Flash flood susceptibility prediction mapping using Remote Sensing and Machine Learning

Abstract

Pakistan has witnessed recurring, devastating floods attributed to extreme rainfall, causing loss of life and significant economic consequences. Studies have been conducted with regards to flood prediction mapping in Pakistan using various remote sensing and GIS techniques, the gap which has been identified is that the findings for the previous studies conducted do not include a simulation aspect, only include results obtained using past data. In our study we used a combination of GIS tools, remote sensing and machine learning techniques to generate susceptibility maps for our region of interest. We have focused on the area of Swat District in Khyber Pakhtunkhwa, Pakistan, a flash flood-prone region, as our study area. Datasets were collected through reliable sources such as Climate Research Unit, Open Topography, Global Flood Database, Geological Survey database, Ensuring the quality of the data, preprocessing was applied to cater for outliers, null data and redundant values. The central research question pertains to flood susceptibility prediction within Swat District. The Frequency Ratio method was employed for feature extraction, demonstrating the influence of factors such as slope, flow accumulation, LULC, distance to rivers and precipitation patterns. After analysis, a wide range of factors were examined to understand the vulnerability of the area to sudden floods. This resulted in the development of a set of characteristics that portrays the regions susceptibility to flash floods. Machine learning models, such, as Random Forest (RF) k Nearest Neighbors (KNN) Support Vector Machine (SVM) and XGBoost (XGB) were then applied on these features. The results of those models based on hyperparameter tuning indicate high performance of those models with recall value of 0.90, f1-score of 0.95 and AUC ROC values of 0.99 for RF and XBG and 0.96, for KNN and SVM showcasing their capabilities. In conclusion, the results obtained used

x

List of Figures

a combination of weighted features, from the Frequency Ratio method and machine learning models to create a map showing the susceptibility of Swat District to floods. The predictions generated by simulating rainfall patterns across our study area we can predict which regions are prone, to flooding and estimate the damage caused by such events. This work will offer valuable guidance and aim to enhance flood risk management strategies, ultimately contributing to the preservation of lives and the reduction of flood-related damages.