Abstract:
Floods, among other natural disasters, pose severe threats to lives and properties.
Accurate prediction and assessment are paramount for effective mitigation.
Conventional flood risk mapping often involves resource-intensive field data
collection, incorporating topographic, hydrologic, and meteorological data.
However, these methods lack the capacity to predict flood probabilities based on
various rainfall scenarios. This study explores the application of machine learning
techniques to address these limitations, focusing on flood risk assessment of
Ratnapura district in Sri Lanka. By leveraging a diverse dataset encompassing
flood records, rainfall data, and satellite imagery sourced from institutions such
as the Disaster Management Center and meteorological observations, we trained
a neural network using Python. The network was executed on cloud computing
platforms, Google Colaboratory, and Google Earth Engine. The results of this
research exhibit considerable promise. The neural network achieved a test
accuracy of 0.7667, indicating its potential for accurate flood probability
predictions following training. Feature importance analysis revealed rainfall as
the most influential factor in predicting flood probabilities, with a relative
importance of 0.191. Other contributors included the normalized difference built
up index (ndbi), clay content, elevation, slope, and drainage density, each playing
a significant role in the predictive model. Additionally, a positive linear
relationship between build-up areas and flood probability was observed.
Nonetheless, it is imperative to recognize that the limited availability of flood and
rainfall data may affect the model's overall accuracy. Despite this limitation, our
study demonstrates the potential for machine learning to significantly enhance
flood risk assessment. This research serves as a valuable step towards more
precise and efficient natural disaster mitigation strategies in the Ratnapura district
and beyond, ultimately contributing to the safeguarding of lives and property in
flood-prone areas.