Thresholding is a widely used technique for classifying populations, particularly in image binarization. However, it has its limitations. Traditional methods can struggle to accurately estimate class sizes, especially when populations have significantly different sizes or variances. This can lead to incorrect estimates of the total number of data points within each class. One major issue with conventional classification methods is that they often produce unsatisfactory results, despite decent classification performance. To address this problem, this paper proposes a novel thresholding method designed to improve the accuracy of class size estimation. The authors test the method by comparing it to existing approaches using both simulated and real-world remote sensing satellite data, as well as truth data. It is important to note that the distribution of each class must be known to apply the proposed method. The comparison experiments in this paper assume that each class follows a normal distribution. The proposed method achieved high counting accuracy in pixel classification, allowing for precise calculation of the number of pixels in the area occupied by the target class from an image. Additionally, because the values of producer accuracy and user accuracy are close when using the proposed method, discrepancies in the evaluation of classification results from the perspectives of both the creator and the user are minimized. Even when the number of pixels in each class and their standard deviations are biased, the proposed method can accurately calculate the number of pixels. Unlike general statistical methods, which only evaluate the number of pixels, the proposed method also provides information on which pixel corresponds to which class.