Multi-temporal burned area mapping using logistic regression analysis and change metrics

Describes a procedure developed for continental-scale mapping of burned boreal forest at 10-day intervals. The basis of the technique is a multiple logistic regression model applied to 1 km SPOT VEGETATION (VGT) clear-sky composites. Independent variables consist of multitemporal change metrics representing 10-day and surrounding 30-day changes in reflectance and in two vegetation indices. The metrics account for seasonal phenological variation by normalizing them to the reflectance trajectory of background vegetation. Three spatial-contextual tests are applied to the per-pixel logistic model output to remove false burned pixels and increase the sensitivity of detection. The procedure was tested over Canada using conventional fire surveys and burned area statistics from the 1998-2000 fire seasons. The area of falsely mapped burns was small (2% commission error over two provinces), and most burns larger than 10 km² were accurately detected and mapped (R²=0.90, P<0.005, n=91). National-level VGT burned areas for 1998-2000 were within 3-17% of fire management agency burned area compiled by the Canadian Interagency Forest Fire Centre (CIFFC).