METHODOLOGICAL APPROACHES TO TREE COVER DETECTION AND EVALUATION IN ARID LANDSCAPES

Abstract

This study addresses the problem of automatic tree cover segmentation in desert environments using high-resolution UAV imagery from the Suhaitu Gacha region (Inner Mongolia, China). Two modeling approaches–a statistical ensemble method (Random Forest, RF) and a deep learning-based semantic segmentation model (U-Net)–were comparatively evaluated. The experimental framework incorporated pixel-level classification accuracy, contour delineation quality, detection of small vegetation structures, and overall segmentation stability. Quantitative assessment was conducted using standard metrics, including mean Intersection over Union (mIoU) and the Kappa coefficient. The results demonstrate that the U-Net model consistently outperforms RF, particularly in complex desert landscapes characterized by low spectral contrast between vegetation and background. U-Net provides superior delineation of fine structures and improved segmentation coherence. However, RF exhibits advantages in computational efficiency, training speed, and robustness, confirming its suitability as a lightweight baseline model. These findings highlight the trade-offs between accuracy and efficiency and support the application of advanced computer vision models for ecological monitoring and desert vegetation analysis.