_potential_energy"], 0))), ("skt", I["skin_temperature_c"]), ("skt_t2", I["skin_temperature_c"] - t2),
110
+ ("gust", I["instantaneous_10m_wind_gust"]),
111
+ ("shear", spd100 - spd10), ("t2m_lapse", t2 - 0.0065 * dz[None])]:
112
+ add(n, a)
113
+ add("wdir_s", u10 / np.maximum(spd10, 1e-3)); add("wdir_c", v10 / np.maximum(spd10, 1e-3))
114
+ for n, a in [("t2m", t2), ("skt", I["skin_temperature_c"]), ("u10", u10), ("v10", v10), ("spd10", spd10), ("tcc", I["total_cloud_cover"]), ("rh", rh), ("ssrd", ssrd / 3600.0), ("tp", tp)]:
115
+ ks = (-3, -1, 1, 3) if n not in ("t2m", "rh") else (-6, -3, -1, 1, 3, 6)
116
+ for k in ks:
117
+ s = shift(a, k); add(f"{n}_d{k:+d}", s - a if n not in ("tp", "ssrd", "tcc") else s)
118
+ add("t2m_mean24", roll_mean(t2, -12, 12) - t2); add("t2m_prev24", roll_mean(t2, -24, -1) - t2)
119
+ tpa = tp
120
+ add("tp_sum_p3", sum(shift(tpa, -k) for k in (1, 2, 3))); add("tp_sum_f3", sum(shift(tpa, k) for k in (1, 2, 3))); add("tp_sum_p12", roll_mean(tpa, -12, -1) * 12)
121
+ add("tcc_mean6", roll_mean(I["total_cloud_cover"], -3, 3)); add("ssrd_mean6", roll_mean(ssrd / 3600.0, -3, 3))
122
+ tm = np.stack([shift(t2, k) for k in range(-12, 13)], 0); add("t2m_max24", np.nanmax(tm, 0) - t2); add("t2m_min24", np.nanmin(tm, 0) - t2); del tm
123
+ iy, ix = p["iy0"], p["ix0"]
124
+ def nb(f, fn):
125
+ patch = np.stack([f[:, iy + a, ix + b] for a in (0, 1) for b in (0, 1)], 0); return fn(patch)
126
+ tpf = np.maximum(E["total_precipitation"], 0)
127
+ add("tp_nbmax", nb(tpf, lambda a: a.max(0))); add("tp_nbmin", nb(tpf, lambda a: a.min(0)))
128
+ wide = lambda f, fn: fn(np.stack([f[:, np.clip(iy + a, 0, f.shape[1] - 1), np.clip(ix + b, 0, f.shape[2] - 1)] for a in (-1, 0, 1, 2) for b in (-1, 0, 1, 2)], 0))
129
+ add("tp_wmean", wide(tpf, lambda a: a.mean(0))); add("tp_wmax", wide(tpf, lambda a: a.max(0)))
130
+ add("cape_wmax", wide(np.sqrt(np.maximum(E["convective_available_potential_energy"], 0)), lambda a: a.max(0)))
131
+ t2f = E["2m_temperature_c"]; add("t2m_nbspread", nb(t2f, lambda a: a.max(0) - a.min(0)))
132
+ add("t2m_gy", bil(np.gradient(t2f, axis=1), p)); add("t2m_gx", bil(np.gradient(t2f, axis=2), p))
133
+ add("tcc_wmean", wide(E["total_cloud_cover"], lambda a: a.mean(0)))
134
+ add("elev", elev); add("dz", dz); add("orog_era", orog_i); add("lsm", bil(g.lsm, p)); add("sdor", bil(g.sdor, p)); add("hv", bil(g.hv, p)); add("lv", bil(g.lv, p))
135
+ for v in TERRAIN_VARS: add("terr_" + v, ss.terr[v])
136
+ add("elev_mean_r10", ss.r10); add("elev_mean_r40", ss.r40)
137
