A robust method to improve the near real-time forest change detection products accuracy

Abstract

Near real-time forest disturbance detection provides timely information for people to understand forest dynamic changes and is an important data source for sustainable forest management. Continuous introduction of near real-time observation products and change detection algorithms facilitates comparative analysis of different products and the possibility of combining different forest disturbance detection products has emerged.

This study collects data products from RADD alerts and GLAD alerts and implements the Breaks for Additive Season and Trend (BFAST) on Sentinel-2 time series to detect forest change. A new product is obtained by their combination and the performance of them are evaluated. The study proves the feasibility of combining different forest change detection products to improve the accuracy.

The validation results showed user accuracies of 96% and producer accuracies 81% on the detection forest loss, which has a higher accuracy than other three individual products. The combined product also has a better performance reducing the time lag of detection (17.5 days).