Authors: Alexander Koltunov, Carlos M. Ramirez, Susan L. Ustin, Michèle Slaton, Erik Haunreiter

The worldwide demand for timely and accurate information about ecosystem dynamics at Landsat spatial scale is growing and as of today still exceeds the availability of information. The diversity of required disturbance metrics and trade-offs between sensitivity, reliability, timelines of information generation, and flexibility toward potential customizations suggests that a single system is not likely to fill such demand in the near future. To address this challenge, the scientific community has been developing and improving various Landsat-based algorithms for land change monitoring. We describe the Ecosystem Disturbance and Recovery Tracker (eDaRT) version 2.9 — a highly automated prototype system in continuous development, which has been operated since 2012 by the USDA Forest Service Pacific Southwest Region to generate most current disturbance maps at Landsat scale and provide customized information services and inputs to science and land management applications in the Region.

The eDaRT processing system utilizes all three dimensions of dense Landsat image time series: spectral, temporal, and spatial. Two anomaly detection algorithms are sequentially applied, one estimating pixels’ disturbance status metrics in every processed image and the other detecting disturbance events, the primary output of eDaRT. The first algorithm initially estimates change relative to a user-defined fixed baseline time period, using a stratified version of the Dynamic Detection Model (DDM; Koltunov et al., 2009) applied to Landsat bands and vegetation indexes that reflect canopy greenness, abundance, and moisture content. Using the model residuals and a probabilistic context analysis, the detected anomalies are further classified as disturbed, cloud/snow, or recovered. The resulting residuals, classification maps, and the associated disturbance confidence values provide the most rapid preliminary snapshot of the current cumulative effect of disturbance and regeneration. The second algorithm detects discrete disturbance events as regime changes in the dense time series of the residuals for each pixel. First, the residuals are compared against a recent baseline window and classified to find candidates for disturbance events. Next, candidate-events are accepted based on temporal consistency of their detection. The standard outputs from this algorithm include disturbance event timing (down to 8–16 day precision) and a detection confidence as a proxy for event magnitude.

We initially evaluated eDaRT performance with high-resolution imagery and airborne LiDAR data in a test area in California for several types of annual disturbance events at 30-m scale. These tests modeled detection probabilities as functions of canopy cover loss, which estimated detection rates of 96%, 87% and 92% respectively for fire, harvest, and tree mortality events, when canopy loss values follow a uniform prior distribution. The error of commission varied between 10- 20% for most forest types (12% on average). Following ongoing optimizations and extended validations, eDaRT will expand beyond Landsat instruments to improve ecosystem monitoring, as an independent system and potentially as a part of an ensemble method.

Full Publication