Crop Type Classification with Sentinel-1 SAR Data

This project classifies crop types from multi-temporal radar imagery using machine learning approaches.

Data

The raw data sources used were:

• Sentinel-1 C-band VV/VH backscatter intensity rasters over an agricultural site spanning 3 months
• Point vector data with 10 crop labels randomly distributed over the site's farmland

Packages

The key Python packages used in this project are:

Data Manipulation and Analysis:

• pandas
• numpy
• xarray
• rioxarray (GeoPandas)

Machine Learning:

• scikit-learn
• TensorFlow/Keras

Model Evaluation:

• matplotlib
• seaborn

Geospatial Data Handling:

• planetary_computer
• rasterio

Preprocessing:

• MinMaxScaler
• StandardScaler

Model Training:

• LogisticRegression
• RandomForestClassifier
• RandomizedSearchCV (hyperparameter tuning)

Miscellaneous Utils:

• warnings
• time
• itertools

The core modeling pipeline relies on Scikit-Learn and Pandas for manipulation and preprocessing of array data, along with geospatial packages like rioxarray to handle satellite raster time series. The deep learning components leverage TensorFlow/Keras.

Hyperparameter tuning was done using RandomizedSearch from Scikit-Learn to find optimal model configurations for the random forest. Evaluation was based on accuracy metrics.

Data Wrangling

1. Spatial Join: A spatial join between the crop labels and radar data extracted average VV/VH values for each farm's time series.
2. Preprocessing: Reshaped to samples x timesteps. Relative vegetation index (RVI) computed for each time series.
3. Train Test Split: 80-20 stratified split of farms to separate train and test sets.

Modeling

Varying ML approaches were tested:

• 1D CNN on RVI
• 1D CNN on radar backscatter
• Random forest on backscatter
• Random forest on RVI

Hyperparameter tuning and cross validation were utilized.

Results

Multiple models were evaluated based on accuracy on a held-out EY test set consisting of Sentinel-1 SAR data over unknown crop types.

The best performing model was a random forest classifier trained on engineered relative vegetation index (RVI) features, which achieved 85% accuracy at predicting crop types on the EY test data.

The next best model, a 1D CNN operating directly on the VV/VH time series, scored 10 percentage points lower at 75% accuracy.

The high performance of the RF+RVI approach can be attributed to:

• The derived polarization ratios captured in RVI provide meaningful geospatial features
• Ensemble modeling reduces overfitting compared to deep CNNs
• Interpretability of RFs allows analysis of important variables

Given the significant jump in accuracy on external industry data, the RF+RVI approach shows good generalization ability even with limited training samples. This model has been saved for real-world crop type mapping applications.

Future iterations could incorporate recent hyperparameter optimization techniques such as Bayesian hyperband to further improve predictive power.

Status

Project is: complete

Contact

• [email protected]
• [email protected]

Acknowledgements

• Data was taken from EY Challenge 2023 and Microsoft Planetary Hub