A practical example of monitoring deforestation from space
This post will explore how remote sensing can be used to to monitor and quantify change detection related to deforestation in the Amazon rain forest.
Exploring what change detection is and an introduction to land cover classifications - click the title to keep on reading.
What is change detection?
A technique often used within Remote Sensing is ‘change detection’ which allows multi temporal imagery of two or more dates to be compared to detect and identify spatial changes to the earth’s surface. Change detection can be produced using different techniques, two of which are detailed below:
- Image to image
- Map to map
Image to image compares two remotely sensed images to understand the spatial change that has occurred overtime. By using image differencing or comparing indices, such as NDVI, between two dates, changes at the earth’s surface can be observed and assessed. Image to image provides a relatively quick and simple approach to understanding change however is limited to a true in-depth analysis.
Map to map performs an independent land cover/land use classification over the two images for both dates and compares the difference between the classifications. This produces detailed land cover/land use classifications of an area and can include as many or few classes as the operator sees fit. However this method is dependent on the classification accuracy for reliable results.
Change Detection maps enable a range of end users, such as environmental scientists, urban planners and agricultural workers, to observe changes to the earth’s surface over a long or short period of time. Information from change detection analysis can be used for a variety of projects such as conservation within the Amazon rainforest to be able to identify and quantify the extent of both deforestation and reforestation efforts across vast areas.
Our case study
The first task was to identify a suitable location where deforestation occurs at a rate of which earth observation can observe and measure the change. A location within the Brazilian state of the Amazonas was chosen due to the large amounts of deforestation that have occurred in this region.
Two Sentinel-2 tiles were downloaded from the ESA Copernicus Hub (tile reference - T19LFL) as close to one year apart as possible. A subset of the two images are shown as a true colour composite below.
ARGANS utilised a 'map to map' approach for this study using a machine learning classification algorithm - random forests. Both images were classified using the same number of classes. In this example only two classes were used, 'forest' and 'non-forest', to create a simple example and classification.
The final step is to compare the two classification maps to each other and to assess for change between the two. This led to four final classes in the results:
- Stable non-forest
- Stable forest
- Forest loss
- Forest growth
This project was completed using all Open Source software and can be freely reproduced by users wanting to gain a further insight into how remote sensing change detection can be used to assess the earth's surface.