Prediction of Freeze/Thaw Transitions in Soil

📌 TL;DR

Comparison of machine learning (ML) driven analysis of Advanced Scatterometer (ASCAT) backscatter vs ERA5 data in predicting freeze/thaw (F/T) transitions in soil, validated against international soil moisture network (ISMN) ground observations.

🚀 Overview

This project investigates whether ML models applied to ASCAT backscatter time-series can accurately predict soil F/T transitions, and how these predictions compare to those derived from ERA5 reanalysis data.

Freeze/thaw transitions are identified using in-situ soil temperature observations from the ISMN, which serve as reference ground truth. For a set of selected stations, ASCAT backscatter observations are processed into time-series features and used to train ML models that predict daily frozen or thawed soil states. Transition dates are then extracted and compared against:

• Ground-based ISMN observations (reference)

• ERA5-derived soil temperature transitions (baseline)

The primary research question guiding this project is:

How does ML-processed ASCAT data compare to ERA5 in accurately predicting soil freeze/thaw transitions?

A set of 10 ISMN stations was created, and the locations of the stations are depicted below.

Repository Structure:

├── ISMN_site_survey.csv
├── requirements.txt
├── notebooks
│   ├── rolling_classification
│   │   ├── 1_ISMN_data_cleaning.ipynb
│   │   ├── 2_ASCAT_ERA5_data_cleaning.ipynb
│   │   ├── 3_assess_methodology.ipynb
│   │   ├── 4_model_selection.ipynb
│   │   └── 5_model_vs_ERA5.ipynb
│   └── simple_classification
│       ├── 1_ISMN_data_cleaning.ipynb
│       ├── 2_ASCAT_ERA5_data_cleaning.ipynb
│       ├── 3_assess_methodology.ipynb
│       ├── 4_model_selection.ipynb
│       └── 5_model_vs_ERA5.ipynb
└── src
    └── freeze_thaw
        ├── data_preparation
        │   ├── ascat_era5.py
        │   ├── feature_engineering.py
        │   ├── general.py
        │   ├── ismn.py
        │   ├── nan_handling.py
        │   ├── outlier_detection.py
        │   └── splitting.py
        ├── data_understanding
        │   └── visualization.py
        ├── evaluation
        │   ├── metrics.py
        │   └── statistics.py
        │   └── visualization.py
        ├── modeling
        │   ├── test.py
        │   └── train.py
        ├── _internal_functions.py
        ├── config.py
        ├── utils.py
        └── validation.py

🎯 Motivation

Soil F/T dynamics play a crucial role in:

• Infrastructure stability (frost heave, subsidence)

• Agricultural productivity and planting cycles

• Hydrological processes and runoff generation

• Permafrost monitoring and climate change assessment

Ground-based measurements can determine F/T transitions with high accuracy, but their spatial coverage is sparse and uneven globally. In contrast, satellite remote sensing offers consistent global observations. The ASCAT microwave scatterometer measures surface backscatter, which is sensitive to changes in soil dielectric properties associated with freezing and thawing. ML may enable direct extraction of F/T transitions from these backscatter time series.

ERA5 is a widely-used source that provides model-based soil temperature estimates derived from data assimilation. Possible downsides of these reanalysis products are that they may smooth or misrepresent local transition timing.

This project evaluates whether ML applied to ASCAT backscatter can match or outperform ERA5 in predicting soil freeze/thaw transitions, providing a reproducible and observation-driven alternative for large-scale monitoring.

📊 Results

Simple Classification

The first set of results are based on classifying an observation solely on its ISMN soil temperature. The boundaries are defined as -1 °C < transition ≤ 1 °C, with frozen and thawed being below and above these bounds respectively. For more details, see notebooks/simple_classification/5_model_vs_ERA5.ipynb.

The two-sided exact paired sign-flipping permutation test ($\alpha=0.05$) was applied to each metric to determine if there's a statistically significant difference in performance across the ten stations between the two methods. The p-values for accuracy, macro F$_1$, and transition F$_1$ are 0.002, 0.003, and 0.002 respectively.

