Seasonal land use/land cover mapping: Accuracy comparison of various band combinations

Abstract

The study determines if seasonal spectral variability significantly affects land-cover classification accuracy of Landsat TM in northern New South Wales, Australia, by comparing accuracies in different seasons. Image transformations such as TC, PCA and NDVI were performed and the resultant images were used alone and in combination with original bands for classification. Results from PC1-3, B1-5 and B1-4, 7 were found better than the others; while B1-4 combination was the best in all dates, producing an overall highest accuracy of 96.7% (Kappa 0.96) in January and overall lowest accuracy of 85.7% (Kappa 0.83) in September. Pair-wise tests showed the January accuracy to be significantly higher than in other months at the 0.001 levels and considered the best period for land-cover mapping in the region. The results were supported by analysis of additional January and September images under a range of environmental conditions in three different years.