Random Forest and deep learning approaches for detecting forest cover changes

Abstract

The forests of the Ukrainian Carpathians, particularly within the National Nature Park “Skole Beskids”, face increasing pressures from climate change and anthropogenic activities. Disturbances ranging from extreme weather events to illegal logging undermine forest health and ecosystem stability. Timely, data-driven monitoring is therefore essential for sustainable forest management. This study applies a multi-sensor remote sensing approach to detect forest cover changes in the Skole Beskids between 2023 and 2024, combining Sentinel-1 synthetic aperture radar (SAR) data, Sentinel-2 multispectral imagery, and a Global Digital Elevation Model (DEM). Sentinel-2 provides detailed optical information on vegetation health and canopy cover, while Sentinel-1 enhances structural change detection under cloudy conditions and dense canopy. A comparison was conducted between a traditional machine learning method (Random Forest, RF) and a deep learning approach (UNet with a ResNet34 backbone) for robust classification and semantic segmentation. The deep learning model demonstrated superior performance in boundary delineation and noise reduction, achieving high accuracy on the validation dataset: an Intersection over Union (IoU) of 0.84, an F1-score of 0.91, a Pixel Accuracy of 0.96, an Overall Accuracy of 83.81%, and a Kappa coefficient of 0.676. Analysis of the 2023–2024 period identified approximately 1.247 square km of forest cover change within the park boundaries. The integration of multi-sensor data and AI-based analysis enabled accurate medium-resolution mapping and improved classification reliability. This research demonstrates the potential of an operational, multi-sensor satellite monitoring framework to support forest governance and conservation strategies in the Ukrainian Carpathians.