The paper presents the results of forest cover change mapping in two study areas in Bulgaria (in mountainous and plain-hilly terrain) for period of about 20 years. A comparison was made of two approaches for classification of multitemporal SPOT HRV/HRVIR data with 20 m spatial resolution. The first approach was the post-classification comparison, i.e. pixel-by-pixel comparison of forest/non forest maps produced by separate classifications of the images from the two ends of the time period. The second approach was a direct multitemporal classification of an image stack comprised of the two-date image data. Following international guidance, instead of counting pixels in the map to obtain the area of forest loss and gain, the areas were estimated by applying an unbiased estimator to sample data collected by stratified random sampling. The map was used to stratify the study areas. Producer’s, user’s and overall accuracy were also estimated using the sample data. A comparison of accuracy and area estimates, and confidence intervals of estimates, showed that the map produced by direct multitemporal classification was more accurate. It yielded consistently higher class-specific accuracies than the map made by post-classification comparison. As expected, the accuracies of the change classes – forest disturbance and reforestation – were significantly lower than that of the stable classes regardless of the change detection approach. Finally, practical issues and guidelines for future forest change detection studies were discussed.
Multi-sensor change detection for within-year capture and labelling of forest disturbance
