Distinguishing natural and agricultural tree systems is critical for monitoring ecosystem services, commodity-driven deforestation, and restoration progress. From a remote sensing standpoint, this task is particularly challenging because:
- Certain trees exhibit high spectral similarity
- Highly heterogeneous, smallholder agricultural landscapes require a small minimum mapping unit to detect differences
- Regions with persistent cloud cover and atmospheric haze reduce optical image quality
The distinction is critical for restoration monitoring applications, as gains in tree cover cannot be meaningfully assessed without understanding whether they result from a successful restoration intervention or agricultural expansion.
This project applies a transfer learning approach to classify tree-based land use systems from satellite imagery. We leverage spatial embeddings extracted from a high-performing convolutional neural network originally trained for tree cover mapping (Brandt et al., 2023) and repurpose them for land use classification.
We train a CatBoost classifier using a combination of Sentinel-1 and Sentinel-2 imagery, gray-level co-occurrence matrix (GLCM) texture features, and extracted spatial embeddings to classify four land use classes: natural, agroforestry, monoculture, and other (background). Through comparative modeling and feature selection, we demonstrate consistent performance gains from incorporating both transfer-learned features and texture information.
In collaboration with Ghana’s Environmental Protection Agency, the method is demonstrated across 26 priority districts, resulting in a 10-meter resolution land use map for 2020. Overall, the findings suggest that spatial embeddings learned for tree detection retain meaningful information about land use structure, offering a scalable pathway for broader monitoring of natural and agricultural tree systems.
Download the paper: Technical Note
View the data: Ghana EPA Monitoring Portal (toggle on WRI Land Use)
Suggested citation:
Ertel, J., J. Brandt, R. Rognstad, and E. Glen (2025). Transfer learning to detect natural, monoculture, and agroforestry tree-based systems in Ghana using remote sensing. Technical Note. Washington, DC: World Resources Institute. doi:10.46830/writn.24.00030
The predictive features for the classification task include Sentinel-1 and Sentinel-2 imagery, spatial embeddings and texture features. Texture features were derived from Sentinel-2 imagery using a GLCM analysis method. The figure below shows the machine learning pipeline, including pre- and post-processing steps.
The src directory is designed to integrate with Data Version Control (DVC) to ensure reproducibility and efficient management of the project's machine learning pipeline.
This directory contains modular scripts and functions organized to support a DVC workflow. Each script performs a specific task within the pipeline, such as data preparation, model training, or evaluation. The pipeline stages are connected through DVC.
- Clear separation of pipeline stages.
- Improved tracking for dependencies, outputs, and metadata.
- Compatible with YAML-based pipeline configurations.
/src
├── stage_load_data.py # Scripts for ingesting raw data
├── stage_prep_features.py # Scripts for data cleaning, transformation, and feature engineering
├── stage_train_model.py # Script to train machine learning models
├── stage_select_and_tune.py # Script to perform hyperparameter tuning and feature selection
├── stage_evaluate_model.py # Script to evaluate model performance
└── transfer_learning.py # Script for running inference
The pipeline is defined in the dvc.yaml file, with dependencies, parameters, and outputs explicitly stated for each stage.
Pipeline parameters are defined in params.yaml. This file centralizes hyperparameters and configuration options for each stage of the pipeline. Update parameters as needed, and rerun the pipeline using dvc repro to propagate changes.