Detection of flood-affected areas using multitemporal remote sensing data: a machine learning approach

Mapping areas affected by flooding requires an efficient and an effective mapping methods because the flood detection methods provided diverse results and bias. A reliable comparison technique between one method and other methods is needed in this case. Furthermore, in-depth analysis regarding the use of data in detecting flood-affected areas also needs to be carried out to improve the performance of flood detection methods. This research aims to compare the application of the machine learning methods, namely random forest (RF), classification and regression tree (CART), and support vector machine (SVM), to detect flood-affected areas. Sentinel-1 Synthetic Aperture Radar (SAR) Ground Range Detected (GRD) data are used in this research through monotemporal and multitemporal approaches. Furthermore, this research also accommodates the use of band combinations as input data. Therefore, this research can provide a highly comparable comparison between machine learning methods. Data from Official Statistics, FloodScan flood-affected area maps, and online news are used as validation data for the detection results of flood-affected areas. The research results show that RF has higher performance than CART and SVM, with an F1-score of 91.54%. Compared to monotemporal data, the use of multitemporal data in flood-affected area detection is proven to increase the performance of the RF, CART, and SVM models, respectively, by 5.20%, 6.34%, and 5.96% on average. The utilization of Sentinel-1 band combinations for machine learning offers an alternative for developing flood detection models. This research is useful for the government to formulate policies related to flood disasters, especially in disaster risk assessment strategies, the distribution of aid, and food security.