Examining the Capability of Supervised Machine Learning Classifiers in Extracting Flooded Areas from Landsat TM Imagery: A Case Study from a Mediterranean Flood

• Main Authors: Gareth Ireland, Michele Volpi, George P. Petropoulos
• Format: Article
• Language: English
• Published: MDPI AG 2015-03-01
• Series: Remote Sensing
• Subjects: flooded area mapping; remote sensing; support vector machines; regularised kernel Fisher’s discriminant analysis; Landsat TM; Evros River; Greece
• Online Access: http://www.mdpi.com/2072-4292/7/3/3372

This study explored the capability of Support Vector Machines (SVMs) and regularised kernel Fisher’s discriminant analysis (rkFDA) machine learning supervised classifiers in extracting flooded area from optical Landsat TM imagery. The ability of both techniques was evaluated using a case study of a riverine flood event in 2010 in a heterogeneous Mediterranean region, for which TM imagery acquired shortly after the flood event was available. For the two classifiers, both linear and non-linear (kernel) versions were utilised in their implementation. The ability of the different classifiers to map the flooded area extent was assessed on the basis of classification accuracy assessment metrics. Results showed that rkFDA outperformed SVMs in terms of accurate flooded pixels detection, also producing fewer missed detections of the flooded area. Yet, SVMs showed less false flooded area detections. Overall, the non-linear rkFDA classification method was the more accurate of the two techniques (OA = 96.23%, K = 0.877). Both methods outperformed the standard Normalized Difference Water Index (NDWI) thresholding (OA = 94.63, K = 0.818) by roughly 0.06 K points. Although overall accuracy results for the rkFDA and SVMs classifications only showed a somewhat minor improvement on the overall accuracy exhibited by the NDWI thresholding, notably both classifiers considerably outperformed the thresholding algorithm in other specific accuracy measures (e.g. producer accuracy for the “not flooded” class was ~10.5% less accurate for the NDWI thresholding algorithm in comparison to the classifiers, and average per-class accuracy was ~5% less accurate than the machine learning models). This study provides evidence of the successful application of supervised machine learning for classifying flooded areas in Landsat imagery, where few studies so far exist in this direction. Considering that Landsat data is open access and has global coverage, the results of this study offers important information towards exploring the possibilities of the use of such data to map other significant flood events from space in an economically viable way.