Nepal-Localized Climate-Smart Crop Advisory from Crop-Recommendation, Cereal-Yield, Climate, Boundary, and Soil-Parent Data: A Small-Data Evaluation

Abstract

This paper evaluates a Nepal-localized climate-smart crop advisory workflow using small public datasets that can be inspected and rerun locally. The workflow separates three tasks: crop-suitability classification from a compact soil-weather crop-recommendation table; district-crop cereal yield regression from Nepal cereal statistics, a fiscal-year NASA POWER climate index, district boundaries, and soil-parent features; and low-yield risk screening. The crop-recommendation dataset contained 2,200 records, seven numeric soil-weather attributes, and 22 crop labels. The Nepal panel contained 11,760 clean district-crop-year yield records for 75 districts and five cereals. A 12,053-row daily POWER point series was converted into 33 fiscal-year climate-index records, and a 75-district boundary layer was overlaid with 942 soil-parent polygons to add static terrain and parent-material descriptors. Using seed 42, Gaussian naive Bayes achieved 0.993 test accuracy and 0.993 macro-F1 on the crop-suitability task. The best temporal yield model, a ridge model with climate, district, crop, geo-soil, and crop-climate interaction features, achieved RMSE 727.90 kg ha-1, MAE 506.70 kg ha-1, and R2 0.486 on fiscal years 2010/11 to 2013/14. The low-yield risk task remained difficult, but the ridge risk classifier improved the screening result to F1 0.330, recall 0.592, precision 0.229, and AUC 0.664. The results support a conservative advisory design: strong crop-suitability screening from the clean benchmark dataset, Nepal-wide district-crop yield screening for cereals, and low-yield risk flags that require field review before intervention.