Konuktor92's picture
expected greenness v2-ca
692b68d verified
|
Raw History Blame Contribute Delete
1.89 kB
---
license: other
library_name: xgboost
tags: [nasa, modis, ndvi, imerg, gldas, pastures, central-asia, tabular-regression]
datasets: [TrendeX90/ca-cube]
---
# TrendeX expected greenness, Central Asia (v2-ca)
Summer (May-Sep) NDVI anomaly of 1 km pasture cells in Kazakhstan, Kyrgyzstan, Tajikistan, Turkmenistan and
Uzbekistan, predicted from weather (IMERG precipitation; GLDAS air temperature, soil moisture, snow) and terrain
only. The gap to observed NDVI is change **not explained by weather**: not a cause, not forage quality.
Protocol locked in git before fitting (`model_v2_ca.yaml`, commit 1bc6ed3). Naryn pilot model: TrendeX90/expected-greenness (v1).
| Test | RMSE xgb | RMSE ridge | RMSE climatology | anomaly corr xgb |
|---|---|---|---|---|
| Temporal (2020-2025) | 0.0530 | 0.0546 | 0.0613 | 0.510 |
| Spatial (unseen super-blocks) | 0.0483 | 0.0515 | 0.0592 | 0.578 |
- Skill vs ridge: temporal 0.055 (95% CI 0.036 to 0.077, 6 units); spatial 0.120 (95% CI 0.117 to 0.123, 4115 units).
- Skill vs climatology: temporal 0.251 (95% CI 0.174 to 0.321, 6 units).
- Seeds (3): 0.053 +/- 0.001.
- **Gate (pre-registered, CI lower bound > 0 on both tests): passed = True; shipped = xgb.**
- P10-P90 coverage on 2020-2025: 76% (nominal 80%).
Limits: trained on 150,000 sampled cells (of 3,432,538 eligible) because of budget; GLDAS
(0.25 deg) is coarse; MCD12Q1 2020 mask; NDVI is greenness, not forage; no field validation; budget-limited tuning
(20 trials) and 3 seeds. Spatial test shares years with training and is easier than the temporal test.
```python
from predict_ca import load # numpy, pandas, xgboost; CPU
m = load(".")
m.predict(features_df) # columns as in feature_schema.json
```
Data: MOD13A3.061, MCD12Q1.061, NASADEM_HGT.001 (LP DAAC), GPM_3IMERGM.07, GLDAS_NOAH025_M.2.1 (GES DISC).
AI use: implementation assisted by Claude (Anthropic).