Abstract

Field-scale rice yield prediction in irrigated coastal systems is constrained by small plot size, persistent cloud cover, and strong interannual variability, which limit the effectiveness of single-sensor remote sensing approaches. This study proposes an integrated framework that combines multispectral time series from Sentinel-2, PlanetScope, and unmanned aerial vehicle (UAV) platforms with phenological metrics derived from the Normalized Difference Vegetation Index (NDVI), climate variables aggregated within phenology-defined windows, and machine learning algorithms to predict rice grain yield at the plot level. The framework was structured in five phases: (i) cross-sensor consistency assessment of red, near-infrared, and NDVI values across platform pairs; (ii) linear harmonization and multisource temporal fusion of NDVI time series at 5-day intervals; (iii) extraction of phenological metrics from smoothed NDVI trajectories using a relative-threshold approach; (iv) aggregation of meteorological variables within crop-stage-specific windows; and (v) yield prediction using partial least squares regression (PLSR), Random Forest, and XGBoost under nested leave-one-out cross-validation. The framework was evaluated on 72 irrigated rice plots (37 in 2022, 35 in 2023) in Lambayeque, northern Peru. Cross-sensor analysis revealed that the PlanetScope–UAV pair achieved the strongest NDVI agreement (R2=0.87, RMSE =0.07), while Sentinel-2–PlanetScope showed higher correlation (R2=0.91) but with systematic bias requiring calibration. Multi-source fusion raised temporal coverage from 53–62% (individual sensors) to 82% in 2022 and 66% in 2023. The best single-season prediction was obtained in 2022 with PlanetScope-derived phenological metrics and XGBoost (Rcv2=0.72, RMSEcv=1.23 t ha−1), while the best cross-season performance was achieved with combined phenological and climate features using the PlanetScope+UAV configuration and XGBoost (Rcv2=0.64, RMSEcv=1.35 t ha−1). SHAP-based interpretability analysis identified post-peak phenological descriptors and climatic conditions during the grain-filling window as the most informative predictors. These findings demonstrate that PlanetScope-centered multi-source fusion, combined with phenology-informed feature engineering, provides a robust basis for rice yield prediction in cloud-prone irrigated environments.

IPC Classification

Keywords