Algoritmo de aprendizado profundo para detecção de manchas de óleo e embarcações em imagens de radar de abertura sintética

Abstract

This dissertation discusses the development of an algorithm for detecting oil spills and vessels in Synthetic Aperture Radar (SAR) images, using a modified version of the U-Net neural network originally proposed by Ronneberger, Fischer, and Brox (2015). This work is distinguished by improvements to the U-Net architecture through modifications such as the use of residual blocks, the inclusion of L2 regularization, adaptation of the final convolutional layer, and others, aimed at enhancing the results of reference studies in the field of detection through semantic segmentation of SAR images. The research encompasses a set of methodological processes, divided into five main stages. The first stage involves the acquisition and construction of SAR image datasets, provided by the European Space Agency's (ESA) Sentinel-1 satellites, containing specific targets of oil spills and vessels. The second stage focuses on the processing of these SAR images, a crucial step to ensure the quality and accuracy of the data for subsequent training. The third stage addresses the implementation of modifications to the U-Net architecture, aiming to increase the model's effectiveness in dealing with the complexities inherent in target detection in SAR images. The fourth stage involves the training and validation of the models, conducted individually for each type of target - oil spills and vessels. This approach ensures that the algorithm is optimized for identifying each type of target individually. The final stage is the development of the detection algorithm, which operates sequentially. Initially, it detects oil spills and then vessels, but only in images where oil spill detection has been confirmed. This sequential strategy allows for a more targeted analysis of the object to be detected. The results obtained in the training were promising. The model for oil spill detection showed an Accuracy of 94%, Precision of 86%, and Jaccard Index (IoU) of 67.51%. The model for vessel detection achieved an Accuracy of 99%, Precision of 89%, and IoU of 76.23%. These results surpass some previous reference approaches that used the U-Net architecture, such as KRESTENITIS et al. (2019) and SHABAN et al. (2021), demonstrating the effectiveness of the implemented modifications and the adopted methodology. The study demonstrates the feasibility of combining SAR remote sensing with artificial intelligence to address contemporary environmental challenges. Future studies are suggested to research improvements in the network with other modifications and hyperparameters, expansion of datasets, use of other data augmentation techniques, exploration of other neural network architectures to evaluate their performance in comparison with U-Net, and the integration of change detection methods to identify targets.