GradCAM++ implementation for SegFormer (Landsat dataset)

[samrat-rm]

This PR introduces a prototype implementation of GradCAM++ for model explainability.

The current implementation is limited to the Landsat dataset and the SegFormer model. It generates activation maps for a specified target class and compares them with selected Landsat bands to analyze how the model attends to different inputs.

Current Implementation :

• GradCAM++ is applied to the SegFormer model trained on the Landsat dataset.
• Activation maps are generated for a specified target class.
• The number of batches processed is controlled through the saved_models_run.json configuration.
• The generated heatmap is compared with three fixed bands:
  • cloudy_Radiance_B10_B11_mean
  • clear_Radiance_B10_B11_mean
  • dem
• Results are visualized as a 2×2 grid:
  • GradCAM++ heatmap
  • Each of the selected bands
• The grid images are saved to the configured output directory.

Pictures :

Limitations:

• Development and testing were performed using a subset of the Landsat test dataset (100–500 images).
• Because of this, the evaluation metrics do not match the metrics reported in results.csv, which were computed on the full dataset.
• The limited dataset size may also affect the quality and reliability of the activation maps.

Next Step :

-After reviewing the GradCAM++ implementation:

• Run experiments on the full Landsat test dataset.
• Recompute metrics and compare results withresults.csv.
• Improve the explainability pipeline and visualization outputs.

Discussion / Help Needed

• Issue link: Contains questions and discussions related to the GradCAM++ implementation and design decisions.

AI Disclosure:

• ChatGPT was used to review portions of the code, spell-check, and improve wording in PR and issue descriptions.
• AI was also used to help write some function descriptions and documentation.