Utilization of the U-Net Convolutional Neural Network and Its Modifications for Segmentation of Tundra Lakes in Satellite Optical Images

Abstract

Tundra lakes are an important indicator of climate change; therefore, the analysis of the dynamics of their size is of particular interest. This paper presents the results of using the U-Net convolutional neural network for tundra lakes segmentation in satellite optical images using Landsat data as an example. The comparative assessment of segmentation accuracy is performed for the original U-Net design and its modifications: U-Net++, Attention U-Net, and R2 U-Net, including with weights derived from a pretrained VGG16 network. The segmentation accuracy is assessed based on the results of manual mapping of tundra lakes in northern Siberia. It is shown that more recent U-Net modifications do not provide a practically significant gain in segmentation accuracy, but increase the computational costs. A configuration based on the classic U-Net gives the best result in most cases (the average Soerens coefficient IoU = 0.88). The technique suggested and the resulting estimates can be used in analysis of modern climate trends.