Terrain-Dependent Generalization Failure in Multimodal Lunar Crater Classification: A Comparative Study Using Optical and Topographic Data

Shreyas Khobragade 

Automated lunar crater detection is an important task in planetary mapping, geological analysis, and future lunar
exploration. Recent studies have demonstrated the potential of combining optical imagery with topographic information; however, the influence of terrain variability on crater classification performance remains insufficiently explored.
This study presents a comparative evaluation of optical imagery, Digital Terrain Model (DTM) data, and multimodal fusion strategies for lunar crater classification across geologically distinct lunar terrains. A manually curated dataset containing 1,904 image patches was developed from co-registered Lunar Reconnaissance Orbiter Camera (LROC) products covering Mare Tranquillitatis and the Tycho Highlands. Four classification approaches were evaluated using a common convolutional neural network architecture: Optical-only, DTM-only, Early Fusion, and
Late Fusion models.
Experimental results showed that DTM-based classification slightly outperformed optical imagery, while both fusion approaches achieved performance comparable to the DTM model, yielding overall accuracies of approximately 76%. Terrain-wise evaluation revealed a more significant finding. Although all models achieved reasonable crater detection performance within Mare Tranquillitatis, every approach failed to identify crater samples within the Tycho Highlands subset.
Analysis of the curated dataset indicated that severe terraindependent sample imbalance and limited representation of highland crater morphologies exerted a stronger influence on performance than input modality or fusion strategy. These findings highlight the importance of terrain-balanced datasets for robust lunar crater classification and suggest that training data diversity may be as critical as advances in multimodal fusion methodology for achieving reliable cross-terrain generalization