Abstract
Abstract
Purpose
Rice production is increasingly affected by climate variability, and many rice-growing regions lack the ground-observation infrastructure needed for timely crop-condition assessment. This study evaluated seven machine learning and deep learning models for predicting rice vegetation condition using satellite-derived spectral indices and environmental variables, with the aim of identifying robust modelling approaches and key predictors for scalable, satellite-based agricultural monitoring.
Methods
Landsat-derived spectral indices (NDVI, NDWI, NDMI) were integrated with climate and terrain variables across 12,544 rice field-seasons in Acadia Parish, Louisiana, from 2008 to 2025. Year fixed effects were applied to control for inter-annual confounds, and seven models; linear regression, ridge regression, random forest, gradient boosting, XGBoost, LightGBM, and a deep neural network, were evaluated using five-fold random cross-validation. Variable contributions were assessed using impurity-based and permutation feature importance.
Results
All seven models achieved high predictive accuracy, with Gradient Boosting yielding the highest coefficient of determination (R² = 0.960 ± 0.002). Permutation feature importance analysis showed that the Normalized Difference Water Index (NDWI) was the dominant predictor (ΔR² = 1.708), far exceeding the Normalized Difference Moisture Index (NDMI; ΔR² = 0.034) and all other environmental variables (ΔR² ≤ 0.002).
Conclusion
In flood-irrigated rice systems, surface-water spectral dynamics captured by NDWI accounted for nearly all predictable variation in rice vegetation condition after temporal confounds were controlled, while ensemble tree-based methods provided a marginal but consistent accuracy advantage over deep learning and linear approaches. These findings offer practical guidance for selecting robust GeoAI models and spectral predictors for satellite-based rice monitoring in irrigated agricultural landscapes.