A Forest Mapping Model for Algeria Using Noisy Labels and Few Clean Data

This study proposes a forest mapping framework for Algeria that addresses the challenge of limited clean data and noisy global land cover labels. The approach combines a small set of manually curated annotations with noisy ESA WorldCover data, leveraging Sentinel-2 multispectral imagery and Digital Elevation Model (DEM) features such as slope, aspect, and the Normalized Difference Vegetation Index (NDVI). A modified ResNet-18 architecture was fine-tuned using both clean and pseudo-labeled noisy data, enabling the model to effectively mitigate label noise. The framework achieved an overall accuracy of 98.5%, demonstrating strong generalization across Algeria’s diverse forest ecosystems. These results highlight the potential of semi-supervised deep learning to improve large-scale forest monitoring, with applications in conservation, sustainable resource management, and climate change mitigation.