Objective Leaf Area Index (LAI) is a key parameter for characterizing vegetation canopy structure and physiological function, with important implications for rubber plantations growth monitoring and ecological assessment. Existing remote sensing studies on rubber plantations LAI lack systematic cross-method comparisons under a unified framework. This study aimed to identify the optimal modelling approach for tropical rubber plantations LAI remote sensing retrieval.

Method Focusing on rubber plantations on Hainan Island, five LAI retrieval models were systematically constructed and compared, including Simple Linear Regression (SLR), Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Machine (SVM), and XGBoost, based on 2017 monthly Landsat 8 imagery, and 551 valid ground-measured LAI records, under the same dataset, identical input features, and a fixed random 10-fold cross-validation framework. The input feature set comprised five vegetation indices, one derived difference feature, and two cyclically-encoded monthly temporal variables.

Result XGBoost achieved the best overall accuracy with an R² of 0.9072, closely followed by RF with an R² of 0.9026, both significantly outperforming SVM and linear models. Global feature importance analysis indicated that cyclically-encoded temporal variables contributed more gain than all vegetation indices. Residual diagnostics revealed a systematic underestimation tendency in the high-LAI range across all five models, associated with optical signal saturation under highly closed canopies, with nonlinear models exhibiting substantially smaller bias than linear models. Monthly LAI products generated by the optimal model is highly consistent with existing knowledge in terms of spatial pattern, topographic differentiation, and seasonal variation.

Conclusion Under the unified comparison framework, XGBoost was the optimal approach for rubber plantations LAI retrieval. Systematic underestimation in the high-LAI range represents an inherent limitation of optical remote sensing-based LAI inversion. Future work should quantify the contribution of temporal variables through ablation experiments and explore multi-source data integration to alleviate signal saturation.