Evaluation of K-Means Algorithm for Faulted Landforms Extraction and Offset Measurement With an Example From the Eastern Kunlun Fault

• Published in: IEEE access Vol. 13; pp. 47848 - 47859
• Main Authors: Zhou, Shengchao, Lin, Zhou
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Clustering algorithms; Correlation coefficients; Datasets; Deep learning; Earthquakes; Eastern Kunlun fault; Fault diagnosis; Geologic measurements; Geomorphology; Image resolution; Image segmentation; K-means algorithm; Lakes; Landforms; Machine learning; offset measurement; Offsets; Quantitative analysis; Satellite imagery; Satellite images; Seismic measurements; Strike-slip fault; Unsupervised learning; Vectors
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3551121

Accurate offset measurement is crucial for recovering the size of past earthquakes and understanding the recurrence patterns of strike-slip faults. Traditional methods, which rely on manual delineation of displaced geomorphic markers from satellite images, often introduce significant uncertainties. This study aims to develop a more objective and accurate method for identifying faulted landforms and measuring offsets. Although supervised deep learning methods have great potential for image recognition and segmentation, due to the absence of data sets, we apply the K-means algorithm, an easy and practical unsupervised machine learning method with minimal parameters, to extract displaced geomorphic markers. Our research is conducted in the Kusai Lake segment of the Eastern Kunlun Fault using high-resolution satellite imagery. Initially, we identify multiple well-preserved geomorphic markers along the fault traces and manually label some to form an imagery dataset. The K-means algorithm demonstrates its efficacy in extracting landforms clearly when the classification number is set to three. Furthermore, a quantitative comparison of several commonly used unsupervised algorithms confirmed that K-means performs best, achieving a recall of 0.70 and an accuracy of 0.906 on the dataset. The validity of our measurements is corroborated by examining two specific sites around Hongshui Gou. Quantitative analysis comparing our imagery-based measurements with field data reveals a strong agreement, evidenced by a Pearson's correlation coefficient of 0.997. Repeated measurements of the same geomorphic markers across different images indicate that the geomorphic features obtained from the images significantly affect the accuracy of offset measurements. This study underscores the potential of our approach for both extracting faulted landforms and accurately measuring their offsets, emphasizing the importance of assessing the integrity of geomorphic markers when using satellite imagery for offset measurements.