Mapping soil properties in tropical rainforest area using UAV-based hyperspectral images and LiDAR points

Abstract Aims This study aimed to integrate UAV-based hyperspectral images and LiDAR points to indirectly estimate the soil nutrient properties in tropical rainforest areas. Methods A total of 175 features, including vegetation indices, texture characteristics, and forest parameters, were extracted from the study area. Five machine learning models, Partial Least Squares Regression (PLSR), Random Forest (RF), AdaBoost, Gradient Boosting Decision Trees (GBDT), and Extreme Gradient Boosting (XGBoost), were constructed to predict soil nutrients. Furthermore, Bayesian optimization algorithm (BOA) was introduced to obtain more optimal model hyperparameters. Results The results showed that BOA can better explain the complex interactions between features and hyperparameters, leading to an average improvement of model performance by 89.38% compared to default parameter models. The GBDT model optimized by BOA outperformed other models in predicting soil pH and TN, with improvements of 512.50% and 36.36%, respectively. The XGBoost model with optimized parameters performed well in predicting SOC and TP (with gains of 206.67% and 95% improvements, respectively). In addition, point cloud features derived from LiDAR data outperformed vegetation indices in predicting soil nutrient properties, enhancing inversion accuracy by effectively characterizing vegetation growth conditions and terrain changes. Conclusions This study indicated that combining the advantages of UAV-based hyperspectral images and LiDAR points will advance the methodology for digitally mapping soil nutrient properties in forested areas, achieving large-scale soil nutrient management and monitoring.