Investigator accuracy: A center-weighted metric for evaluating the location accuracy of image segments in land cover classification

• Published in: International journal of applied earth observation and geoinformation 2023-08, Vol.122, p.103402, Article 103402
• Main Authors: Wu, Bingxiao, Gu, Zhujun, Zhang, Wuming, Fu, Qinghua, Zeng, Maimai, Li, Aiguang
• Format: Article
• Language: English
• Subjects: Accuracy evaluation; Digital image processing; Image segmentation; Location accuracy; Machine learning models
• ISSN: 1569-8432
• DOI: 10.1016/j.jag.2023.103402

[Display omitted] •A new evaluation metric reflects the location accuracy of single image segments.•A simple and efficient embedded patch extraction method is proposed.•A parallel evaluation scheme is proposed for the distance-to-center weight per pixel in single patches in image segmentation results. Accuracy evaluation is an essential step in validating image segmentation results. Existing metrics, such as overall accuracy and F1-score, mainly concern the range and consistency of the semantic labels of image segments, which may not be sufficiently sensitive to detect missing or segmentation errors in small patches. To address this issue, this study proposes the investigator accuracy (IA) metric, which focuses on the location accuracy of single patches by evaluating the capture accuracy of their near-center subregions and category weight to determine the image segmentation quality. Before evaluating the IA metric, we optimize the grayscale dilation algorithm to separate each identified patch from the image without converting the data format and then distinguish each patch as embedded or nonembedded. Next, we use an iterative grayscale erosion approach to assess the distance-to-center weight, which is a crucial parameter for evaluating the IA of each pixel in a single patch. In addition, we apply a parallel analysis scheme to improve the efficiency of the IA evaluation. The results indicate that the capture accuracy of near-center subregions and the category weight of a single patch affect its IA. Unlike commonly used metrics, the IA is independent of the area ratio and the number of patches belonging to multiple landcover types. The output of the intermediate analysis steps can be used to produce thematic maps showing the distribution density of target patches.