This repository contains the end-to-end pipeline for rts_mapping_v2, a project dedicated to creating a comprehensive, high-resolution inventory of Retrogressive Thaw Slumps (RTS) across the Arctic. This work builds upon the proof-of-concept U-Net model developed in our initial paper and scales it up by leveraging the new, large-scale Arctic Retrogressive Thaw Slumps (ARTS) dataset.
Retrogressive Thaw Slumps are accelerating with Arctic warming, impacting landscapes, ecosystems, and global carbon cycles. Mapping their distribution is critical, but has been limited by the lack of large, diverse training datasets.
This project addresses that limitation by implementing a complete, three-step workflow:
Generate a massive, analysis-ready training dataset by pairing the 43,000+ polygons from the ARTS dataset with modern, high-resolution satellite imagery.
Train and benchmark state-of-the-art deep learning models, including both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), to identify the optimal architecture for RTS detection.
Deploy the best model at a pan-Arctic scale to produce the most extensive RTS inventory to date.
This project is structured as a sequential pipeline, with each step building upon the last.
-
Step 1: Training Data Generation The foundation of this project is the creation of a new, large-scale raster dataset called ARTS_image. This process is automated using a Google Earth Engine (GEE) pipeline.
Input Labels: The ARTS dataset, containing 23,529 RTS-present and 20,434 RTS-absent vector polygons.
Input Imagery:
PlanetScope: High-resolution RGB imagery.
Sentinel-2: Multispectral data used for deriving indices like NDVI.
ArcticDEM: High-resolution elevation data.
Process: For each polygon in ARTS, the pipeline queries GEE to find the latest available, cloud-free imagery. It then clips the imagery and DEM to the polygon's extent, rasterizes the label, and exports a multi-band GeoTIFF training tile.
Output: A collection of GeoTIFFs stored in a Google Cloud Storage (GCS) bucket, ready for model training.
-
Step 2: Model Training & Development With the ARTS_image dataset, we will train and evaluate various deep learning models for semantic segmentation. This work will be conducted in a GPU-enabled environment like Google Colab.
Architectures: We will compare the performance of a U-Net based CNN against a modern Vision Transformer (ViT) architecture to determine the state-of-the-art for this specific task.
Class Imbalance Strategy: A core challenge in remote sensing is the extreme rarity of the target class (RTS) compared to the background. We will address this by utilizing the extensive RTS-absent polygons from the ARTS dataset and employing data augmentation and advanced loss functions (e.g., Focal Loss, Dice Loss) to ensure the model learns robust features for both classes.
Experiment Tracking: All training runs will be systematically logged to compare model performance metrics (e.g., IoU, F1-score) and select the optimal model for deployment.
-
Step 3: Pan-Arctic Inference & Inventory Generation The final step is to use the best-performing model from Step 2 to generate the pan-Arctic RTS map.
Platform: This large-scale inference will be run on Google Earth Engine to leverage its parallel processing capabilities.
Process: The trained model weights will be used to run inference on a pan-Arctic mosaic of the latest available PlanetScope, Sentinel-2, and ArcticDEM data. The model will output a probability map indicating the likelihood of an RTS at each pixel.
Output: The final product will be a vector inventory (e.g., GeoPackage or Shapefile) of all detected RTS features across the Arctic, derived from thresholding and vectorizing the model's output probability map.
Data Processing: Google Earth Engine (GEE)
Storage: Google Cloud Storage (GCS) & Google Drive
Model Training Environment: Google Colab, VS Code, Google cloud computing
Core Libraries: geopandas, rasterio, earthengine-api, tensorflow / pytorch, scikit-learn
This project is made possible by and builds directly upon the following foundational work:
The ARTS Dataset: Yang, Y., Rodenhizer, H., Rogers, B.M. et al. A Collaborative and Scalable Geospatial Data Set for Arctic Retrogressive Thaw Slumps with Data Standards. Sci Data (2025).
The proof-of-concept paper: Yang, Y., Rogers, B.M., Fiske, G. et al. Mapping retrogressive thaw slumps using deep neural networks. Remote Sensing of Environment 288 (2023) 113495.