Abstract: Based on the unsupervised pixel-based Spectral Rule-based Classifier (SRC) algorithm, an effective Modified Spectral Rule-based Classifier (MSRC) was proposed considering the influence of atmospheric correction on spectral reflectances of remote sensing images. MSRC modifies rule sets according to ground object spectrum curves and spectral indices, optimizes spectral categories through refined and supplementary rules as well as modified thresholds. The Landsat 8 remote sensing images of islands (Jiapeng, Qi'ao) and coastal zones (Quanwan, Huidong) in the Pearl River Delta were chosen as the experimental data. The band reflectances and ground object spectrum curves before and after atmospheric correction process were contrasted. Classification results and accuracy of the MSRC algorithm were analyzed and compared with those of other six classification algorithms: the original SRC algorithm, Minimum Distance Classification (MDC) algorithm, Maximum Likelihood Classification (MLC) algorithm, Support Vector Machine (SVM) algorithm, Neural Network Classification (NNC) algorithm and spectral indices-based classification methods. The overall accuracy (OA) of MSRC algorithm using experimental data were respectively 87.66%, 82.38%, 77.67% and 80.05%, which were all higher than those of the original SRC algorithm, MDC, MLC and spectral indices-based classification methods, and closed to the accuracy of supervised algorithms (SVM and NNC) without the requirement of manually labelling the training dataset. MSRC performs well in land-cover type scenarios of islands and coastal zones on Landsat 8 multispectral remote sensing images.

Key words: remote sensing images, classification, spectral rule, atmospheric correction