Detection of Natural Features and Objects in Satellite Images by Semantic Segmentation Using Neural Networks

Abstract

In recent years, Neural Networks have become one of the most research focused areas of Artificial Intelligence. From detecting objects in real time to the classification of images, these Neural Networks are efficient and are achieving maximum possible accuracies based on the given inputs. In this work, we use Neural Networks for detecting features in satellite images. Using image segmentation and object detection techniques, we find objects, like roads, buildings, trees, and other resources, in the satellite images. In this work, Neural Network architecture used for segmentation of the images is ConvNet also called Convolutional Neural Network. U-Net which has a convolutional autoencoder architecture maps the layers to find the features and resources in the given satellite images. U-Nets do per-pixel semantic alignment for finding objects and features which result in segregation of resources. By using these, each feature or resource in the satellite image is segmented in different colors with regards to the distinct features allowing us to estimate the resources.