+ hh = times.hour.values[:, None] * np.ones((1, S)); doy = times.dayofyear.values[:, None] * np.ones((1, S))
138
+ add("hr_s", np.sin(2 * np.pi * hh / 24)); add("hr_c", np.cos(2 * np.pi * hh / 24)); add("doy_s", np.sin(2 * np.pi * doy / 365.25)); add("doy_c", np.cos(2 * np.pi * doy / 365.25))
139
+ add("hr_utc", hh)
140
+ sel = solar_elev(times, p["lat"], p["lon"]); add("sun_elev", sel)
141
+ asp_s = ss.terr["aspect_sin"][None]; asp_c = ss.terr["aspect_cos"][None]; sl_ = ss.terr["slope_deg"][None]
142
+ add("down_slope_flow", (u10 * asp_s + v10 * asp_c) * sl_)
143
+ add("flow_x_tpi2k", spd10 * ss.terr["tpi_2km"][None]); add("flow_x_tpi10k", spd10 * ss.terr["tpi_10km"][None])
144
+ add("night_x_tpi2k", np.maximum(-sel, 0) * ss.terr["tpi_2km"][None]); add("night_x_dz", np.maximum(-sel, 0) / 90 * dz[None])
145
+ n_v2 = len(names)
146
+ # ------------- Sun*+ 追加特徴 (全国版) -------------
147
+ st = ss.st
148
+ for n, col in [("lat", "lat"), ("lon", "lon"), ("region", "region"), ("t_elev_max_r10", "t_elev_max_r10"), ("t_elev_min_r40", "t_elev_min_r40"),
149
+ ("t_sea_frac_20km", "t_sea_frac_20km"), ("t_sea_frac_50km", "t_sea_frac_50km"), ("t_sea_frac_150km", "t_sea_frac_150km")]:
150
+ add("x_" + n, st[col].values)
151
+ add("x_dem_minus_stn", (st.t_dem_pt_m.values - st.elevation_m.values))
152
+ # 地形 x 気象 の追加相互作用 (沿岸・斜面)
153
+ cd = ss.terr["coast_dist_km"][None]
154
+ add("x_coast_x_sun", np.clip(cd, -20, 20) * np.maximum(sel, 0) / 90) # 日射と海岸距離 (海風)
155
+ add("x_coast_x_spd", np.exp(-np.abs(cd) / 10) * spd10) # 海岸付近の風
156
+ add("x_t2m_dem", t2 - 0.0065 * (st.t_elevation_m.values - orog_i)[None]) # DEM標高での気温補正
157
+ add("x_valley_night", np.maximum(-sel, 0) / 90 * ss.terr["tpi_10km"][None])
158
+ add("x_cold_pool", np.maximum(-sel, 0) / 90 * (st.t_elev_min_r40.values[None]) * (1.0 / (1 + spd10)))
159
+ add("x_abs_lat_x_doy", np.abs(st.lat.values[None] - 36) * np.cos(2 * np.pi * (doy - 15) / 365.25))
160
+ F = np.stack(cols, 2).astype(np.float32)
161
+ return F, names, base, n_v2
sunstar_plus/model.py ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Sun*+ 推論ラッパ。
2
+ SunstarPlus.downscale(...) : 観測なしダウンスケーリング (リード0, ERA5/再解析/NWP を地点へ)
3
+ SunstarPlus.forecast(...) : 観測同化 (時刻 t0 までの AMeDAS 観測 + 近傍局, リード 1/3/6/12/24h)
4
+ 入力 fields: ERA5 互換の 0.25° 格子場 dict {変数名: (T,H,W)}。単位 = tp [mm/h], 気温/露点/skin [°C], 気圧 [hPa], 日射 [J/m2/h] (ERA5 の生単位)。
5
+ 前後 >=24 h (観測特徴は >=200 h) の余白時刻を含めること。"""
6
+ import gzip, json, pickle
7
+ from pathlib import Path
8
+ import numpy as np, pandas as pd, lightgbm as lgb
9
+ from .features import *
10
+ from .obsfeats import obsfeats, neighbor_table, nb_mean, NB_NAMES, DB_KEYS
11
+
