A Remote Approach to Cashew Orchard Detection: Leveraging Active Learning with Satellite Imagery in Guinea-Bissau

Can active learning and medium-resolution satellite imagery be combined to accurately map cashew orchards across an entire country where no nationwide agricultural database exists, thereby overcoming the limitations of previous regional-level studies?

The researchers utilized a 360-feature raster derived from Copernicus Sentinel-2 satellite imagery spanning 2017 to 2021, incorporating spectral composites, temporal harmonic regression coefficients from the CCDC algorithm, and spatial Gray-Level Co-Occurrence Matrix statistics. To efficiently build training data without extensive on-site fieldwork, they employed a pool-based active learning framework using Margin Sampling alongside a Support Vector Machine classifier. Experts annotated pixel patches using historical high-resolution imagery in Google Earth Pro to classify land cover into eight distinct categories across Guinea-Bissau and northern Guinea.

Active learning dramatically optimized the training data collection, requiring only 1,809 points for the Margin Sampling dataset compared to 4,498 points for traditional Random Sampling while maintaining superior performance. The final machine learning model produced a nationwide 2021 cashew orchard map with a 10-meter spatial resolution and an overall balanced accuracy of 94.0%. All generated datasets and the complete countrywide map are openly accessible via GitHub to support future environmental monitoring.

Unregulated cashew production is a primary driver of deforestation and biodiversity loss in West Africa, yet regional conservation efforts have been severely hindered by a lack of comprehensive geospatial data. By demonstrating that high-accuracy national mapping can be achieved entirely off-site using scalable machine learning and active learning workflows, this study provides a vital tool for governments and conservationists to monitor land-use change and protect vulnerable ecosystems.