A suite of scripts to process and analyze Murtel Rock Glacier images, focusing on:

• Weather filtering for TIR (thermal infrared) and RGB images
• RGB spatial alignment (SuperGlue-based)
• TIR-RGB overlay or matching

1. Clone this repository

2. Install requirements:

   pip install -r requirements.txt

3. Add SuperGlue Model: Place the models directory from magicleap into aligning directory in order for the rgb aligning to work. Make sure to adhere to their License.

Below are sample commands for each key functionality:

Trains an svm using labeled data:

python weather_filter/rgb_filter_model_training.py \
        ../RGB_images \
        data/filtering/rgb_labels.csv \
        --extracted_features data/filtering/all_extracted_features.csv \
        --train_features data/filtering/train_extracted_features.csv \
        --test_features data/filtering/test_extracted_features.csv \
        --train_csv data/filtering/train.csv \
        --test_csv data/filtering/test.csv \
        --model svm \
        --model_dir data/filtering \
        --re_extract

(This orchestrates feature extraction, splitting, model training and testing.)

To predict unlabeled data:

python weather_filter/rgb_filtering.py \
        data/RGB_DL_images \
        --scaler_path scaler.joblib \
        --model_path model.joblib \
        --output_features data/filtering/predictions.csv

Trains a threshold entropy-based classifier on TIR images:

python weather_filter/tir_filter_train.py \
        data/TIR_images \
        data/filtering/tir_labels.csv \
        --outputs_csv data/filtering/tir_filtering.csv \
        --threshold_file data/filtering/tir_entropy_threshold.joblib

(This calculates entropies, selects and saves decision boundary.)

For classification on unlabeled TIR data:

    python weather_filter/tir_filtering.py \
        data/TIR_images \
        data/filtering/tir_entropy_threshold.joblib

Full Preprocess example: Ensure to have a data_dir/scenario in which rgb and tir dirs are filled with the respective parts of image pairs.

python preprocessing/full_tir_and_rgb_proproccess.py \
        scenario \
        --data_dir ../data/align_test_dirs \
        --do_rgb \
        --do_tir

(Both rgb and tir data of a input dir with the directories rgb and tir is compressed, filtered and masked. Several different strategies are available)

SuperGlue Workflow (align a directory of images to the first as reference): Ensure to have added the SuperPoint and SuperGlue models.

python aligning/superglue_workflow.py \
        data/align_test_dirs/fullmix \
        data/fullmix_aligned

(This script estimates an affine warp for each image to match the reference image and saves aligned images.)

Using cross-correlation with template matching:

python tir_matching/xcorr_rgbtir.py \
        data/align_test_dirs/fullmix \
        --output_dir data/tir_overlays_fullmix

(This matches and overlays TIR images with RGB spatially)

In the utils dir, bash helper scripts for image pairing and gif creation is found.

Labeling Helpers are in labelers. Some legacy code in trial_code.

The directory aligning also includes visualization code for SuperGlue operations as well as surface movement estimation.

• SuperPoint and SuperGlue from magicleap
• The images were acquired in the framework of the PERMA-XT project run by University of Fribourg/GEOTEST/PERMOS.
• Special thanks goes out to the Project Partners Dr. Dominik Amschwand and Prof. Jan Beutel.