Exploration of the influence of ambiguity pixels on image classification reliability

Ambiguity pixels and unreliable classification are easily generated in remote sensing images (RSIs) due to the complexity of landscape sand imaging conditions. At present, few studies focus on the quantification, influence, and propagation mechanism of ambiguity pixels in RSI classification. To investigate these problems, this paper extracts the ambiguity pixels in the image, and explores their influence on classification reliability. First, the pixels of an RSI are classified into ambiguity and unambiguity classes on the basis of uncertainty. Second, the reliability of classification results on ambiguity and unambiguity pixels is evaluated and analyzed using defined reliability indices. Finally, three experiments are designed to reveal the influence of ambiguity pixels on the reliability of RSI classification. Experimental results show that the indicator values of the unambiguity pixels are much higher than those of the ambiguity pixels, illustrating the effects of ambiguity pixels on the reliability of RSI classification results.