12
+ VARS = ["temp_c", "wind_u_ms", "wind_v_ms", "humidity_pct", "wind_speed_ms", "precip_mm"]
13
+ LABEL = {"temp_c": "気温 [°C]", "wind_u_ms": "風 u [m/s]", "wind_v_ms": "風 v [m/s]", "humidity_pct": "相対湿度 [%]", "wind_speed_ms": "風速 [m/s]", "precip_mm": "降水量 [mm/h]"}
14
+
15
+ def _post(var, x):
16
+ if var == "humidity_pct": return np.clip(x, 0, 100)
17
+ if var in ("precip_mm", "wind_speed_ms"): return np.maximum(x, 0)
18
+ return x
19
+
20
+ class SunstarPlus:
21
+ def __init__(self, root="."):
22
+ self.root = Path(root); mdir = self.root / "models"
23
+ self.cfg = json.load(open(mdir / "config.json"))
24
+ self.models = {t: {v: [self._load(mdir / f) for f in self.cfg["models"][t][v]] for v in VARS} for t in "AB"}
25
+ self.sigma = {}
26
+ for t in "AB":
27
+ for v, d in self.cfg.get("sigma", {}).get(t, {}).items():
28
+ self.sigma[(t, v)] = dict(gbm=self._load(mdir / d["file"]), k=d["k"], ratio={int(k): x for k, x in d["ratio"].items()})
29
+ @staticmethod
30
+ def _load(p):
31
+ p = Path(p)
32
+ return lgb.Booster(model_str=gzip.open(p, "rt").read()) if p.suffix == ".gz" else lgb.Booster(model_file=str(p))
33
+ # ------------------------------------------------------------------ helpers
34
+ def _ens(self, models, X): return np.mean([m.predict(X) for m in models], 0)
35
+ def _predict(self, task, X, base, lead, sigma, heldout):
36
+ out = {}
37
+ for v_i, var in enumerate(VARS):
38
+ r = self._ens(self.models[task][var], X); mean = _post(var, base[:, v_i] + r)
39
+ sg = None
40
+ if sigma and (task, var) in self.sigma:
41
+ d = self.sigma[(task, var)]; raw = np.maximum(d["gbm"].predict(np.concatenate([X, r[:, None]], 1)), 0.05) * np.sqrt(np.pi / 2) * d["k"]
42
+ if heldout: raw = raw * d["ratio"].get(lead, 1.0)
43
+ sg = np.maximum(raw, 0.05)
44
+ out[var] = (mean, sg)
45
+ return out
46
+ # ------------------------------------------------------------------ public
47
+ def downscale(self, times, fields, grid, stations, sigma=True, heldout=True):
48
+ """times: DatetimeIndex (hourly UTC, 前後24h余白込み); fields: ERA5互換; grid: Grid; stations: DataFrame (stations_nationwide.csv 形式)
49
+ returns {var: (mean (T,S), sigma (T,S) or None)} (余白 24h は NaN になりやすいので使わないこと)
50
+ heldout=True: 学習に使っていない地点向けに σ を拡大 (既知地点では False)"""
51
+ ss = StaticSet(stations, grid); F, names, base, n_v2 = build_features(times, fields, ss)
52
+ T, S, nf = F.shape; flag = (np.asarray(times) >= np.datetime64("2024-10-01")).astype(np.float32)
53
+ X = np.concatenate([F.reshape(T * S, nf), np.repeat(flag, S)[:, None]], 1)
54
