Deep Learning Approach for Studying Forest Types in Restored Karst Rocky Landscapes: A Case Study of Huajiang, China

Abstract

Forest restoration landscapes are vital for restoring native habitats and enhancing ecosystem resilience. However, field monitoring (lasting months to years) in areas with complex surface habitats affected by karst rocky desertification is time-consuming. To address this, forest structural parameters were introduced, and training samples were optimized by excluding fragmented samples and those with a positive case ratio below 30%. The U-Net instance segmentation model in ArcGIS Pro was then applied to classify five forest restoration landscape types: intact forest, agroforestry, planted forest, unmanaged, and managed naturally regenerated forests. The optimized model achieved a 2% improvement in overall accuracy, with unmanaged and intact forests showing the highest increases (7%). Incorporating tree height and age improved the model’s accuracy by 3.5% and 1.9%, respectively, while biomass reduced it by 2.9%. RGB imagery combined with forest height datasets was most effective for agroforestry and intact forests, RGB imagery with aboveground biomass was optimal for unmanaged naturally regenerated forests, and RGB imagery with forest age was most suitable for managed naturally regenerated forests. These findings provide a practical and efficient method for monitoring forest restoration and offer a scientific basis for sustainable forest management in regions with complex topography and fragile ecosystems.