Data and code for: Neighborhood-Aware Retrieval of a Spatial Floe Fragmentation Index for Arctic Sea Ice Using Sentinel-1 SAR and AMSR2

Overview

This repository contains the information for the paper "Neighborhood-Aware Retrieval of a Spatial Floe Fragmentation Index for Arctic Sea Ice Using Sentinel-1 SAR and AMSR2".

Author: Bo Li, Xi Zhao

Contact: libo97@mail2.sysu.edu.cn

Environment: PyCharm 2021, Python 3.10

Last Updated: July 2026

Data

The archived data include metadata and quicklook images for the Sentinel-1 SAR scenes, DSI-Net-derived ice–water classification and floe contour products, the matched SAR–AMSR2 retrieval samples, and the prediction results obtained with the best-performing model:

• Sentinel1_416scenes_manifest.csv — This file provides the inventory of the 416 Sentinel-1 SAR scenes, including the scene identifiers and associated acquisition and file information required to link the quicklooks, DSI-Net outputs, and SAR-derived SFFI samples.
• Sentinel1_416scenes_HH_quicklooks.zip — This archive contains HH-polarization quicklook images for the 416 Sentinel-1 SAR scenes. The images are provided for scene visualization and qualitative inspection and should not be treated as substitutes for the original analysis-ready Sentinel-1 GRD products.
• DSI-Net_sentinel1_416Scenes_ice_water_classification_results.zip — This archive contains the DSI-Net-derived ice–water classification maps for the 416 Sentinel-1 SAR scenes used to generate the SFFI reference dataset.
• DSI-Net_sentinel1_416Scenes_floe_contour_extraction_results.zip — This archive contains the individual floe contour extraction results generated by DSI-Net for the 416 Sentinel-1 SAR scenes. These contours provide the floe area, spatial separation, and geometric information used to calculate the SAR-derived SFFI reference values.
• SFFI_SAR_AMSR2_matched_samples_52074.csv — This file contains the 52,074 spatially and temporally matched SAR–AMSR2 samples used for model training and evaluation. Each sample includes a SAR-derived SFFI reference value, AMSR2 multifrequency dual-polarization brightness temperatures and derived radiometric variables, ASI sea ice concentration, 3 × 3 neighborhood statistics, and associated spatial and acquisition identifiers.
• G3-XGBoost_SFFI_test_samples_pred.csv — This file contains the SFFI predictions produced by the best-performing G3-XGBoost model for the randomly partitioned test samples. It includes the SAR-derived reference SFFI values, model predictions, and prediction errors used to reproduce the model evaluation and observed-versus-predicted analysis.

Code

The archived code implements the principal processing, retrieval, and evaluation procedures used to derive the Spatial Floe Fragmentation Index (SFFI) from Sentinel-1 SAR observations and retrieve it from AMSR2 data:

• calculate_amsr2_radiometric_features.py — calculates polarization differences (PDs), polarization ratios (PRs), and gradient ratios (GRs) from spatially aligned AMSR2 brightness-temperature GeoTIFFs at 23.8, 36.5, and 89.0 GHz. The script verifies grid consistency, propagates invalid observations, and records processing status in a manifest.
• extract_complete_amsr2_cells_within_sar.py — identifies AMSR2 grid cells that are fully contained within each Sentinel-1 SAR scene. It exports the selected cell locations and their corresponding bounds on the SAR grid for subsequent SFFI calculation.
• calculate_sffi_from_floe_contours.py — calculates grid-scale SFFI reference values from individual floe contours extracted from Sentinel-1 SAR imagery. The calculation combines floe area composition, distance between floes, and floe circularity according to the SFFI formulation presented in the associated manuscript. Both single-scene and manifest-based batch processing are supported.
• train_random_forest_sffi_retrieval.py — trains and evaluates the random forest (RF) retrieval model using the G1, G2, or G3 feature configuration.
• train_xgboost_sffi_retrieval.py — trains and evaluates the extreme gradient boosting (XGBoost) retrieval model using the same feature configurations and exports the fitted model in the native XGBoost format.
• train_mlp_sffi_retrieval.py — trains and evaluates the multilayer perceptron (MLP) retrieval model using standardized input variables and an internal validation subset for early stopping.
• predict_sffi_with_trained_model.py — applies a trained retrieval model to a feature table, writes SFFI predictions, and optionally evaluates the predictions when reference SFFI values are available.
• batch_predict_daily_sffi.py — applies a trained model to daily AMSR2 feature tables over a user-defined date range. It records successful, failed, skipped, and missing dates in reproducibility manifests.
• evaluate_and_plot_sffi_predictions.py — calculates R², RMSE, MAE, and bias from observed and predicted SFFI values and produces an observed-versus-predicted scatterplot colored by absolute error.
• analyze_sffi_feature_importance.py — extracts individual and grouped feature importance from a trained tree model and exports tabular results and publication-ready figures.
• XGB_SFFI_model.pkl — contains the archived best-performing G3-XGBoost model used for direct reproduction of the reported SFFI predictions. Because Joblib files can execute serialized Python objects when loaded, this file should be used only when obtained from the trusted study archive.

The archive begins with preprocessed AMSR2 GeoTIFFs, SAR-derived individual-floe contours, or the released retrieval sample tables; it does not download the source satellite products or execute DSI-Net inference. The complete workflow, required inputs, command-line examples, feature definitions, model-format notes, and data links are provided in README.md.

The scripts require Python 3.10 or later and use NumPy, pandas, SciPy, scikit-learn, XGBoost, Joblib, Matplotlib, OpenCV, Rasterio, Shapely, and Affine. Exact installation instructions and compatible package ranges are provided in requirements.txt.

If you use these scripts or modified versions in your research, please cite this repository and acknowledge the author.
For any questions or collaboration inquiries, please contact:

Bo Li, libo97@mail2.sysu.edu.cn