A comprehensive evaluation of disturbance agent classification approaches: Strengths of ensemble classification, multiple indices, spatio-temporal variables, and direct prediction

Landsat time series images are used for the detection of forest disturbance and the classification of causal agents. Various studies have classified disturbance agents with respect to forest disturbance detected using Landsat time series images. However, the accuracy of the finally classified disturbance agents in different approaches is rarely evaluated. In this study, we investigated the effectiveness of using ensemble classification, and multiple spectral and spatio-temporal information for the accuracy of the classification of disturbance agents in two-stage prediction (i.e., disturbance agents are classified with respect to the detected disturbance) and direct prediction (i.e., disturbance agents are directly classified from Landsat temporal information). Predictor variables were derived from the results of the trajectory-based temporal segmentation of five spectral indices using an annual Landsat time series (2000–2018). We compared six approaches of classifying disturbance agents. For two-stage prediction, we investigated four disturbance detection approaches: threshold-based detection with a single spectral index, random forest (RF) model with a single spectral index, RF model with multiple spectral indices, and RF model with spatio-temporal variables. The detected disturbance pixels were aggregated to disturbance patches and classified into disturbance agents. For direct prediction, two RF models one with only temporal variables and the other with spatio-temporal variables were constructed to classify pixel-based disturbance agents. The overall accuracy of the RF model using spatio-temporal variables for direct prediction was 92.4% and significantly higher than that of the RF model for two-stage prediction (90.9%). The use of an RF model based only on a single spectral index in disturbance detection was not effective for improving accuracy compared with threshold-based detection; however, the use of an RF model based on multiple spectral indices in disturbance detection improved the accuracy of the final classification of disturbance agents. Introducing spatial variables in RF models was effective for improving the overall classification accuracy in pixel-based direct prediction. However, it was not necessary in two-stage prediction because of spatial information contained in the patches. Although a spatially discontinuous appearance was observed for the RF model for directly classifying disturbance agents, this could be an alternative approach to two-stage prediction when considering the relative classification performance and simplicity of implementation.