Classification of ASAS multiangle and multispectral measurements using artificial neural networks

Because the anisotropy of the Earthʹs surface reflectance is strongly influenced by vegetation cover, multidirectional remotely sensed data can be highly effective in discriminating among land cover classes. This article explores the use of multiangle and multispectral data from the Advanced Solid-State Airay Spectroradiometer (ASAS) in land cover mapping using artificial neural networks. A multilayer feed forward neural network is trained to identify five land cover classes in Voyageurs National Park, Minnesota. Multiangle data achieve 89% of accuracy when applied to a single band (774–790 nm), 7-directional imagery and 88% accuracy when applied to multispectral nadir data. Analysis of error using the confusion matrix indicates that the higher classification accuracy is obtained primarily for three classes: deciduous forest, wetlands, and water. The results suggest that 1) directional radiance measurements contain much useful information for discrimination among land cover classes, 2) the incorporation of more than one spectral multiangle band improves the overall classification accuracy compared to a single multiangle band, and 3) neural networks can successfully learn class discriminations from directional radiance data and/or multidomain data.