Deep Learning for Satellite Picture Classification

Abstract

   Applications such as emergency management, law enforcement, and environmental monitoring all rely heavily on satellite images. For these purposes, human facilitation and object identification are required. The enormous number of potential search areas and the scarcity of available analysts make automation a necessity. Yet, standard approaches to item identification and categorization are limited in their ability to provide a solution due to their emphasis on correctness and precision. Some progress has been made on automating these processes using the supervised neural class of machine learning algorithms. Evidence suggests that neural networks based on convolution may improve image recognition as well as comprehension. Here, we employ them to determine how to distinguish man-made features from natural ones in multispectral, high-resolution satellite images. Using the IARPA Functional World Map (fMoW) dataset, we offer a deep learning method for categorizing features and infrastructure into 63 categories. Visual features and satellite data are integrated using neural networks based on convolution and other artificial neural that form the backbone of the system. Built in Python with the help of the deep learning frameworks Keras and TensorFlow, it is deployed on a Linux server equipped with a Geforce Titan X graphics card. In the fMoWTopCoder competition, the system is now ranked second. Its F1 score is 0.797, and its total accuracy is 83%, with 15 out the of 16 classes properly classified (95% or higher).