Differentiating between natural and agricultural trees using remote sensing is essential for assessing ecosystem services, commodity-driven deforestation, and restoration progress. Many existing approaches focus on identifying individual tree commodities, rather than classifying tree systems in a way that is agnostic to species and management practices. This research presents a transfer learning approach for classifying tree-based systems, leveraging spatial embeddings extracted from a pre-trained neural network to improve performance in label-scarce environments.

We apply a CatBoost classifier to a combination of Sentinel imagery, gray-level co-occurrence matrix (GLCM) texture features, and learned 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. The approach evaluates whether spatial representations learned for tree detection can be repurposed to distinguish broader tree-based land use systems.

The method is demonstrated across 26 priority districts in Ghana, resulting in a 10-meter resolution land use map for 2020. The findings indicate that learned spatial embeddings retain meaningful information about land use structure beyond their original task, offering a scalable path forward for broader monitoring efforts.

Code Repositories: