Techniques for agricultural land cover classification from GIS-enhanced AVIRIS data

Remotely-sensed data have been used for assessment of land cover since remote sensing originated, but not to their full potential partially because of unsatisfactory spectral and spatial resolutions and exclusion of topographic and soil information. Integration of the data acquired by the Airborne Visible-Infrared Imaging Spectrometer (AVIRIS) instrument with geographical information system (GIS) data could overcome these problems allowing the potential of remotely-sensed data to be realized. The AVIRIS data were collected for the Indian Pine Natural Resources Field Station, Purdue University, West Lafayette, Indiana, U.S.A., on 6 September 1991 and have 224 spectral bands, each with 10 nm band widths in the electromagnetic region of 0.4 to 2.45 $\mu$m. An approximately 16 km$\sp2$ study area was chosen. Soil series and field boundaries were incorporated with the AVIRIS data in an attempt to enhance the data. The AVIRIS data were lumped based on the TM spectral characterization to create simulated TM data. Minimum distance and binary-coded and Gray-coded back-propagation neural networks were used to classify all data sets, and the neural network classifiers provided average training and testing accuracies of 99.6% and 80.4%. The maximum likelihood algorithm was used to classify the original AVIRIS data and the simulated TM data because only these two data sets satisfied the Gaussian assumption on which this algorithm is built, and it provided better than 90% training and testing accuracies. Training and testing results were smoothed with pseudo-counts and then were normalized with the iterative proportional fitting procedure. Classification algorithms were compared simultaneously with the Tukey multiple comparison method. After field boundaries were incorporated with the AVIRIS data, classifications were improved. Moreover, the classifications were further improved by addition of soils. Incorporating soils alone did not improve the classifications. The integration of GIS information can improve the classification of AVIRIS data, but the order of integration of GIS information is important