Interpretable Semantic Image Segmentation Using U-Net and Visual Diagnostics
Students & Supervisors
Student Authors
Supervisors
Abstract
Semantic image segmentation is crucial for the accurate interpretation of geological imagery, facilitating informed decision-making in resource management, environmental monitoring, and geological exploration. Predictive segmentation models using deep learning techniques enable automated analysis and enhance decision-making processes. To precisely segment geological imagery, this research uses a U-Net-based segmentation model that includes a pre-trained convolutional neural network (CNN) encoder (VGG16), enhanced by Batch Normalization and dropout regularization. Intersection over Union (IoU) and total pixel accuracy metrics are used to assess the model's performance. The model achieves an overall pixel accuracy of 96.98% and an overall mean IoU of 90.28%, according to the results. The IoU scores per tile ranged from 88% to 93%, and the Dice Coefficient Score was 89.98%, indicating robust segmentation performance across diverse image tiles. The U-Net-based model, combined with visual diagnostics, including GradCAM heatmaps and pixel-wise error analysis, shows a powerful capability to provide both high accuracy and interpretability. Thus, this method offers valuable information to improve the analysis of geological data, resource allocation, and decision-making in relevant industries.
Keywords
Publication Details
- Type of Publication:
- Conference Name: IEEE 2nd International Conference on Computing, Applications and Systems (COMPAS 2025)
- Date of Conference: 23/10/2025 - 23/10/2025
- Venue: Kushtia, Bangladesh
- Organizer: IEEE Computer Society Bangladesh Chapter