Replacing Yield Detrending with Direct Spatiotemporal Inputs Improves LSTM-Based Rice Yield Estimation

Accurate rice yield estimation is essential for food security. Two key factors affecting estimation accuracy are the long-term upward trend in yield over time and regional heterogeneity across space. Current studies predominantly employ statistical detrending methods (e.g., moving averages, linear regression) to isolate temporal trends. However, such methods rely on prior assumptions about the time–yield relationship and may introduce systematic bias when these assumptions break down. Meanwhile, the individual contributions of temporal and spatial information, and their interactive effects, have not been systematically evaluated within a unified framework. We selected 112 rice-growing counties across six U.S. states (2000–2021), using vegetation index (Normalized Difference Vegetation Index), meteorological indicators (growing degree days, killing degree days, and cumulative precipitation), and spatiotemporal variables (year, longitude, and latitude). We designed six input configurations to compare conventional detrending against direct temporal variable inclusion, testing across four model architectures (Long Short-Term Memory, Random Forest, XGBoost, and Transformer). Results showed that: (1) directly inputting year significantly outperformed detrending across all models, with the combined spatiotemporal configuration achieving the best performance (LSTM R2 = 0.61 vs. 0.54 for detrending); (2) year was the most important predictor in SHAP analysis, with spatiotemporal variables ranking higher than most meteorological and remote sensing variables; (3) spatial information consistently improved accuracy and mitigated systematic bias for extreme yield regions; (4) the combined configuration performed best across different states, years (including extreme climate events), and yield levels, achieving near-end-of-season accuracy at the grain-filling stage (1.5–2 months before harvest). This study demonstrates that integrating raw spatiotemporal data directly into deep learning models is more effective than statistical detrending, offering a simpler and more robust approach for large-scale crop yield estimation.