DEFORESTATION DETECTION IN SAR IMAGES USING DEEP NEURAL NETWORKS
SAR, optical, data, fusion, convolutional, neural, network, transformer, unet, remote, sensing
Given the limitations of optical satellite images, synthetic aperture radar (SAR) stands out for its resistance to adverse weather conditions. However, accurate recognition of deforested areas in SAR images is required due to speckle noise and object variability. In this work, we carried out an online experiment with volunteer participants who identified deforested areas in SAR images. To achieve this, we developed software that allows participants to annotate the SAR images, delimiting deforested areas. With the results of this experiment, it was possible to analyze the relationship between the participants’ self-declared experience level and the accuracy in detecting deforested areas. We also compare human performance and the performance obtained with an automatic model based on the UNet architecture. The results show that greater knowledge in remote sensing or SAR does not guarantee quality grades. Furthermore, UNet’s performance surpasses the performance obtained with humans on the task. To explore the segmentation of SAR images in more depth, a second experiment was carried out using state-of-the-art models for segmentation on fused SAR and optical data. The study reinforces the potential of deep learning in detecting deforestation, emphasizing the need for continuous improvement of architectures and training of specialists