Deep Learning Approach for Accurate Segmentation of Sand Boils in Levee Systems

Abstract

Sand boils can contribute to the liquefaction of a portion of the levee, leading to levee failure. Accurately detecting and segmenting sand boils is crucial for effectively monitoring and maintaining levee systems. This paper presents SandBoilNet, a fully convolutional neural network with skip connections designed for accurate pixel-level classification or semantic segmentation of sand boils from images in levee systems. In this study, we explore the use of transfer learning for fast training and detecting sand boils through semantic segmentation. By utilizing a pretrained CNN model with ResNet50V2 architecture, our algorithm effectively leverages learned features for precise detection. We hypothesize that controlled feature extraction using a deeper pretrained CNN model can selectively generate the most relevant feature maps adapting to the domain, thereby improving performance. Experimental results demonstrate that SandBoilNet outperforms state-of-the-art semantic segmentation methods in accurately detecting sand boils, achieving a Balanced Accuracy (BA) of 85.52%, Macro F1-score (MaF1) of 73.12%, and an Intersection over Union (IoU) of 57.43% specifically for sand boils. This proposed approach represents a novel and effective solution for accurately detecting and segmenting sand boils from levee images toward automating the monitoring and maintenance of levee infrastructure.