Rolling Classification

The second set of results are based on classifying an observation based on its current ISMN soil temperature along with the temperatures for the preceding 72 hours. An observation is classified as frozen if all temperatures are at or below 0 °C. Vice versa for the thawed state. The transition state is if the temperatures cross the boundary. For more details, see notebooks/rolling_classification/5_model_vs_ERA5.ipynb.

The same statistical test was applied to each metric to determine if there's a statistically significant difference in performance across the ten stations between the two methods. The p-values for accuracy, macro F$_1$, and transition F$_1$ are 0.002, 0.002, and 0.003 respectively.

The results suggest that the lightGBM model trained on the ASCAT data significantly outperforms ERA5. However, there's a weakness in the methodology that may be unfairly punishing ERA5. Specifically, large area measurements are being evaluated against a single point measurment from an ISMN station. The lightGBM model trained on ASCAT is also affected by this, but the model is specifically trained against labels derived from a single ISMN point measurement. A better approach may be to select multiple ISMN stations that fall under one ASCAT and ERA5 footprint, then taking the average temperature. See notebooks/../3_assess_methodology.ipynb for more details.

🛠️ Tech Stack

• Python (3.12)
• Pydantic
• Data manipulation, analysis, & visualization
  • pandas
  • NumPy
  • Matplotlib
  • Seaborn
  • Plotly
• Model training & evaluation
  • LightGBM
  • Scikit-learn
• Jupyter Notebook

▶️ How to Run

Setup

It's recommended to use Python 3.12. Create virtual environment:

python3.12 -m venv .venv

If using Mac/Linux:

source .venv/bin/activate

If using Windows:

.venv\Scripts\activate

Install requirements:

pip install -r requirements.txt

• Download the raw ASCAT and ERA5 data files from and move them to ../data/raw/ASCAT_ERA5.
• Download the ISMN data and move them to ../data/raw/ISMN/{station_name}.
• Run notebooks 1-4.
• Execute ../src/freeze_thaw/modeling/lgb_train.py; be mindful of the parameters under if __name__ == '__main__':
• Run notebook 5.

ℹ️ Sources

The raw ASCAT and ERA5 data files can be found at .

The individual swath orbit files from the ASCAT Surface Soil Moisture (SSM) Climate Data Record (CDR) v8 at 12.5 km sampling (H121, https://doi.org/10.15770/EUM_SAF_H_0011 ) were stacked and converted into time series format. Since the data are provided on a fixed Earth grid, this processing step involved only restructuring the data without altering the original values. For each in situ station, the time series of the nearest ASCAT grid point was extracted. During this transformation, data from the corresponding Intermediate Climate Data Record (ICDR, H139) were appended as well.

ERA5 data has been downloaded from the Copernicus Climate Data Store (CDS) (https://doi.org/10.24381/cds.adbb2d47) and converted into time series format. The ERA5 data are provided on a 0.25 degree grid and data has not been altered during data conversion. The time series of the closest grid point of the ERA5 dataset has been extracted of each in situ station.

ISMN T&Cs forbids the re-export or transfer of the original data to third parties. Therefore, the ISMN data used for this analysis is not included in this repo or elsewhere. The data used for the analysis can be downloaded from the data provider using the following steps:

1. Create an account at ismn.earth.
2. Click on "Data Access" on the home page, then set the initial filters:

3. Search for stations:

4. Create and execute an area filter around the station of interest, then click "Download" using the four steps depicted below:

5. Select the following parameters, then click on "your requests" to download the data.

🙏 Acknowledgements

• Main Supervisor: Univ.Prof. Wolfgang Wagner
• Co-supervisor: Sebastian Hahn
• Co-supervisor: Prof. Nysret Musliu

©️ Licensing

The source code in this repository is licensed under the MIT License.

• ASCAT data are © EUMETSAT OSI SAF and licensed under CC-BY-4.0.
• ERA5 data are © Copernicus Climate Change Service / ECMWF and licensed under CC-BY-4.0.

The data are not covered by the MIT license.