GENERAL DEEP LEARNING SEGMENTATION PROCESS USED IN REMOTE SENSING IMAGES

Abstract

In the present research, we aim at constructing a general segmentation process for different kinds of remote sensing images and various use cases. We focus on the differences in characteristics of the remote sensing and ordinary images, such as irregular shape, lack of labeled images, and normalization issues. The process includes labeling, preprocessing, augmentation, test data sampling, model building, as well as prediction and merging steps. Labeling serves to identify target objects represented in remote sensing images efficiently. The preprocessing step can be applied to reshape an image aiming to fit the requirements of the general artificial intelligence (AI) model and to accelerate steps. Augmentation mitigates the shortage of labeled images. Test data sampling is performed to evaluate the model performance. Finally, prediction and merging are applied to output a full-sized remote sensing image prediction result. In this research, the landslide segmentation, crop farmland segmentation, and cloud segmentation tasks are considered to evaluate the process. Intersection of union (IOU) is employed as evaluation metric. Eventually, we achieve the performance of 72% IOU in the landslide segmentation task, 83% IOU in the crop farmland recognition task, and the 86% IOU in cloud segmentation task by using the proposed process. This supports that the developed process can by further applied considering different remote sensing images and use cases.