Near real-time monitoring of tropical forest disturbance: New algorithms and assessment framework

Near real-time monitoring systems of forest disturbance are essential to reduce illegal logging and global deforestation, but the assessment of their performance varies greatly and rarely adheres to recommended assessment practices. Current assessment protocols recommended in the published literature focus on sample-based estimation of area and accuracy of features in remote sensing-based maps, which is less relevant for assessing the performance of near real-time monitoring systems. Rather than area bias and accuracy of mapped features, the objective of near real-time monitoring is fast detection of forest disturbance events. Here we present a new assessment framework as well as two new algorithms for near real-time monitoring of tropical forest disturbance. The first algorithm (NRT-CCDC) fits time series models based on MODIS data to predict future observations. Detections of forest disturbance are labeled as “low-probability”, “high-probability”, and “confirmed” based on the number of consecutive observations that deviate substantially from the prediction. The second algorithm (Fusion2) predicts MODIS observations at a high temporal frequency by building a model based on a time series of Landsat observations. Fusion2 utilizes daily MODIS observations from both Terra and Aqua to achieve rapid detection of disturbance events. A framework for assessing the performance of near real-time monitoring is presented that focuses on the timing and minimum detectable size of forest disturbance events while still being based on probability sampling and design-based inference. Central to the framework is a new metric we refer to as the “alert-lag relationship”, which characterizes the frequency of omitted disturbance events as a function of the time lag between the dates of disturbance and detection. When applied to three different near-real monitoring systems representing three levels of operational readiness (Fusion2, NRT-CCDC and Terra-i), we found that Fusion2 achieved an alert accuracy (event-based agreement) of 50% after 82 days and NRT-CCDC achieved 50% after 126 days. Terra-i never achieved an alert accuracy of 50% but achieved 30% after 109 days. In terms of event size, all three systems showed better performance in detecting large events compared to small events. The smallest event detected by the three systems was 6.5 ha. Conventional accuracy assessment (pixel-based agreement) showed that the user's accuracy, which characterizes commission errors, of the disturbance class for Fusion2, NRT-CCDC, and Terra-i were 87%, 79%, and 30% respectively. The two more complex systems, Fusion2 and NRT-CCDC were consistently more accurate than the operational Terra-i system. In general, the higher the cost of implementing the system corresponded to a higher accuracy of the detection system. A primary objective moving forward is to make the NRT-CCDC and Fusion2 more operational over larger areas.