Normalisation of within-scene optical depth levels in multispectral satellite imagery using National Forest Inventory plot data.

An automated method for normalisation of local variations of atmospheric optical depth levels within Landsat TM and ETM+ images has been developed and implemented in an operational system for the production of large area forest type classification and forest parameter maps. A prerequisite for the method, is the existence of a large number of evenly distributed geo-referenced field plots within each scene to allow for accurate modelling of the spectral response. First, a linear index from the radiance values observed in the six reflective bands of the Landsat sensor is calculated. Using pixels for dark vegetation only, the index image is then smoothed to a low resolution image which is proportional to the local variations in haze. In a second step the haze index image is calibrated by linear modelling of the haze index for each plot against the residuals obtained from an empiric spectral response model. The calibrated haze image is then subtracted from the original image, resulting in a TM image with homogeneous haze levels. The method contributes in combination with other correction procedures, both to better results for classifications and estimation results for large areas, and to a better visual appearance of the images. It has been applied operationally on more than 50 satellite scenes over Sweden.