Machine Learning Meets Martian Geomorphology: A GPT-4 Approach to Image Alignment

Abstract:
<p class="MsoNormal">The precise alignment of high-resolution Context Camera (CTX) imagery with a resolution of 6 m/px and the lower-resolution Thermal Emission Imaging System (THEMIS) global mosaic at 100 m/px is a persistent issue in Martian geology and geography. Although both datasets use a Mars-centric simple-cylindrical (plate Carrée) projection and come with preset georeferencing, misalignment between individual CTX images and the THEMIS base layer is common. To address this, we harnessed the capabilities of GPT-4 for Python script development tailored to this geospatial challenge.</p><p class="MsoNormal">The first script leverages the Scale-Invariant Feature Transform (SIFT) algorithm to identify keypoints in both image sets, which are then serialized into .npy files for subsequent processing. The second script, optimized for a system equipped with an i7-4930K CPU and an RTX 2080 Super GPU, employs the Facebook AI Similarity Search (FAISS) library to match these keypoints based on their unique descriptors. A third script, currently under development, intends to utilize these matched keypoints to calculate an affine transformation for georeferencing correction in ArcGIS Pro.</p><p class="MsoNormal">A distinguishing facet of this project is the iterative consultation with GPT-4. Through focused Q&amp;A sessions, we honed the chatbot's understanding of the complexities involved in geospatial alignment. This led to the generation of skeleton Python scripts, which were further refined with the help of GPT-4 to include error logging, improved computational efficiency, and contingency plans for potential failures. This approach not only expedites script development but also offers a novel problem-solving paradigm in planetary science and GIScience.</p>

Keywords: Mars, geomorphology, planetary, georeferencing, image alignment, automation, machine learning, GPT-4, GPT, Chat GPT

Authors: