Dataset Viewer
Auto-converted to Parquet Duplicate
district_id
stringclasses
200 values
year
int64
2.01k
2.03k
yield_kg_rai
float64
120
563
D001
2,008
347.2
D002
2,008
176.2
D003
2,008
392.9
D004
2,008
321.2
D005
2,008
361.3
D006
2,008
219.8
D007
2,008
434.1
D008
2,008
275.6
D009
2,008
369.4
D010
2,008
302.3
D011
2,008
240.3
D012
2,008
268
D013
2,008
376.2
D014
2,008
301.3
D015
2,008
400.9
D016
2,008
333.9
D017
2,008
302.8
D018
2,008
272.7
D019
2,008
270.7
D020
2,008
302.5
D021
2,008
361.3
D022
2,008
360.4
D023
2,008
416.2
D024
2,008
291.7
D025
2,008
378.7
D026
2,008
350.2
D027
2,008
351
D028
2,008
315.6
D029
2,008
330.3
D030
2,008
497
D031
2,008
120
D032
2,008
375.7
D033
2,008
328.9
D034
2,008
315.4
D035
2,008
309
D036
2,008
364
D037
2,008
341.6
D038
2,008
265
D039
2,008
120
D040
2,008
269.1
D041
2,008
293.6
D042
2,008
305.6
D043
2,008
277.7
D044
2,008
363.8
D045
2,008
358.8
D046
2,008
366.2
D047
2,008
263.9
D048
2,008
153.9
D049
2,008
403.8
D050
2,008
338.3
D051
2,008
215.2
D052
2,008
395.4
D053
2,008
463.5
D054
2,008
405.5
D055
2,008
422.2
D056
2,008
298.8
D057
2,008
301
D058
2,008
365.8
D059
2,008
292.1
D060
2,008
341.4
D061
2,008
238.1
D062
2,008
368.3
D063
2,008
308.7
D064
2,008
321.7
D065
2,008
289.3
D066
2,008
382.2
D067
2,008
350.1
D068
2,008
324.6
D069
2,008
382.7
D070
2,008
448.2
D071
2,008
361.9
D072
2,008
300.5
D073
2,008
413.1
D074
2,008
393.1
D075
2,008
305.8
D076
2,008
308.8
D077
2,008
277.7
D078
2,008
307.4
D079
2,008
321.1
D080
2,008
328.4
D081
2,008
440.6
D082
2,008
295.5
D083
2,008
359.8
D084
2,008
317.8
D085
2,008
278.5
D086
2,008
437.2
D087
2,008
305
D088
2,008
279.6
D089
2,008
419.4
D090
2,008
350.3
D091
2,008
297.9
D092
2,008
408.3
D093
2,008
330.9
D094
2,008
309.8
D095
2,008
132.7
D096
2,008
420.5
D097
2,008
439
D098
2,008
445.7
D099
2,008
355.7
D100
2,008
294.7
End of preview. Expand in Data Studio

Isan rice yield benchmark

All data in this dataset are synthetic. They come from a documented data-generating process whose true yield components are stored next to the observations, so forecasts and their explanations can be checked against the truth. Nothing here is an empirical statement about Thai rice yields.

Data for the preprint Leakage-safe, explainable multi-source machine learning for district-level rice yield forecasting: a ground-truth simulation benchmark for Northeast Thailand by Thanatchon Chongmankhong (read online · PDF). Main-season rice in 200 simulated districts of Northeast Thailand (Isan) over 18 seasons, 2008-2025, with drought, flood, heat, disease, salinity, price and competing-crop effects, persistent management and satellite-only pest outbreaks. Code: https://github.com/thanatchon36/isan-rice-yield-benchmark; models: https://huggingface.co/thanatchon36/isan-rice-yield-benchmark; archived on Zenodo: 10.5281/zenodo.24036674.

Independent personal research. Does not represent the views of any employer or institution.

Files

path content
data/yields.csv district-season rice yield, yield_kg_rai (kg per rai; 1 rai = 0.16 ha, multiply by 0.00625 for t/ha); 200 districts x 18 seasons
data/districts.csv static district attributes: province, coordinates, basin, terrain, soil, salinity, irrigation and paddy share (noisy map layers), flood frequency, rainfall normal, distances to river, sugar mill and market
data/climate_monthly.csv May-November weather: rainfall, mean and maximum temperature, relative humidity, soil moisture
data/satellite_monthly.csv May-November NDVI, EVI, LSWI and land-surface temperature; NaN = no cloud-free composite
data/prices_annual.csv 2006-2025 farm-gate prices of rice, cassava, sugarcane and maize, and the urea price
data/reservoir_annual.csv reservoir storage (% of capacity) by basin and year
data/truth_components.csv the ground truth: the additive contribution of every mechanism (base, trend, drought, flood, heat, disease, salinity, price, management, pest, noise) to each district-season yield, in kg/rai
data/meta.json seed, scenario and every parameter of the data-generating process
results/e1_* E1 accuracy: metrics with season-block bootstrap CIs and Holm-adjusted Diebold-Mariano tests, out-of-sample predictions for 2018-2025 with conformal 80%/90% intervals, errors by season, the selected configuration
results/e2_leakage.csv E2: the leakage-safe rolling origin against random 5-fold CV, same-season prices, current neighbour yields and anomalies from the full-period mean
results/e3_ablation.csv E3: change in RMSE when each of the eight feature groups is removed
results/e4_* E4: how well grouped SHAP values recover the true components
results/e5_leadtime.csv E5: in-season forecasts from July to November, with and without satellite features
results/e6_* E6: conformal interval coverage and width, overall and by season
results/e7_robustness.csv E7: the model race repeated in 10 replicate worlds of the base scenario and in 4 alternative scenarios (nonlinear, clouds, map error, no pest)
results/diag_* post-hoc diagnosis of the Ridge failures (row-wise vs season-blocked penalty tuning)
results/feature_groups.csv the 81 features and their groups
SHA256SUMS checksums of every file above

Test seasons are 2018-2025 under a rolling origin: the forecast for season t is trained on seasons before t only. docs/DATA_SCHEMA.md in the code repository gives the schema that real data must follow to replace these files.

Licence

CC BY 4.0.

Citation

@misc{chongmankhong2026isanrice,
  author = {Chongmankhong, Thanatchon},
  title  = {Leakage-safe, explainable multi-source machine learning for district-level rice yield forecasting:
            a ground-truth simulation benchmark for Northeast Thailand},
  year   = {2026},
  doi    = {10.5281/zenodo.24036674},
  url    = {https://thanatchon36.github.io/isan-rice-yield-benchmark/paper/},
  note   = {Preprint; all results are from a synthetic benchmark}
}
Downloads last month
124

Models trained or fine-tuned on thanatchon36/isan-rice-yield-benchmark