MODTrendr
MODTrendr
Project Summary
The MODIS-based detection of trends in disturbance and recovery (MODTrendr) project is based on a similarly named algorithm called LandTrendr developed by Robert Kennedy and Zhiqiang Yang at Oregon State University. Their approach is to temporally segment a Landsat time series into distinct periods representing stable, disturbance, and recovery. We adapted this approach to MODIS time series data which have much lower spatial resolution. The modified algorithm was tested over a large forested area in the Pacific Northwest of the USA where an existing LandTrendr map already existed. The MODTrendr adaptation was shown to be very successful at detecting disturbances even those that affected areas smaller than a MODIS pixel. This algorithm is now being tested on a variety of other regions and with other types of inputs including time series of posterior probabilities from land cover classifications and biomass predictions from similar remote sensing data.
To download the MODTrendr codes please contact the author.
These pictures were taken during a trip to Oregon in September 2012 with Garrett Meigs and Scott Goetz. We saw a lot of dead trees and even visited a live forest fire near Bend, Oregon.
Related Publications
Sulla-Menashe, D, RE Kennedy, Z Yang, J Braaten, ON Krankina, and MA Friedl. 2014. Detecting forest disturbance in the Pacific Northwest from MODIS time series using temporal segmentation. Remote Sensing of Environment (in press).
(A) Spatial distribution of the Landsat Cumulative Disturbance Magnitude metric across the study area for the period 2001-2008. (B) Spatial distribution of MODIS dNBR for the period 2001-2008. (C) Spatial distribution of Landsat Cumulative Disturbance Magnitude for the historical period 1985-2000. (D) Omission (purple) and commission (green) errors in MODTrendr results relative to the NWFP-DB for the period 2001-2008. Panels A-C are colored according to their dNBR values: a small change (cyan) is defined to be less than 0.2 dNBR, a moderate change (orange) as between 0.2 and 0.5 dNBR, and a large change (red) as greater than 0.5 dNBR. Panels A and C consider a pixel stable forest if the Disturbed Area ≤ 0.05 MODIS pixels and Panel B considers a pixel stable forest if MODIS dNBR ≤ 0.08.
Abstract:
Fire, insects, and human activities are the dominant drivers of forest disturbance at the global scale. Because forests are geographically extensive and are often remote, the Moderate Resolution Imaging Spectroradiometer (MODIS) is uniquely suited to monitor the state and health of forested ecosystems. However, the extent to which coarse-resolution remote sensing data can accurately capture spatial and temporal patterns of disturbance is unclear. To investigate this, we developed an 11-year time series of MODIS Normalized Burn Ratio images corresponding to peak-growing season conditions for a study area located in the Pacific Northwest of the conterminous United States. Using a temporal segmentation algorithm that was originally developed using Landsat TM and ETM data, we created annual maps of forest disturbance from these time series. We then compared these maps to a database of annual forest disturbance that was compiled using Landsat TM/ETM data for the same region. Results from this comparison revealed that about half of all pixels affected by disturbances that occupied more than 5% of a MODIS pixel were correctly identified as disturbed, including 79% of those that were affected by disturbances larger than one-third of a MODIS pixel. Our results also show that the size, severity, and timing of disturbance events, along with gridding artifacts inherent to MODIS data, interact in complex ways that influence the signature of forest disturbance events in MODIS data. These results demonstrate both the utility as well as the limitations of MODIS and other coarse spatial resolution sensors for monitoring forest disturbance at regional to global scales.
Example of a segmentation result from LandTrendr. In a sequence of some spectral index for a single Landsat pixel through time, a disturbance segment (highlighted in red) can be summarized by its timing, severity, and duration.