Forest change detection methods demonstrated on Landsat images of tropical forest in Brazil

NR has applied change detection methods on a time series of Landsat images of Brazillian tropical rain forest. The results from this preliminary study are very promising.

Automatic co-registration

The first step is automatic coregistration of repeated acquisitions of the same scene. This was done with subpixel accuracy on most images, and with no more than 1-2 pixel offsets on the few remaining. One of the purposes of this step is to eliminate false change detections due to misaligned images.

The 14 images (green and orange frames) were coregistered to the reference image (black frame). The green frames indicate subpixel accuracy, and the orange frames indicate 1-2 pixel errors.

To get a visual impression of the coregistration accuracy, please see Some results.

Automatic change detection

The second step is change detection on a time series spanning many years. The key point is that change detection should not be done by classifying images from two different dates, and subtract the results. This will lead to misclassifications in the individual images to accumulate in the change map. Rather, the entire time series including the two dates should be analyzed to detect consistent changes.

Experiments on a time series of 24 images spanning 22 years demonstrated that the Markov chain-based method has a great potential, but that it needs to be further improved to provide reliable results. The strength of the method is that it models the sequence of forest changes, thereby eliminating unlikely change events.

Left: forest change from 29 July 1986 till 23 June 2008. Right: forest change from 23 July 2007 till 23 June 2008.

For more change detection results, please see Some results.

For more information contact: Øivind Due Trier

Back to Section for Earth Observation