Métodos de segmentação em regiões com erosão e solo exposto: análise comparativa entre modelos Deep Learning

Abstract

Soil erosion is considered one of the greatest natural risks because it impacts various economic sectors worldwide, causing damage especially in the civil and electrical sectors. The high costs of maintaining infrastructures affected by this type of phenomenon, as well as the blocking of different tasks in areas affected by the appearance of these artifacts, are widely documented in the literature. In this sense, it is essential to identify and monitor the appearance of large-scale erosion to minimize economic losses. Recently, there has been the emergence of a range of works that employ the use of convolutional neural networks (CNNs) and satellite images to monitor different instances on earth. However, the literature still lacks work that makes a more in-depth comparison of different CNN architectures in detecting erosion and exposed soil in satellite images. From this perspective, this work aims to evaluate and compare the performance of four semantic segmentation architectures based on CNNs: U-Net, Mask RCNN, RED-CNN and Googlenet. The experimental results carried out on the reference dataset show that the U-Net and Mask RCNN models achieved the best results, with accuracy values of 98% and 92% respectively. In the IoU metric, these models obtained the highest values with an average value of 44%.