Deep learning models to map deforestation based on Sentinel 1 coherent features in the southern border of Amazon

Artikel i vetenskaplig tidskrift, 2025

The Amazon is the largest continuous area of tropical forest on Earth, yet it remains under significant threat from deforestation, degradation, wildfires, and the expansion of agriculture, livestock, and illegal mining. While optical and radar-based monitoring systems provide accurate long-term data, such as Land Use and Land Cover (LULC) and deforestation alerts, their effectiveness is largely confined to the dry season, with some requiring extensive manual effort to detect forest disturbances. This study aims to improve LULC and deforestation monitoring by developing deep learning (DL) classifiers using Synthetic Aperture Radar (SAR) coherent features. The models were trained on three distinct Amazonian landscapes, such as flat, undulated, and hilly, through Sentinel-1 data with a 30 m minimum mapping unit and 12-day revisit time. U-Net, Semantic Flow Network (SF-Net), and Long Short-Term Memory (LSTM) architectures were adapted and enhanced with residual learning, dilated convolutions, attention mechanisms, and squeeze-and-excitation blocks, with hyperparameter tuning conducted via the Optuna framework. The model Sentinel 1 scene 54 622/627, U-Net model, and 4 classes reached the highest overall accuracy and intersection over-union (IoU) in order of 0.95 and 0.66, respectively. The less precise mapping was noticed by Sentinel 1 scene 83 617/622, LSTM model, and 4 classes with a global accuracy of 0.61 and IoU of 0.36. The deep learning model that achieved the lowest error was U-Net, with an RMSE of 0.43 and a standard deviation of 0.43, and it was considered a random error. On the contrary, the SF-Net and LSTM showed systematic error, which reached RMSE between 0.38 and 0.83 and a standard deviation between 56.53 and 114.69. The most precise LULC classes were provided by Forest (Fo) and Deforestation (De), which achieved the highest values of F1-Score with 0.97 and 0.92, respectively. On the opposite way, it was the Non-Forest (NF) and Water (Wa) classes that obtained a lower F1-score in order of 0.51 and 0.72, respectively. Taylor and Target diagram analyses indicated that scene 83 617/622 was particularly well-suited for U-Net-based DL modeling, aligning closely with Ground Control Points (GCPs). This research introduces a novel DL approach leveraging Sentinel-1 coherent features for effective LULC mapping across varied terrain in the southern Amazon during the dry season.