Sentinel-1 enhanced deep-learning deforestation prediction model; A case study of Equatorial Guinea, Gabon, and the Republic of the Congo

By Eitan Buffaz

Abstract
This study evaluates the integration of Sentinel-1 synthetic aperture radar (SAR) imagery into the World Wildlife Fund’s Forest Foresight (FF) deep-learning prediction framework to enhance deforestation predictions across Equatorial Guinea, Gabon, and the Republic of the Congo. Two new models are implemented and compared to the baseline model from FF. One uses 10-meter spatial resolution Sentinel-1 VH ground range detected data, while the other uses 40-meter resolution. The two new models consist of the original encoder from the FF baseline model with an additional encoder for the Sentinel-1 data. The features are then concatenated and decoded together to get a prediction map. Performance is assessed using the area under the precision-recall curve (AUC) and F0.5 score for each of the countries, the full study area, and “new” deforestation pixels (where there were no alerts in the preceding six months). Results indicate that the 10-meter fusion model achieves a marginal improvement of 1% in AUC over the baseline model for the full area and 2% in new deforestation areas. The 40-meter model is closer in performance to the baseline model, but its performance is slightly better on new deforestation areas while keeping the computational and storage costs substantially lower. The 40-meter resolution model strikes a balance between performance and costs that could make it the preferred model for FF’s stakeholders in the Congo Basin.