Deep learning-based object detection and geographic coordinate estimation system for GeoTiff imagery

Abstract

A deep learning-based system has been created to autonomously analyze GeoTiff aerial imagery in order to retrieve information about objects type and their geographic coordinates. This research focuses on applying a Convolutional Neural Network (CNN) to detect objects and estimate the geographic coordinate of airplanes, ships and cars in those images. The system prototype was tested to measure the accuracy and precision for object detection. Furthermore, a Mean Absolute Error (MAE) analysis is done to the system to measure object coordinate estimation performance. The accuracy and precision for object detection of the system prototype are 81,05% and 93,29%, respectively. The system has MAE values which vary from 0,000012° to 0,000034° for object coordinate estimation.