Abstract
Tropical dry forests (TDFs) provide essential ecosystem services yet are notoriously difficult to map using remote sensing data due to their spectral similarity to open fields, variability in forest regrowth, and factors such as seasonal leaf phenology and landscape fragmentation that hinder their discrimination. These limitations are critical in data-scarce regions where traditional classification methods struggle. This study introduces a novel semi-supervised deep learning (DL) framework for land use and land cover (LULC) change detection, combining synthetic aperture radar (SAR) and optical satellite imagery for TDF change detection. The proposed framework combines unsupervised pseudo‑labeling and a custom Y‑Net architecture to fuse optical and radar imagery, enabling accurate change detection with limited labeled data. The framework achieves state-of-the-art results, with a mean overall accuracy of 95.3% and a mean Intersection over Union (mIoU) of 88.1%, outperforming established models like standard U-Net and PSPNet. Even in scenarios where only 60% of the dataset is labeled, the semi-supervised method maintains accuracy above 90%, demonstrating its robustness in limited-data conditions. The proposed semi-supervised framework is applied to reveal TDF changes in the Cauca River Valley in Antioquia (Colombia) using satellite images between 2017 and 2021. These findings provide a valuable foundation for advancing remote sensing applications in environmental monitoring, conservation planning, and sustainable resource management.