Why does accuracy assessment and validation of multi-resolution-based satellite image classification matter? A methodological discourse

Abstract

This study presents a methodological discourse about how to validate the reliability of thematic maps derived from multi-resolution satellite-based image classification. Besides, the paper examines unbiased estimates of accuracy assessment using known sampling units. Landsat and spot images were used for lulc thematic layer extraction. These thematic layers together with reference data extracted from panchromatic aerial photo interpretation and ground survey were used as input datasets for accuracy assessment and validation analysis. For each lulc unit, a minimum of 50 reference samples were derived using a stratified random sampling scheme. Consequently, error matrices were generated to validate the quality of the 1973, 1995 and 2007 lulc maps. To improve sampling biases introduced due to the stratified random sampling reference data collection scheme, accuracy assessment indices including the producer’s, user’s and overall accuracy as well as Kappa coefficient of agreement were adjusted to the known areal proportion of map categories. The computed overall accuracy, corrected for bias using known marginal proportions of the 1973, 1995 and 2007 thematic layers were 88.12%, 89.95% and 92.27%, respectively. Also, 81.20%, 82.17% and 83.11% of Kappa coefficient of agreement were achieved from the 1972, 1995 and 2007 classifications, respectively. The findings show that high resolution aerial photos are good sources of  reference datasets in the absence of historical ground truth data for accuracy assessment analysis and the lulc classifications fulfilled the minimum of lulc classification standards of overall accuracy and Kappa coefficient of agreement. Consequently, all the lulc classifications could be used as an input for policy options for integrated land resource management practices in the watershed studied.