Thermal Infrared Image Georeferencing using an Ensemble of Land Cover Predictions

Abstract

Precise georeferencing of remote sensing data is usually implemented in a post-processing fashion and is a crucial step for Earth observation applications such as change detection, natural hazard management, and ground target tracking. It is particularly important for small satellites intending to perform temperature monitoring and wildfire detection on a global scale wherein the precise location of fire is to be communicated. The cost-efficient navigating systems on board such satellites are often not capable of providing accurate geolocation information directly due to space and power limitations. Therefore, it is very important to have a globally applicable georeferencing refinement framework that is robust against illuminational and time-relevant scene changes. In this paper, we propose a georeferencing framework for thermal infrared images that consists of ensemble matching of deep learning-based land cover predictions to archival, well-georeferenced land cover maps. We verify the proposed framework on the georeferencing of single-band Landsat thermal imagery. Experimental results show the efficiency and practicality of the method with 72% of the test images geolocated within 1-pixel accuracy with no trajectory information available.