+ B = np.stack([base["t2m"], base["u10"], base["v10"], base["rh"], base["spd10"], np.zeros_like(base["tp"])], 2).reshape(T * S, 6)
55
+ res = self._predict("A", X, B, 0, sigma, heldout)
56
+ return {v: (m.reshape(T, S), None if s is None else s.reshape(T, S)) for v, (m, s) in res.items()}
57
+ def forecast(self, times, fields, grid, stations, obs, lead, neighbors=True, sigma=True, heldout=True):
58
+ """obs: (T,S,7) 時刻 times に揃えた AMeDAS 実測 [気温,u,v,相対湿度,降水,風速,気圧(NaN可)] (欠測は NaN)。
59
+ 初期時刻 t0=times[i] の観測から times[i+lead] を予測。returns {var: (mean (T,S) 有効時刻 index, sigma)}。 t0 以降の観測は使わない。
60
+ neighbors=True: 近傍局 (obs が有る他地点) の残差特徴を使う。stations に列 'has_obs' (bool) があればそれを近傍プールとする。"""
61
+ ss = StaticSet(stations, grid); F, names, base, n_v2 = build_features(times, fields, ss)
62
+ T, S, nf = F.shape; assert lead in (1, 3, 6, 12, 24)
63
+ bb = dict(t2m=base["t2m"], d2m=base["d2m"], u10=base["u10"], v10=base["v10"], rh=base["rh"], spd10=base["spd10"])
64
+ G, gnames, res = obsfeats(np.asarray(obs, np.float32), bb)
65
+ dbase = np.stack([bb[k] for k in DB_KEYS], 2)
66
+ pool = stations["has_obs"].values.astype(bool) if "has_obs" in stations else np.isfinite(obs[:, :, 0]).any(0)
67
+ if neighbors:
68
+ idx, w = neighbor_table(stations, pool)
69
+ pk = obs[:, :, 4]; rr = [res["T"], res["U"], res["V"], res["RH"], res["SPD"]]
70
+ nbf = []; wsum = None
71
+ for r_ in rr:
72
+ m0, den = nb_mean(r_, idx, w); nbf.append(m0); wsum = den if wsum is None else wsum
73
+ nb6 = [nb_mean(roll_mean(r_, -5, 0), idx, w)[0] for r_ in rr]
74
+ P0, _ = nb_mean(pk, idx, w); P3, _ = nb_mean(roll_mean(pk, -2, 0), idx, w)
75
+ NB = np.stack(nbf + nb6 + [P0, P3, P3, wsum], 2)
76
+ else: NB = np.full((T, S, len(NB_NAMES)), np.nan, np.float32)
77
+ ti = np.arange(lead, T); t0 = ti - lead
78
+ flag = (np.asarray(times)[ti] >= np.datetime64("2024-10-01")).astype(np.float32)
79
+ parts = [F[ti].reshape(-1, nf), np.repeat(flag, S)[:, None], G[t0].reshape(len(ti) * S, -1),
80
+ (dbase[ti] - dbase[t0]).reshape(len(ti) * S, -1), np.full((len(ti) * S, 1), lead, np.float32), NB[t0].reshape(len(ti) * S, -1)]
81
+ X = np.concatenate(parts, 1)
82
+ B = np.stack([base["t2m"][ti], base["u10"][ti], base["v10"][ti], base["rh"][ti], base["spd10"][ti], np.zeros_like(base["tp"][ti])], 2).reshape(-1, 6)
83
+ res_ = self._predict("B", X, B, lead, sigma, heldout)
84
+ out = {}
85
+ for v, (m, s) in res_.items():
86
+ full = np.full((T, S), np.nan, np.float32); full[ti] = m.reshape(len(ti), S)
87
+ sf = None
88
+ if s is not None: sf = np.full((T, S), np.nan, np.float32); sf[ti] = s.reshape(len(ti), S)
89
+ out[v] = (full, sf)
90
+ return out
