Effect of spatial resolution on the accuracy of satellite-based fire scar detection in the northwest of the Iberian Peninsula

In this work, an empirical study was carried out to evaluate the impact of the spatial resolution of satellite images on the accuracy and uncertainty of burned area detection using classification techniques based on neuro-fuzzy NF models. The study area was situated in the northwest of the Iberian Peninsula, where in the summer of 2006, a large number of fires occurred, razing a surface area of more than 100,000 ha. A set of 12 zones containing a burned area in their central part were selected. Landsat Thematic Mapper TM, Terra Moderate Resolution Imaging Spectroradiometer MODIS, Advanced Very High Resolution Radiometer Local Area Coverage AVHRR-LAC, and Advanced Very High Resolution Radiometer Land Long Term Data Record AVHRR-LTDR images with a spatial resolution of 30, 250, 1100 m, and 0.05° ∼5000 m, respectively, obtained on 20 August 2006, were used. An NF classifier at pixel level for every image was constructed, taking into account only the spectrum bands red and near-infrared NIR common to all of them. The results in the study region suggest that burned areas of ∼1200 ha could be detected with a mean relative error less than 30% only in the MODIS image. In the case of the LAC and LTDR images, a minimum burned area size of >1800 ha and >3600 ha, respectively, is required to find similar errors. Burned areas greater >3600 ha can be detected in MODIS imagery with a mean relative error of ∼15%. A regression model of commission and omission error intervals compared with spatial resolution is presented. The conclusion is that in regard to the conditions of the study area, both error intervals increase symmetrically and linearly with the logarithm of the pixel size. The results also suggest that red and NIR spectrum bands could be used to detect burned area in post-fire images in Iberia, but with a relative error depending on burned area size for different spatial resolutions.