Towards tree-based systems disturbance monitoring of tropical mosaic landscape using a time series ensemble learning approach

Abstract

Tree clearing and degradation inside and outside forest ecosystems in Africa are important contributors to global carbon budget and emissions. Part of the uncertainties in emission estimates, among other things, is related to the non-inclusion of disturbances across all tree-based systems and limitations in the capacity of existing methodologies to detect subtle changes or degradation. Here, we present a tree-based system (TBS) disturbance monitoring approach in a mosaic landscape in East Africa. For this purpose, high- and low-intensity disturbance detection time series algorithms, namely Breaks for Additive Season and Trend monitor (BFASTmonitor), Continuous Degradation Detection (CODED), and Landsat-based detection of Trends in Disturbance and Recovery (LandTrendr), were used. These algorithms were tested in three approaches based on reference canopy cover change data from repeated airborne laser scanning. Data processing was performed using Collection 2 Landsat (5, 7, and 8) time series in Google Earth Engine. The approaches were based on testing the accuracy of 1) the algorithms' magnitude of change outputs using multispectral information (i.e., several vegetation indices and bands); 2) multispectral ensemble learning approach applied at each of the three algorithms; and 3) combining the multiple algorithms in an ensemble learning approach. Our results showed an improvement in the accuracy of detecting a TBS disturbance from approach 1 to 3. The best result (44% omission error and 31% commission error against reference airborne laser scanning data) was obtained when BFASTmonitor and CODED were combined in an ensemble learning approach. The shortwave infrared bands were more important for detecting TBS disturbance than green, red, and near-infrared bands. The application of the best model in TBS disturbance monitoring in the region revealed that disturbances were more dominant outside forest ecosystems than inside. Improved disturbance monitoring in TBS can contribute to more accurate emission estimates in Africa and elsewhere in the tropics.