Accurate mapping of forest disturbances is challenging and costly. Time series change detection is commonly used to detect disturbances, but is prone to high rates of error. Here, we introduce a method that uses regionally-distinct machine learning models, making use of spatial and temporal information, to both detect and attribute disturbances. This approach should reduce errors by accounting for differences in disturbance characteristics across varying forest types. For this study, we used summer NDVI data from 7 Landsat scenes located across the contiguous United States. Times series from 2000-2016 were used to create 47 spatial and temporal variables. We trained 350+ points for each region, deciding disturbance condition using time series images and values, and recent satellite imagery. Random forest models were trained and used to predict disturbance locations. An additional 250+ training points were selected from the detected points and classified by type. These data were used to train another random forest model, which was used to predict disturbance types across the area. Validation was completed using historical imagery in addition to detection training methods. Using our developed methods, we then asked, how do variables important for disturbance detection and attribution vary among regions?
Results/Conclusions
Our methods show promise as a new framework for detecting and attributing forest disturbances across a diverse array of forest types. Validation showed false positive error rates of 2-23%, and false negative rates of 3-40%. Incorrect disturbance attribution occurred at rates of 1-13%. Lower-severity biotic disturbances in Northern Colorado led to the highest false negative rate (40%). The highest false positive rate (23%) occurred in Eastern Pennsylvania, potentially due to the deciduousness of regional forests. Variables important to disturbance detection and attribution differed regionally, with no clearly discernable geographic patterns. Temporal variables were more important for the detection of disturbances, while spatial variables were more important for the attribution of disturbances. However, all regions had at least one spatial and one temporal variable in their final detection and attribution models. The magnitude of NDVI decline in the time series was important across many regions, as was the minimum slope over the time series. The variation in both variables in the surrounding area was also frequently important. Currently, the developed method is highly useful for small to moderate-scale studies and provides a relatively quick and easy tool to identify forest disturbances for use in ecological analyses of disturbance impacts and interactions.