A Transformer-based network for deforestation detection from bitemporal optical satellite images

The expansion of deforestation rates in the Brazilian Amazon has raised attention to forest monitoring initiatives. Currently, the PRODES project employs human specialists to detect new deforestation areas from bitemporal optical satellite images. Previous works showed that convolutional neural networks (CNNs) achieved excellent results to automatically detect deforestation. Recently, models based on Swin Transformers layers outperformed the CNN models in many computer vision tasks. In this paper, we investigated the employment of the Swin Transformer-based model to detect new deforestation areas, compared to a traditional CNN-based network. Swin Transformer-based models increased the F1-Score by 0.055. The code can be found here.