Weakly Supervised Semantic Segmentation With Class Balancing Strategies Applied to Remote Sensing Imagery

• Ano de defesa: 2024
• Autor(a) principal: Meira, Pascoassis Souza Santos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Laboratório Nacional de Computação Científica; Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA); Brasil; LNCC; Programa de pós-graduação em Modelagem Computacional
• Palavras-chave em Português: Aprendizagem supervisionada; Segmentação semântica; Controle remoto de detecção; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://tede.lncc.br/handle/tede/394

Obtaining comprehensive labeled data for remote sensing applications can be costly and time-consuming. So using traditional supervised learning can become infeasible. In this dissertation, we explore Weakly Supervised Learning (WSL) as a compelling alternative, focusing on multi-class semantic segmentation with limited label availability. More specifically, we delve into a Remote Sensing Imagery (RSI) application, leveraging the ISPRS Potsdam dataset as a benchmark for comparing our WSL approach with the traditional fully supervised method. Furthermore, we address the significance of class imbalance challenges inherent to real world RSI and explore the problem of machines learning from them. To mitigate these issues, we investigate newly introduced loss functions, designed specifically for weakly annotated and imbalanced datasets. Additionally, we introduce a novel "seed generation" technique for the proposed WSL system, which makes use of superpixels methodology. Our experiments demonstrate that the proposed WSL systems achieve improvements in both overall accuracy and class-specific F1-Score compared to traditional training with limited labeled data. Our analysis indicates that some widely used methods for handling class imbalance in deep learning are less effective for WSL tasks compared to fully supervised learning.