Contributions of Multiple UAV Features to Cotton SPAD Estimation from the Perspective of Explainable Machine Learning

Reliable estimation of crop chlorophyll status, a key indicator of photosynthetic activity and nutritional condition, is essential for supporting informed field management decisions. Recently, unmanned aerial vehicle (UAV) remote sensing has attracted considerable attention in crop chlorophyll estimation. However, research on integrating spectral indices (SI) with texture and structural information derived from high-resolution UAV imagery to estimate cotton chlorophyll remains limited, and the relative contributions of these different types of features are still unclear. This study utilized multispectral UAV imagery of cotton during the flowering stage at flight altitudes of 60 m, 80 m, and 100 m, from which the features of 12 SI, eight texture indices (TI), and four structural indices (STI) were derived. The Soil–Plant Analyzer Development (SPAD) provides an indirect yet relatively reliable assessment of leaf chlorophyll status. Accordingly, the Boruta algorithm was subsequently employed to identify variables that contribute significantly to SPAD-based estimation. For each flight altitude, SPAD estimation models were constructed based on three distinct machine learning algorithms. Subsequently, the SHapley Additive exPlanations (SHAP) framework was applied to determine key variables influencing SPAD estimation and to examine how the contributions of the three index types varied across different UAV flight altitudes. The results showed that combining UAV-derived SI, TI, and STI enables accurate estimation of cotton SPAD values. SHAP analysis further revealed the three feature types’ relative contributions to the RF model predictions. Among them, SI had the highest average model-attributed importance (59.36%), followed by STI (23.38%) and TI (17.25%). Moreover, with increasing UAV altitude, the importance of SI gradually increased, with its contribution rising from 58.79% at 60 m to 63.06% at 100 m; in contrast, the contribution of TI showed a decreasing trend, dropping from 20.42% to 12.82%. This study reveals the contributions of spectral, texture, and structural features to cotton SPAD estimation at different UAV flight altitudes, providing a clearer understanding of the relative roles of different feature types in cotton SPAD estimation.