River body extraction using convolutional neural network

Abstract

Hydrologists use various machine learning approaches for segmenting water bodies for carrying out various panning activities. Conventional approaches are mainly clustering or classification based and hence often these approaches are subjective to image properties like texture, colour, pixel intensity value etc. Therefore, fails to achieve acceptable level of accuracy, automation level and ability of generalization. Convolutional Neural Network (CNN) models on the other hand have proven to be robust and successful in achieving state-of-the-art performance in the field of natural image processing. In this work, a Deep Learning framework known as SegNet has been used for extraction of river from digital images. Mean Intersection over Union (MIoU) and the Pixel-wise Accuracy (PA) is been used for performance evaluation purpose. Experimental results show that adopted approach can overcome the problems with conventional approaches and achieve segmentation result with an acceptable level of accuracy.

Subject Classification:

• Information and communication
• Circuits

Keywords:

• Deep Learning
• river
• segmentation
• CNN
• SegNet