Training data development strategy for applying deep learning in remote sensing applications

Abstract

Deep Convolution Neural Networks (DCNNs) are playing a very important role in remote sensing applications. However, one of the major challenges in utilizing DCNNs is, of access to the training data. Firstly, there are very few training data available in various fields such as in natural disaster area, secondly, even if it’s available it may not be suited to the area we are planning to implement. In such a case creating training data by oneself becomes very important. However, we need to understand that there is a big difference between computer vision dataset and remote sensing dataset. As in the latter case, one scene may cover thousands of Kilometers and the total number of scenes are limited. This is why there is a concept of ‘chips’ used in remote sensing domain which means a subset of the satellite scene to be used as an ‘image’ in computer vision sense. This study is comparing the various possible strategies to make the chips from the ALOS-2 scenes and recommending the best after utilizing these chips with popular segmentation network U-Net and SegNet.