MarsDeepNet: A Hybrid Deep Learning Approach for Mars Terrain Classification using Modified GoogLeNet Architecture

This study introduces MarsDeepNet, a hybrid deep-learning method for Martian landscape categorization. Based on GoogLeNet, the last five layers are updated with fifteen additional layers to improve feature extraction. To solve deep networks' vanishing gradient problem, we added a leaky ReLU activation function to the feature map to make the model more expressive. MarsDeepNet performed well on a publically accessible dataset, producing an accuracy of 97.94%, precision of 97.57%, recall of 96.80%, and F1-score of 94.21%. The model outperforms ResNet50, Darknet53, ShuffleNet, GoogLeNet, SqueezeNet, ResNet101, and simple CNN architectures in categorizing Martian terrains. Results show that MarsDeepNet outperforms earlier classification models and provides a solid foundation for planetary exploration research. A dependable tool for alien landscape research, this work advances remote sensing and planetary science by using modern deep-learning algorithms. Following Mars, this work may be used to study other planets and understand their geology. This study represents an advanced step towards the integration of the application of deep learning methods for the categorisation of planetary landscapes and provides a basis for the development of more advanced models/approaches.