Tropical dry forest land use/land cover change detection using semi-supervised deep learning algorithms and remote sensing

Background

Tropical dry forests (TDFs) are among the most threatened biomes and provide critical ecosystem services such as carbon sequestration, water regulation, soil conservation, and biodiversity support. These forests experience strong seasonal variability and fragmentation, making them difficult to classify using remote sensing. In regions like the Cauca River Valley in Colombia, agricultural expansion, urbanization, and logging accelerate degradation. Optical satellite imagery often struggles with cloud cover and phenological variation, while spectral similarities between dry forests and open fields reduce classification accuracy. Synthetic aperture radar (SAR) offers structural information that complements optical imagery and improves monitoring potential.

Goals and Methods

This research develops a semi-supervised deep learning framework to detect land use and land cover change in tropical dry forests. The study integrates optical imagery from Sentinel 2 and PlanetScope with Sentinel 1 radar data using a dual encoder Y Net architecture. Semi-supervised learning combines labeled and pseudo-labeled data generated through clustering algorithms to overcome limited training data. Experiments compare U Net, PSPNet, and Y Net models and analyze temporal changes between 2017 and 2021 across the Antioquia region.

Conclusions and Takeaways

Results of this study show that the semi-supervised Y Net improves classification accuracy and mapping performance, achieving high accuracy while reducing dependence on labeled datasets. The study identifies significant declines in tropical dry forest cover and spatial hotspots of loss. For practitioners, the framework demonstrates how multi-sensor remote sensing and deep learning can support monitoring, conservation planning, and adaptive management in data-scarce dryland ecosystems.