A semantic segmentation approach to river sediment identification

Abstract

Soil erosion is an ecological hazard that, if left unchecked, poses wider threats to the environment. These threats range from inconveniences such as the ruining of landscapes and the reduction of water quality to hazards such as floods and landslides. This thereby necessitates a method to monitor soil erosion, one of which is by monitoring the formation of river sediments. Computer vision techniques have matured in recent years and have been used in many different fields of applications. One form of computer vision technique is called "semantic segmentation," which is a technique that seeks to associate every pixel of an image with its own object class. This presented an opportunity where images of river sedimentation could be analysed and identified accurately to the pixel. In examining further the use of semantic segmentation for river sedimentation purposes, this project looked at three publicly available network architectures: Unet, Linknet, and Feature Pyramid Network (FPN). All these three networks belong to a type of architecture called fully convolutional networks. Three prediction models, one from each architecture, were trained and tested against 100 images of various river sediment formations along the course of the IIUM river. The images are divided into 75 images for training and 25 images for validation. Meanwhile, the model is assessed both quantitatively by Intesection over Union (IoU), and label predictions assessed qualitatively. After training, the sediment IoU scores obtained were as follows: 0.83446103 for Unet, 0.8188789 for Linknet, and 0.20392573 for FPN. The qualitative results outputted however were mixed: the architectures are able to perform somewhat well in identifying sedimentation when the formation of those sediments is uniform, with Unet performing the best, followed by Linknet and then FPN. However, all the architectures struggled in identifying the sediment when non-uniform sedimentation formations are present. One recommendation proposed is to add object classes to reduce intraclass differences and hopefully reduce class confusion by the prediction models. Another recommendation is to develop novel architectures that are able to accommodate intraclass differences while still producing accurate sediment identification.