sunstar_plus/obsfeats.py ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """チャンク(10日)単位の特徴行列生成 (学習行サンプリング / 評価ストリーミング共通)"""
2
+ import numpy as np, pandas as pd
3
+ from .features import *
4
+
5
+ LEADS_B = np.array([1, 3, 6, 12, 24])
6
+
7
+
8
+ def obsfeats(O, base):
9
+ """V2 と同一 (feats.obsfeats): 観測履歴特徴 G (T,S,ng)。O: (T,S,7) [temp,u,v,rh,precip,speed,pressure]"""
10
+ T_, U_, V_, RH_, P_, SP_, PS_ = [O[:, :, i] for i in range(7)]
11
+ spd = np.hypot(U_, V_)
12
+ res = {"T": T_ - base["t2m"], "U": U_ - base["u10"], "V": V_ - base["v10"], "RH": RH_ - base["rh"], "SPD": spd - base["spd10"]}
13
+ obsv = {"T": T_, "U": U_, "V": V_, "RH": RH_, "SPD": spd, "P": P_}
14
+ names, cols = [], []
15
+ for k, a in obsv.items():
16
+ names.append(f"o_{k}0"); cols.append(a)
17
+ for lag in (1, 2, 3, 6, 12, 24): names.append(f"o_{k}_l{lag}"); cols.append(shift(a, -lag))
18
+ for k, a in res.items():
19
+ names.append(f"r_{k}0"); cols.append(a)
20
+ for lag in (1, 2, 3, 6, 12, 24): names.append(f"r_{k}_l{lag}"); cols.append(shift(a, -lag))
21
+ for win in (3, 6, 24, 72, 168): names.append(f"r_{k}_m{win}"); cols.append(roll_mean(a, -win + 1, 0))
22
+ names.append(f"r_{k}_sd24"); cols.append(np.sqrt(np.maximum(roll_mean(a ** 2, -23, 0) - roll_mean(a, -23, 0) ** 2, 0)))
23
+ names.append("o_T_trend3"); cols.append(T_ - shift(T_, -3))
24
+ names.append("o_P_sum3"); cols.append(sum(np.nan_to_num(shift(P_, -k)) for k in range(0, 3)))
25
+ names.append("o_P_sum24"); cols.append(sum(np.nan_to_num(shift(P_, -k)) for k in range(0, 24)))
26
+ G = np.stack([np.asarray(c, np.float32) for c in cols], 2)
27
+ return G, names, res
28
+
29
+ NB_NAMES = ["nb_T0", "nb_U0", "nb_V0", "nb_RH0", "nb_SPD0", "nb_T_m6", "nb_U_m6", "nb_V_m6", "nb_RH_m6", "nb_SPD_m6", "nb_P0", "nb_P_m3", "nb_Pmax3", "nb_w"]
30
+ DB_KEYS = ("t2m", "u10", "v10", "rh", "spd10", "d2m")
31
+
32
+ def hav_km(lat1, lon1, lat2, lon2):
33
+ p = np.pi / 180; a = np.sin((lat2 - lat1) * p / 2) ** 2 + np.cos(lat1 * p) * np.cos(lat2 * p) * np.sin((lon2 - lon1) * p / 2) ** 2
34
+ return 12742 * np.arcsin(np.sqrt(a))
35
+
36
+ def neighbor_table(st, pool_mask, K=8, L=30.0, Z=250.0, rmax=90.0):
37
+ """各地点の近傍 K 局 (pool の他局): index (S,K), weight (S,K). 自局は除外。"""
38
+ S = len(st); lat, lon, el = st.lat.values, st.lon.values, st.elevation_m.values
39
+ pool = np.where(pool_mask)[0]; idx = np.zeros((S, K), np.int32); w = np.zeros((S, K), np.float32)
40
+ for s in range(S):
41
+ d = hav_km(lat[s], lon[s], lat[pool], lon[pool]); d[pool == s] = 1e9
42
+ o = np.argsort(d)[:K]; j = pool[o]; dd = d[o]
43
