This project was developed as part of my Bachelor’s Thesis (TFG) and focuses on applying semantic segmentation to multispectral remote sensing images using deep learning techniques with PyTorch.
Multispectral images contain information beyond the visible RGB spectrum, capturing additional bands that provide more accurate insight into the characteristics of observed objects.
In this work, several deep neural network architectures—originally designed for RGB inputs—have been successfully adapted to handle these more complex multispectral formats.
After evaluating multiple architectures, we found that lighter models, such as ResNet18dilated + C1 deepsup, not only offer better computational efficiency but also achieve competitive accuracy of 87.74% in semantic segmentation.
These results confirm that, with proper adaptation, existing RGB-based models can be effectively reused in the multispectral domain—sometimes even outperforming heavier alternatives like ResNet101.
From left to right:
Original image, ground truth mask, and prediction generated by the model.
Achieved accuracy: 87.74%
📄 For full technical details and results, see the GID-semantic-segmentation.pdf document.
📘 For installation, usage, or training instructions, go to Appendix B – User Manual inside the PDF.
This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
You are free to use, modify, and share the code for non-commercial purposes, provided you give appropriate credit.