Generative Models in Remote Sensing: A Hybrid Approach with Synthetic and Real Images for Landslide Scar Segmentation

• Grant number: 25/04787-2
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: June 01, 2025
• End date: February 28, 2029
• Field of knowledge: Physical Sciences and Mathematics - Geosciences - Geology
• Principal Investigator: Carlos Henrique Grohmann de Carvalho
• Grantee: Lucas Pedrosa Soares
• Host Institution: Instituto de Astronomia, Geofísica e Ciências Atmosféricas (IAG). Universidade de São Paulo (USP). São Paulo , SP, Brazil

Abstract

The year 2023 was the hottest on record, marking a critical turning point in climate change. According to the Emergency Events Database, 399 disasters occurred, resulting in 86,473 deaths and estimated damages of $202.7 billion. In Brazil, heavy rains triggered landslides in São Sebastião (SP), causing 63 deaths and leaving thousands homeless, while in 2024, large-scale floods affected 90% of the municipalities in Rio Grande do Sul. With over 8 million Brazilians living in risk areas, the implementation of effective disaster monitoring approaches, particularly for landslides, has become urgent. In this context, remote sensing combined with artificial intelligence algorithms emerges as a promising strategy. However, the heterogeneity of satellite images and the lack of diverse data sets limit the effectiveness of current models. The generation of synthetic landslide images through generative models presents a feasible solution, allowing the creation of images that closely resemble real ones. Therefore, this project proposes the development of methodologies for generating synthetic images specifically for landslides, aiming to enhance accuracy and improve the generalization ability of semantic segmentation models in different areas. The research is divided into three main objectives: (1) generate and evaluate the quality of synthetic images, (2) develop semantic segmentation models using these data, and (3) explore the use of Visual Language Models (VLMs), which combine textual inputs to improve segmentation interactively. The research results will be made available on a WebGIS platform, integrating the models in an accessible way for users and contributing to the improvement of natural disaster prediction and supporting the formulation of public policies.