A machine learning and remote sensing approach for accurate forest sub-compartment level vegetation cover change monitoring.

Introduction: Accurate detection of vegetation cover type changes in forest sub-compartments (FSCs) is essential for supporting informed forest management decisions. Although various forest change detection algorithms have been developed, fine-scaledetection at the FSC level has received limited attention. Methods: This study addresses this gap by developing an FSC-scale vegetation cover type change detection method that couples spectral and texture information from Sentinel-2 multispectral imagery with forest management planning and design investigation (FMPI) data. Original spectral bands, vegetation indices, and texture features were extracted and used to construct a classification model based on a particle swarm optimization-back propagation neural network (PSO-BPNN). To evaluate performance, the proposed PSO-BPNN method was compared with random forest (RF), support vector machine (SVM), and conventional back-propagation neural network (BPNN) models. Results: Results indicate that PSO-BPNN consistently outperformed the other algorithms in change detection at the FSC scale. Specifically, the method achieved an overall accuracy of 91% for change identification, with a Kappa coefficient of 0.86. In the validation dataset, it successfully detected approximately 80% of the changed FSCs. Discussion: The proposed approach offers a robust and accurate solution for fine-scale forest change monitoring, it enhances the scientific basis for sustainable forest resource management.