Ter-Net: A Dual-Branch Ensemble Network for Accurate and Interpretable Terrain Type Classification
Shusmita Anjum Aziz, Towhidul Alam, M. S. Rahman, Md Arifur Rahman, B M Taslimul Haque, Tze Hui Liew · 2025
Accurate and automated terrain classification is vital for critical applications including disaster management, autonomous navigation, and environmental monitoring. While deep learning models have shown significant promise, their "black box" nature often limits the interpretability of the mechanism of a model, as well as real-world adoption. This paper introduces Ter-Net (Terrain Net), a novel dual-branch ensemble architecture designed for high-accuracy and interpretable terrain classification. Ter-Net integrates a custom Convolutional Neural Network (CNN) featuring Squeeze-and-Excitation (SE) blocks for fine-grained feature extraction, in parallel with a pre-trained DenseNet121 backbone renowned for its feature reuse. These distinct branches are fused using an ensemble stacking technique and refined by a final classification head. Evaluated on a challenging four-class satellite imagery dataset (Desert, Forest, Mountain, and Plains), Ter-Net achieves state-of-the-art performance, securing a test accuracy and an F1-Score of 97.29%, outperforming established architectures such as ResNet18, MobileNetV3Small, and EfficientNetB0. Furthermore, we employ various Explainable AI (XAI) techniques including Grad-CAM, Saliency Maps and LIME to validate Ter-Net. These visualizations confirm that Ter-Net makes its predictions based on relevant geological and vegetative features, significantly enhancing the model’s transparency and trustworthiness.