top of page

Detecting Deforestation with Google Earth Engine

Analyzing a new method of identifying land use change

Google Earth Engine & the IW Method

   Google Earth Engine is an open-access, cloud-based platform for analyzing geospatial information. Earth Engine's catalog provides access to an extensive library of satellite imagery and tools to enable analysis of large datasets. Earth Engine has proven useful for global analysis of land cover change, like the forest loss study conducted by researchers at the University of Maryland that produced this map of global tree cover change.

   In this analysis, I test out a new method for detecting land cover change developed by the Defenders of Wildlife. The IW method calculates normalized changes in 5 different spectral metrics:

  • CV: Change Vector measures total change in reflectance between two images.

  • NBR: Normalized Burn Ratio identifies areas that had recently burned.

  • NDSI: Normalized Difference Soil Index identifies changes in the concentration of bare ground.

  • NDVI: Normalized Difference Vegetation Index identifies changes in concentration of vegetation.

  • RCVmax: Relative CV Maximum measure total change in each band.

Study area highlighted.

Source: Google Earth Engine

Methods

   In order to test the IW, I used the growing suburbs just west of the San Antonio, Texas city limits as my sample study area. I used imagery of a growing unincorporated community known as Rio Medina, which had a population of 60 at the time of the 2010 census but now has over 600 residents according to an online community page.

   I ran the Defenders of Wildlife's IW script on Sentinel-2 imagery of the Rio Medina area between August of 2016 and March of 2017. This script produced a layer for each of the five spectral metrics, and I manually selected areas of visible land change to compare the results. The goal of this procedure was to analyze the IW method's ability to recognize forest loss due to development of land for residential use, as compared to human visual analysis of the imagery.

   The analysis of the IW's accuracy is based only on the pixels in areas that I selected as true or not-true land cover change.

   

Visual Analysis of Rio Medina-area Imagery (change areas highlighted)

Sentinel-2 Imagery for Rio Medina, Texas, 7/17/2016.

Areas with significant forest loss highlighted for comparison.

   

Sentinel-2 Imagery for Rio Medina, Texas, 3/7/2017.

Areas with significant forest loss highlighted for comparison.

   

Results & Discussion

   For the Rio Medina study area, the IW method most accurately identified true land cover change on the CV index, the NDSI, and the RCV. It was less successful at identifying true change in the NBR and NDVI indices. Since this area did not have forest fire activity during the study period, it is unsurprising that the Normalized Burn Ratio (NBR) failed to identify burned areas. However, the fact that the NDVI failed to consistently distinguish true land cover change is interesting, and likely due to the fact that the area has large amounts of agricultural land use, some of which was vegetated in the first image (August 2016) and cleared in the second image (March 2017). In the top left of the above images, you can see the agricultural land in different stages of the production cycle. The IW correctly identified this as a change in land cover, but it is not a change in land use- and since the IW results were compared against my manually-identified areas of land use change, the reported accuracy of its NDVI results are poor.

    The IW was most successful at detecting true change on the CV (71% of change identified), NDSI (94%), and RCVmax (90%), with minimal false-positive changed detected (CV: 8% falsely identified as change; NDSI: 36%; RCVmax 43%). These categories had false-negative (pixels incorrectly identified as no-change) rates of 29%, 6%, and 2%, respectively.

    As previously noted, the NBR and NDVI were less successful at detecting change. The NBR incorrectly identified 68% of unchanged pixels in the study area as showing change, and incorrectly identified 46% of changed pixels as not showing change. The NDVI identified 85% of the true change pixels in the study area, but also identified 90% of the pixels that I marked as unchanged as showing change. This can be explained by the fact that several of the regions I marked as no change were in the agricultural areas in the upper left corner of the image.

   Some limitations of this analysis include:

  • The areas used to compare the accuracy of the IW method were selected manually, and choosing to select different areas of the imagery (for example, places that were obviously no-change instead of areas that were heavily agricultural and subject to seasonal differences in land cover) would change the results.

  • This analysis is not appropriate for determining the accuracy of the Normalized Burn Ratio, since the study region does not include burned areas.

  • The results of this analysis are limited in scope to the San Antonio area, and the accuracy of IW in detecting land cover change may be very different in other geographic locations.

    Based on the NDSI, CV and RCVmax, the IW was successful at identifying areas of land cover change in the Rio Medina area. Most of the deforestation in this area is caused by development of land for residential use.

True Change
1: CV   2:NBR     3:NDSI     4:NDVI       5:MCVmax

Follow

bottom of page