Spectral Representation Learning in Color Space for Deforestation Mapping with Limited Supervision

Deforestation poses a severe threat to global ecosystems, contributing significantly to biodiversity loss, climate change, and the degradation of essential ecosystem services. Traditional methods of forest monitoring are often limited in scale, frequency, and accuracy, necessitating more robust and automated solutions. Current research explores the application of image segmentation techniques, particularly deep learning-based methods, such as U-Net and its variants, have improved deforestation detection but still face challenges in handling complex spectral features and achieving high precision. In this paper, we propose a novel color-space representation learning for deforestation mapping by using a U-Net architecture with an EfficientNet-B2 encoder for better feature extraction. Our approach utilizes four channel input combining color-space channel with Near-Infrared (NIR) bands to capture richer spectral information. By employing color-space representation learning with a composite loss function, our method effectively addresses satellite imagery artifacts. Evaluations on Amazon and Atlantic rainforest datasets demonstrate superior performance in only 1% supervision. This work advances remote sensing-based deforestation monitoring by delivering higher accuracy and robustness for real-world applications.