Well Detection in Satellite Images using Convolutional Neural Networks
Published in Proceedings of the 5th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM), 2019
Abstract
The Government of India conducts a well census every five years. It is time-consuming, costly, and usually incomplete. By using transfer learning-based object detection algorithms, we have built a system for the automatic detection of wells in satellite images. We analyze the performance of three object detection algorithms - Convolutional Neural Network, HaarCascade, and Histogram of Oriented Gradients on the task of well detection and find that the Convolutional Neural Network based YOLOv2 performs best and forms the core of our system. Our current system has a precision value of 0.95 and a recall value of 0.91 on our dataset. The main contribution of our work is to create a novel open-source system for well detection in satellite images and create an associated dataset which will be put in the public domain. A related contribution is the development of a general purpose satellite image annotation system to annotate and validate objects in satellite images. While our focus is on well detection, the system is general purpose and can be used for detection of other objects as well.
Download here