Deep Convolution Neural Network Using TMC-2 DEM Images in Chandrayaan-2

Abstract

Abstract The terrestrial planets, including the Moon, Earth, and Mars, have impact craters, contributing significantly to the solar system's complex geomorphology. However, conventional crater identification methods struggle with the accuracy, of their varied shapes, locations, and sizes. Our main aim is to locate lunar craters using Digital Elevation Model (DEM) images from Terrain Mapping Camera-2 (TMC-2) onboard the Chandrayaan-2 mission. Employing a crater-based U-Net model, CNN, Resnet18, and Image Net are utilized for weight training. The custom semantic segmentation network based on the U-Net model proves effective. The methodology involves Canny Edge Detection, pre-trained models, and bounding boxes for crater localization. Fully Convolutional Neural Networks (FCNN) and U-Net are applied to assess and recognize lunar craters in complex scenarios. The proposed model comprises a neural network, feature extractor, and optimization technique for lunar crater detection. The model achieves 80.95% accuracy using unannotated data and precision and recall are much better with annotated data and accuracy 86.91% in object detection with Chandrayaan-2's DEM photos. As we have only considered 2000 images as annotation is a time-consuming process, in the future we will use more image data sets so that our result is comparatively better for this.