A multi-source remote sensing and machine learning framework for maize mapping and yield estimation in fragmented Loess gully regions

Résumé fourni par la source

Accuracy crop distribution mapping and reliable yield estimation are essential for overcoming fragmentation and decentralization in smallholder farming systems of the Loess Plateau gully region. Multi-source remote sensing data, ancillary datasets, and machine learning techniques were integrated to map maize distribution and estimate yield. First, Sentinel-2 temporal spectral features, vegetation indices, and topographic variables were integrated to identify the optimal maize mapping model by a comparing machine learning algorithms: Random Forest (RF), Extra Trees (ET), Gradient Boosting Decision Tree (GBDT), and Histogram-Based Gradient Boosting Decision Tree (HGBDT). Subsequently, Sentinel-2 optical data and ERA5-Land meteorological data were dynamically resampled and spatiotemporally fused. A maize yield estimation model was then developed by integrating these fused predictors with in-situ measured maize yield samples. Finally, SHapley Additive exPlanations (SHAP) analysis was applied to quantify the feature contribution to both crop mapping and yield estimation, improving model transparency and interpretability. The results indicate that RF model achieved superior performance for maize identification in heterogeneous agricultural landscapes, with an overall Accuracy of 0.825, Precision of 0.849, Recall of 0.933, and F1-Score of 0.889. In the multi-source fusion-based yield estimation task, the HGBDT model yielded the highest predictive accuracy, with an R² of 0.6, RMSE of 1.07 t/ha, relative RMSE (rRMSE) of 11.49%, and MAE of 0.86 t/ha. Here, a methodological advancement is presented toward accurate and interpretable crop mapping and yield estimation in the ecologically complex and topographically fragmented Loess Plateau.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.