A Semantic Segmentation Approach to River Sediment Identification

This study investigates the use of semantic segmentation to identify river sedimentation as part of soil erosion monitoring. Three neural network architectures, namely U-Net, LinkNet, and Feature Pyramid Network (FPN), were evaluated using 2,800 non-overlapping patches generated from 100 images of sediment formations from the IIUM River. After patch extraction, the image-mask pairs were divided into training and validation subsets using a 75:25 ratio, producing 2,100 training patches and 700 validation patches. Based on sediment-class Intersection over Union (IoU) under patch-level validation, U-Net achieved 0.8345, followed by LinkNet at 0.8189 and FPN at 0.2039. U-Net and LinkNet identified relatively uniform sediment formations more successfully, whereas all models struggled with irregular formations or sediment that was visually similar to river water. Because the data split was performed at the patch level, patches originating from the same source photograph may have been present in both the training and validation subsets. Therefore, the reported results should be interpreted as internal patch-level validation rather than as an estimate of generalisation to completely unseen source river images.