zicohasan/paddy-yield-prediction: Paddy Yield Prediction v1.0.0

Accurate paddy (Oryza sativa) yield prediction is a cornerstone of food security monitoring and precision agricultural management. While satellite-based yield modeling offers broad spatial coverage, it suffers from mixed-pixel limitations in smallholder systems characterized by fragmented, sub-hectare fields. Conversely, Unmanned Aerial Vehicle (UAV) imagery provides ultra-high spatial resolution but lacks broad temporal coverage and regional weather context. This study presents a multi-source data fusion framework that integrates plot-level UAV imagery (from Vegetative, Reproductive, and Ripening phases) with Sentinel-2 satellite time-series and daily weather covariates (NASA POWER) for yield prediction across 514 tropical smallholder plots in Lamongan, Indonesia. We benchmark traditional machine learning algorithms (Support Vector Regression, Random Forest, CatBoost) against the state-of-the-art TabPFN algorithm, a zero-shot transformer pre-trained on tabular data, which we adapt for regression via bin-expected discretization. Under standard 5-fold cross-validation, the UAV-only configuration achieves a high predictive performance, with Random Forest obtaining R2 =0.2126 (MAE=0.1292 kg/m²), followed closely by TabPFN (R2 =0.2086). Fusing UAV features with Sentinel-2 spectral indices and NASA POWER weather covariates maintains competitive performance (Random Forest Fused R2 =0.2114, TabPFN Fused R2 =0.1968). Under Leave-One-Petak-Out spatial block cross-validation, where spatial blocks are held out one-by-one, all traditional models exhibit a significant performance drop. However, TabPFN (Fused) demonstrates exceptional spatial robustness, achieving the best overall spatial validation performance with R2 =0.1150 (MAE=0.1395 kg/m²), outperforming Random Forest (R2 =0.0658) and CatBoost (R2 =0.0350) by a wide margin. These results demonstrate that TabPFN is a powerful, robust, and zero-shot alternative for agricultural modeling under spatial domain shifts, validating the agronomic benefits of multi-temporal and multi-source remote sensing fusion.