UAV-based LiDAR and multispectral sensors fusion for cotton yield estimation: Plant height and leaf chlorophyll content as a bridge linking remote sensing data to yield

Abstract

Accurate crop yield prediction is essential for enhancing agricultural sustainability and guiding economic policy decisions. It is effective to fuse multi-source remote sensing data to predict crop yields, but difficult to reveal the effects of physiological processes on yield estimation models, and challenging to guide crop field production and management. In this study, an innovative framework was introduced to construct plant height (PH) and leaf chlorophyll content (LCC) inversion models for UAV LiDAR and multispectral data through different strategies. PH and LCC, two key growth features affecting cotton yield, were evaluated using multiple linear regression (MLR), partial least squares regression (PLSR), and extreme gradient boosting (XGBoost) algorithms for single-feature and multi-feature fusion, respectively. The multi-feature fusion model based on the XGBoost algorithm was significantly better than the single-feature model (R²=0.744). Further optimization of the multi-feature fusion model revealed that multi-temporal growth features as input variables significantly improved the accuracy of the multi-feature fusion model compared with that based on single-temporal (R²=0.802). Shapley additive explanations (SHAP) analysis revealed the key contribution of LCC to yield formation at the flowering and boll development stage in different cotton varieties. Cluster analysis confirmed that the dynamic trends of PH and LCC were closely related to yield, indicating that PH and LCC could be used as a bridge between remote sensing data and yield. This study highlights the value of UAV-based multi-dimensional and multi-temporal data fusion of growth features in yield estimation models, enabling a deeper understanding of yield formation mechanisms and providing novel methodological tools for phenomics research and precision agriculture management.