Integrating mRMR and PSO–machine learning with SHAP interpretation for UAV-based maize aboveground biomass and carbon estimation

Accurate estimation of crop aboveground biomass (AGB) and aboveground carbon (AGC) across growth stages is pivotal for evaluating agroecosystem carbon sequestration. Nevertheless, the integration of feature selection, model optimization, and interpretability for unmanned aerial vehicle (UAV)-based crop AGB estimation remains underexplored. We propose a comprehensive interpretable machine learning framework for whole-season crop AGB estimation. Maize was selected as the study crop, and multispectral UAV imagery was acquired at key growth stages. The minimum redundancy maximum relevance (mRMR) algorithm was applied to screen spectral bands, vegetation indices (VIs), and texture features (TFs) to formulate various feature subsets. These subsets were coupled with three machine learning algorithms (random forest (RF), extreme gradient boosting (XGBoost), and support vector regression (SVR)) optimized with particle swarm optimization (PSO) to develop stage-specific AGB prediction models. SHapley Additive exPlanations (SHAP) analysis was used to determine the feature contributions. The optimal models were identified for AGB and AGC quantification. PSO-optimized RF with optimal feature combinations achieved the best AGB predictions [jointing: Bands+VIs+TFs; silking/maturity: Bands+VIs+TFs+plant height (PH)]. VIs such as the Normalized Near Infrared Index (NNIRI), the Normalized Difference Red Edge Index (NDRE), the Chlorophyll Index-Red Edge (CIreg), and the Chlorophyll Vegetation Index (CVI), along with the Red_Variance and PH, contributed significantly. AGB spatial distributions were mapped, and AGC was derived using carbon conversion coefficients. This framework provides a promising tool for crop carbon sequestration assessment and sustainable agricultural management.