Flood Inundation Mapping of the 2018 Kerala Floods: A ComparativeStudy of Traditional Remote Sensing, Machine Learning, and Deep Learning Methods.

The exponential increase in flood intensity that causes loss of life and economic and structural
damage to their connected environment calls for strategic rescue and response solutions for
risk mitigation. This study focuses on flood mapping using satellite imagery combined with
machine learning (ML) and deep learning (DL) techniques. Remote sensing and Geographic
Information Systems (GIS) serve as vital tools in this process, enabling the effective utilization
of satellite data.
While academics consistently contribute to novel flood mapping approaches, a research gap
that requires a discussion about the comparative performances of these ML and DL
techniques exists, which this paper aims to address. This comparison is crucial as it highlights
the strengths and limitations of each method, contributing valuable insights to the literature on
flood risk management. The study focuses on the Ernakulam District of Kerala, chosen due to
its frequent flooding and the availability of diverse datasets.
The methodology involves the use of satellite imagery for flood analysis, employing an array
of techniques: a thresholding method recommended by the UN-SPIDER Office for Outer
Space Affairs, and statistical ML methods including Random Forest, Support Vector
Classification (SVC), Real AdaBoost, alongside a deep learning semantic segmentation
method, UNet. Modelled using JavaScript and Python languages, the models and the
packages are completely reusable. The dataset comprises two before and after floods satellite
images: the thresholding method uses Sentinel-1 SAR images, and the ML and DL method
uses Sentinel-2 MSI Level 1C, a digital elevation model image from SRTM for feature
engineering, processed to identify flood-affected areas. The data is normalized and cleaned
to account for cloud and missing data before the analysis. Alongside, we sourced the labelled
flood data from the Kerala State Disaster Management Authority (KSDMA) and filtered and
rasterized it on QGIS.
The results emphasize the varied effectiveness of these methods, with Random Forest
outperforming others with a 96.61% accuracy rate. At the same time, the UNet-Linear Model
lags at 75% accuracy, indicating the significant impact of hyperparameter tuning and dataset
size on model performance. This comparative analysis not only delineates the strengths and
weaknesses of traditional and advanced techniques but also sets a precedent for future
studies to build upon an understanding of flood risk management and rapid response
strategies.