+ ww = np.exp(-(dd / L) ** 2 - ((el[s] - el[j]) / Z) ** 2); ww[dd > rmax] = 0
44
+ idx[s, :len(j)] = j; w[s, :len(j)] = ww
45
+ return idx, w
46
+
47
+ def nb_mean(a, idx, w):
48
+ """a (T,S) -> weighted nan-mean over neighbours (T,S), and weight sum"""
49
+ v = a[:, idx] # (T,S,K)
50
+ fin = np.isfinite(v); ww = w[None] * fin
51
+ num = np.where(fin, v, 0) * w[None]
52
+ den = ww.sum(2)
53
+ return np.where(den > 1e-3, num.sum(2) / np.maximum(den, 1e-6), np.nan).astype(np.float32), den.astype(np.float32)
54
+
sunstar_plus/terrain.py ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """任意地点の地形・地理特徴量を計算 (stations_nationwide.csv の t_* 列と同じ定義)。
2
+ 標高: AWS Terrain Tiles (terrarium z12) / 海岸距離: global-land-mask / 土地被覆: ESA WorldCover 2021 (COG)
3
+ 要: requests, pillow, scipy, global-land-mask, (任意) rasterio"""
4
+ import io, math
5
+ import numpy as np, requests
6
+ from PIL import Image
7
+ from scipy import ndimage as ndi
8
+ Z = 12
9
+
10
+ def _ll2tile(lon, lat, z=Z):
11
+ n = 2 ** z; return (lon + 180) / 360 * n, (1 - np.arcsinh(np.tan(np.radians(lat))) / np.pi) / 2 * n
12
+
13
+ def _tile(x, y):
14
+ r = requests.get(f"https://s3.amazonaws.com/elevation-tiles-prod/terrarium/{Z}/{x}/{y}.png", timeout=30)
15
+ if r.status_code == 404: return np.zeros((256, 256), np.float32)
16
+ r.raise_for_status()
17
+ a = np.asarray(Image.open(io.BytesIO(r.content)).convert("RGB")).astype(np.float32)
18
+ return a[..., 0] * 256 + a[..., 1] + a[..., 2] / 256 - 32768
19
+
20
+ def _coast(lat, lon):
21
+ from global_land_mask import globe
22
+ la = np.arange(lat - 2, lat + 2.0001, 0.01); lo = np.arange(lon - 2, lon + 2.0001, 0.01)
23
+ LO, LA = np.meshgrid(lo, la); land = globe.is_land(LA, LO)
24
+ ky = 1.11; kx = 1.11 * math.cos(math.radians(lat))
25
+ dl = ndi.distance_transform_edt(land, sampling=(ky, kx)); ds = ndi.distance_transform_edt(~land, sampling=(ky, kx))
26
+ i = int(round(2 / 0.01)); sgn = dl[i, i] if land[i, i] else -ds[i, i]
27
+ yy, xx = np.mgrid[-200:201, -200:201]; dist = np.hypot(yy * ky, xx * kx)
28
+ return dict(t_coast_dist_km=float(sgn), t_sea_frac_20km=float((~land)[dist <= 20].mean()), t_sea_frac_50km=float((~land)[dist <= 50].mean()), t_sea_frac_150km=float((~land)[dist <= 150].mean()))
29
+
30
+ def _landcover(lat, lon):
31
+ import rasterio
32
+ from rasterio.windows import Window
33
+ lat0 = int(math.floor(lat / 3) * 3); lon0 = int(math.floor(lon / 3) * 3)
34
+ url = f"/vsicurl/https://esa-worldcover.s3.eu-central-1.amazonaws.com/v200/2021/map/ESA_WorldCover_10m_2021_v200_N{lat0:02d}E{lon0:03d}_Map.tif"
35
+ with rasterio.Env(GDAL_DISABLE_READDIR_ON_OPEN="EMPTY_DIR"), rasterio.open(url) as src:
