Teaching AI to see permafrost: an early study

The mapping bottleneck

Rock glacier inventories are a prerequisite for permafrost mapping at regional scale, but they are produced entirely by expert visual interpretation of aerial or satellite imagery. Across the French Alps alone, this produced an inventory of over 3,000 landforms, each requiring individual review - an effort that took years. Large areas of the globe remain incompletely charted because the human effort required is not available. Automated detection tools, if accurate enough to serve as first-pass filters, can support this process.

Training a CNN to recognize rock glaciers

The approach treats rock glacier recognition as an image classification problem. A dataset of 954 labeled rock glacier images was assembled from SPOT 6 orthoimages (grayscale, 1.5 m resolution) covering the French Alps, selecting only landforms with clearly visible creep features (frontal lobes, ridge-furrow morphology) and no snow or shadow obscuring the surface. A matching set of 4,147 generic mountain terrain images provided the negative class, drawn from areas near but not overlapping the inventoried landforms.

Rock glacierTerrain954 patches in training set

SPOT 6 panchromatic · 150×150 px · 1.5 m resolution · French Alps

The CNN architecture used three convolutional blocks with 32, 64, and 128 filters respectively, each followed by ReLU activation and 2×2 max pooling, then two fully connected layers before a softmax output. Input image size was set at 150×150 pixels, determined by hyperparameter optimization. Training used the Adam optimizer with categorical cross-entropy loss for 5 epochs, after which further training caused overfitting on the validation set.

Performance was evaluated on an independent test set from the Italian side of the Alps (127 rock glacier images, 156 generic terrain images), a region the model had never seen during training. The result: 88% true positives (112 of 127 rock glaciers correctly classified) and 79% true negatives (124 of 156 terrain images correctly rejected). Failures concentrated on landforms with atypical illumination, fresh snow, or subdued surface morphology.

Sliding-window detection over a test landscape

Classifying pre-cropped image patches is one thing; detecting rock glaciers within a large unprepared orthoimage is another. The detection algorithm extracted overlapping 200×200 pixel windows from the full SPOT 6 scene at a step of 20 pixels (30 m on the ground), classified each window with the CNN, and recorded the probability score at each position. The test area covers approximately 80 km² of the Argentera massif in the southwestern Italian Alps. The algorithm correctly identified the majority of the larger rock glaciers in the area, particularly those with pronounced lobes and furrows and minimal vegetation. False positives appeared mainly over stream thalwegs and other debris-covered terrain with high morphological complexity.

The study was framed as a proof of concept: demonstrating that a CNN trained on a few hundred images could generalize to an independent region with meaningful accuracy. Open extensions noted at publication included multiscale detection windows, integration of DEM derivatives as additional input channels, and transfer learning from larger remote sensing datasets.