Restoration of Inoperative Terrain Data on Planetary Surfaces Using Deep Generative Models for DEM Reconstruction and Morphometric Analysis Kavitha A. Dr. C.Heltin Genitha
St. Joseph^s college of Engineering, Chennai, TamilNadu, India
Abstract
The ultimate problem of data inoperative terrain ridged planetary surfaces exists in all less illuminated areas, causing combination shadows from the oblique solar angles, topographic occlusion, and objects in the form of rocks, dust clouds, or artificial noise. Such shadows corrupt the Digital Elevation Model and affect thorough surface feature identification and morphometric analysis. The application of deep generative models, with an emphasis on Generative Adversarial Network-(GAN)s and Variational Autoencoder-(VAE)s, for restoring or reconstructing elevation data most times corrupted or missing. By learning elevation propagation models, these executions enhance the quality of Digital Elevation Model-(DEM)s through the Global Accuracy Index (GAI) for surface classification. With the aid of high-resolution DEMs, point cloud data, and multi-sensor imagery, datasets from the Mars and Lunar reconnaissance orbiter missions are being worked on. Pre-processing involves damage detection using histogram analysis and edge filters such as Sobel and Laplacian. Binary masks are created to identify missing zones and serve as conditional inputs for model training. The elevation data is normalized and then co-registered with auxiliary thermal or multispectral imagery. The constraint VAE loss encodes the terrain into latent distributions for probabilistic reconstruction with cold fusion. The GAN-based U-Net branch reconstructs masked DEMs while ensuring realistic outputs through a discriminator. The entire outcome is again compiled into a single DEM through confidence-weighted fusion. Morphometric parameters such as slope, aspect, and curvature are derived from the reconstructed DEMs. The parameters are classified using Support Vector Machines (SVMs) and Deep Convolutional Neural Networks (DCNNs). Model performance evaluation is carried out using RMSE, SSIM, PSNR, and GAI. Model validation is performed against ground-truth and legacy datasets.
