Landform Segmentation in Terrain Images Using Image Translation Neural Network Architectures

Satellites have made it easier for scientists to study various aspects of Earth and however, it is still difficult to interpret the complexities of Earth’s terrains accurately and efficiently. One of the key endeavors of satellite imagery is landform classification and terrain analysis which can be performed by the recent developments in areas of machine learning, especially deep learning architectures. Therefore, in this work, we proposed variants of three most widely used Neural Network Architectures: Convolutional Neural Networks (CNN), Autoencoders, and U-Net for mapping terrain maps to landform segmentation maps. The output is validated by using popular metrics like accuracy and structural similarity index (SSIM). A comprehensive comparative analysis of each of these models’ training patterns and how these parameters change over repeated training iterations is also presented. Observations state that the proposed U Net performed well in decreasing the loss and increasing the structural similarity. On the other hand, Convolutional Autoencoder has better accuracy and lesser computational cost compared to U Net and CNN.