MoonFuji/algerian_forest_mapping_paper_dataset

# Forest Mapping Dataset — Algeria (Sentinel-2) This repository contains the data used in our paper on AI-powered forest mapping in Algeria (abstract and metadata below). The dataset consists of Sentinel-2 image patches, high-quality manually annotated masks, and noisy WorldCover-derived labels used for training and evaluation of semantic segmentation models. **Abstract** This paper presents an AI-powered approach for mapping Algerian forests using Sentinel-2 satellite imagery and advanced deep learning techniques. We leverage the ESA WorldCover dataset for initial training while addressing its inherent noisy labels through a robust methodological framework. Our approach employs a DeepLabV3+ architecture for forest segmentation, incorporating six spectral bands and vegetation indices (NDVI, EVI, SAVI). To handle uncertain pixel classifications and label noise, we implement a custom composite loss function combining Deep Abstaining Classifier (DAC) Loss with categorical focal and Dice loss components, specifically weighted to prioritize accurate forest class detection. Model validation is performed using a high-quality, manually annotated dataset created via Google Earth Pro imagery. Our methodology achieves superior performance (Accuracy: 95.9%, Dice: 91.43%, IoU: 83.45%, Recall: 95.4%), substantially outperforming baseline U-Net architecture. The model successfully learns to ignore training label noise while producing spatially coherent forest predictions with well-defined boundaries, offering a reliable, scalable method for forest mapping in regions with data quality challenges. **Keywords:** forest mapping; Sentinel-2; deep learning; noisy labels; DeepLabV3+; DAC Loss; semantic segmentation; remote sensing; Algeria **Dataset organization** - `manually_extracted_patches/` — folder with image patches, ground truth, and validation subsets. - `sentinel2_images/` — Sentinel-2 image patches saved as NumPy arrays (`*.npy`). - `clean_masks/` — high-quality manually a …