Development of Artificial Intelligence Models to Assess Flood Exposure of the Paraguay River in Chaco Using Remote Sensing Tools

Abstract

Agriculture and livestock farming are the main economic activities in the Province of Chaco, Argentina. Due to environmental characteristics such as high regional precipitation, flooding directly affects rural populations and agricultural production.
This study focuses on developing artificial intelligence (AI) models to improve the identification of flood-prone areas, using the Department of Bermejo as a case study. Although global surface water maps exist, they are not always reliable at regional scales. Therefore, generating accurate information on surface water is essential for effective flood monitoring and regional planning. We developed remote sensing tools based on optical satellite imagery and implemented three image segmentation models: Support Vector Machine (SVM) and Random Forest (RF), based on traditional machine learning algorithms, and a deep learning model using the U-Net architecture. The goal of this study is to enhance surface water detection using Convolutional Neural Networks (CNNs) that incorporate spatial pixel context, in contrast to traditional models that ignore this aspect. We used Landsat satellite imagery to train and evaluate model performance. The U-Net-based model achieved a 24% increase in the DICE coefficient and a 17% improvement in accuracy compared to the SVM and RF models. These findings confirm that deep learning techniques that account for spatial context significantly improve the identification of surface water.