Physically interpretable feature learning and model robustness analysis for remote sensing LAI retrieval

Leaf Area Index (LAI) estimation via remote sensing is critical for large-scale vegetation monitoring. To clarify feature utilization, model generalization, and deep-learning interpretability, this study designed two input schemes: “raw reflectance bands” and “reflectance bands + vegetation indices (VIs)”. We systematically compared LAI-estimation performance of two traditional machine-learning models, Random Forest (RF) and Light Gradient Boosting Machine (LightGBM), and two deep-learning models, Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN). Taking CNN as a representative, we analysed how input-feature configurations shape internal feature-learning behavior via feature-correlation analysis, activation-weighted encoding-intensity quantification, and SHapley Additive exPlanations (SHAP) attribution analysis. The results reveal: 1) Using only raw bands, RF and LightGBM suffer severe overfitting without manual feature engineering, whereas MLP and CNN show good stability and relative robustness in this study. 2) The inclusion of VIs mitigates overfitting and improves the training stability of RF and LightGBM. In contrast, MLP and CNN maintain high robustness, demonstrating insensitivity to such feature variations on our dataset. 3) CNN internal feature representations are strongly shaped by input composition. When VIs are provided, internal features correlate strongly with these artificial features; when only raw reflectance is available, they anchor to spectral bands and show weak alignment with external VIs. Under raw-band-only input, the NIR band delivers the largest SHAP predictive contribution for LAI estimation, confirming the essential predictive role of NIR in this context. This study constructed a multi-perspective interpretability framework to disentangle CNN feature-learning behavior under different input settings. It clarifies model boundaries, generalization, and applicable scenarios, supporting model-and-feature selection for LAI estimation and deep-learning interpretability in remote-sensing applications.