Near real-time monitoring of tropical forest disturbance by fusion of Landsat, Sentinel-2, and Sentinel-1 data

The ability to detect forest disturbance rapidly relies heavily on the frequency of usable satellite observations. Combining optical and SAR data can provide the most usable observations in the tropics and hence has the potential to achieve more timely detection of forest disturbance than using any individual sensor system. A new algorithm (Fusion Near Real-Time, or FNRT) for near real-time monitoring of tropical forest disturbance by fusion of Landsat, Sentinel-2, and Sentinel-1 data fits a time series model to the data of each individual sensor system and produces a change score for each clear observation during the monitoring period. The change scores from the three different kinds of sensors are then combined for disturbance monitoring. Applying FNRT using all three data sources in eight selected sites in the Amazon Basin detected 69.8 & PLUSMN; 5% and 84.6 & PLUSMN; 4.1% of the forest disturbances within the first 30 and 60 days, respectively. FNRT using all three sensors reached a peak producer's accuracy of 91.6 & PLUSMN; 3.2% at approximately a 100-day lag time, compared to 86.4 & PLUSMN; 3.8% using the two optical sensors combined, and 71.7 & PLUSMN; 4.7% using Landsat only (the best-performing single-sensor). The user's accuracy of FNRT using all three sensors increased to 90.9 & PLUSMN; 3.3% when excluding false detections in non-forest areas. Overall, the model result using all three data sources had the highest detection rate with shorter lag times. Although modest, including Sentinel-1 data in the multi-sensor fusion approach improved the detection speed of some disturbance events. Given the abundance of optical data during the dry season in our study region, Sentinel1 data would be more valuable during the wet season or in persistently cloudy regions such as Madagascar or coastal West Africa.