An object-based burnt area detection method based on landsat images - a step forward for automatic global high-resolution mapping

This study presents an algorithm for automatically mapping burnt areas using high-resolution images. It is applied to the Landsat 4, 5, 7 and 8 Land Surface Reflectance product; specifically, images acquired before and after (or during) the same fire season. It is also possible to extend the timeframe and use reference images acquired in the preceding year. This approach was adopted as cloudiness can make the acquisition of long time series impossible. A second advantage is that it avoids huge data transfers. The algorithm combines traditional, pixel-based image processing (calculation of spectral indexes and image differentiation) with object-based procedures (segmentation, reclassification, neighbourhood analysis) and consists of four steps. First, spectral indices (the Normalized Difference Vegetation Index and Normalised Burnt Ratio), and differences between image layers are calculated. The second is a multi-resolution segmentation, which uses the Normalised Burnt Ratio and near infrared layers. At this phase, masking of clouds, water and deserts takes place using atmospherically-corrected Landsat images. This is followed by the classification of ‘core’ burnt areas based on automatically-adjusted thresholds. The characteristics of the whole image (excluding clouds, deserts and water bodies) are analysed to develop functions that establish these thresholds. The fourth step consists of neighbourhood analysis. This focuses on objects that have not been classified as burnt areas, but whose spatial and spectral distances suggest that they may be part of them. The algorithm was tested in various areas (e.g. Spain, Greece, Siberia, California, Australia and Zambia). Comparisons with manual interpretation show that the fully-automated classification is very accurate (80–100%). The algorithm can be also applied to MODIS and Sentinel-2 data. It was developed within the framework of the Advanced Forest Fire Fighting (AF3) project, and the results have been used for damage and risk assessment.