36
+ row, col = src.index(lon, lat); a = src.read(1, window=Window(col - 13, row - 13, 26, 26), boundless=True, fill_value=0)
37
+ return {f"t_lc_{n}_frac": float((a == c).mean()) for c, n in {10: "tree", 20: "shrub", 30: "grass", 40: "crop", 50: "built", 80: "water"}.items()}
38
+
39
+ def compute_terrain(lat, lon, elevation_m, landcover=True):
40
+ """returns dict of t_* features for one point. elevation_m: 観測点の標高 (AMeDAS 台帳値など)"""
41
+ R_KM = 12.0; dlat = R_KM / 111.0; dlon = R_KM / (111.0 * math.cos(math.radians(lat)))
42
+ xa, ya = _ll2tile(lon - dlon, lat + dlat); xb, yb = _ll2tile(lon + dlon, lat - dlat)
43
+ xa, xb, ya, yb = int(xa), int(xb), int(ya), int(yb); P = 256
44
+ H = np.zeros(((yb - ya + 1) * P, (xb - xa + 1) * P), np.float32)
45
+ for x in range(xa, xb + 1):
46
+ for y in range(ya, yb + 1): H[(y - ya) * P:(y - ya + 1) * P, (x - xa) * P:(x - xa + 1) * P] = _tile(x, y)
47
+ land = H > 0; Hf = np.where(land, H, 0.0).astype(np.float32)
48
+ px = 40075016.686 * math.cos(math.radians(lat)) / (2 ** Z * 256)
49
+ gy, gx = np.gradient(Hf, px); slope = np.degrees(np.arctan(np.hypot(gx, gy))); asp = np.arctan2(-gx, gy)
50
+ fx, fy = _ll2tile(lon, lat); cx = int(round((fx - xa) * P)); cy = int(round((fy - ya) * P))
51
+ def bm(a, r):
52
+ k = int(round(r / px)); return float(a[max(cy - k, 0):cy + k + 1, max(cx - k, 0):cx + k + 1].mean())
53
+ def box(a, r):
54
+ k = int(r / px); return a[max(cy - k, 0):cy + k + 1, max(cx - k, 0):cx + k + 1]
55
+ o = {}
56
+ o["t_elevation_m"] = bm(Hf, 125); o["t_dem_pt_m"] = float(Hf[cy, cx]); o["t_slope_deg"] = bm(slope, 125)
57
+ o["t_aspect_sin"] = bm(np.sin(asp), 125); o["t_aspect_cos"] = bm(np.cos(asp), 125)
58
+ c = bm(Hf, 125)
59
+ o["t_tpi_500m"] = c - bm(Hf, 500); o["t_tpi_2km"] = c - bm(Hf, 2000); o["t_tpi_10km"] = c - bm(Hf, 10000)
60
+ b1 = box(Hf, 500); b5 = box(Hf, 2500)
61
+ o["t_relief_1km"] = float(b1.max() - b1.min()); o["t_relief_5km"] = float(b5.max() - b5.min())
62
+ o["t_land_mask"] = float(bm(land.astype(np.float32), 125) >= 0.5)
63
+ o["t_elev_mean_r10"] = bm(Hf, 2500) - elevation_m; o["t_elev_mean_r40"] = bm(Hf, 10000) - elevation_m
64
+ o["t_elev_max_r10"] = float(box(Hf, 2500).max()) - elevation_m; o["t_elev_min_r40"] = float(box(Hf, 10000).min()) - elevation_m
65
+ o.update(_coast(lat, lon))
66
+ if landcover:
67
+ try: o.update(_landcover(lat, lon))
68
+ except Exception as e: print("WorldCover 取得失敗 (土地被覆は NaN):", repr(e)[:80]); o.update({f"t_lc_{n}_frac": np.nan for n in ("tree", "shrub", "grass", "crop", "built", "water")})
69
+ return o