Navigating the Flood Vulnerability: Data-Driven Analysis for Assam’s Revenue Circles
Zeyu Chen, Bo Long, Qiuyi Wei
Every year, the state of Assam in India faces a serious problem with floods. Nearly 40% of the state is at risk, leading to significant damage to people’s lives, their livelihoods, and the infrastructure. This constant threat of flooding is a major concern for the people of Assam. To face this challenge, significant funds are allocated for disaster relief. However, there is a need to allocate these funds effectively to revenue circles in need and enhance flood preparedness across Assam.
We have developed two classification models to categorize the 180 revenue circles in Assam into six groups based on their vulnerability to floods. These models aim to assess the flood vulnerability of each revenue circle and provide recommendations for allocating disaster relief funds more efficiently. The goal is to enhance overall flood preparedness and resilience throughout the state of Assam, ultimately ensuring better protection for its residents and resources.
We have a dataset that contains information about floods in different areas from May 2021 to August 2023. This data includes 64 different aspects related to floods, and we’ve organized them into six categories to better understand the vulnerability, risk, resilience, and readiness in each area:
- Flood Proneness Variables: These include factors like elevation, slope, and drainage density that help us assess how susceptible an area is to flooding.
- Socio-economic Vulnerability Variables: This category looks at aspects like the age distribution in the population, particularly focusing on older and younger residents who may be more vulnerable during floods.
- Demographic Variables: These variables consider factors like the total population and the ratio of males to females in the area.
- Environmental Vulnerability Variables: This category examines the presence of critical infrastructure such as roads, schools, and hospitals, which can be affected by floods and impact the community’s response and recovery.
- Government Responses Variables: We look at indicators like the number of relief camps and the financial assistance provided by the government in response to floods.
- Damages and Losses Variables: This category includes data on the human toll of floods, such as lives lost, as well as damage to roads and bridges.
By organizing these flood-related variables into groups, we are able to extract valuable insights from each group using a technique called Principal Component Analysis. This helps us gather crucial information from various aspects of data related to Assam state, enabling us to make better decisions. We’ve identified 11 key factors related to floods, each representing a different aspect: damage, government investment, inundation, rainfall, river behavior, population distribution, road conditions, terrain features, land characteristics, drainage systems, and electricity availability.
In the next step of our work, we’ll use these factors to create the classification models. We’ll begin our flood damage analysis by applying the Random Forest method to the results of above extracted indices. This Random Forest model is designed to predict the potential damage that each revenue circle might experience if a flood occurs.
The upper left plot demonstrates our random forest predictions at June, 2022. A darker color indicates a potentially higher level of damage in the event of a flood. The plot in the upper right corner displays the structure of our random forest model. Initially, PCA is applied to the row dataset to extract indices. Subsequently, the random forest is applied using these indices (excluding the damage index) to make predictions for the damage index. The lower table is our random forest model’s outcome. The random forest model seems to achieve a relatively high accuracy rate, approximately 82%. However, it’s important to note that this accuracy is somewhat skewed due to the model’s excellent performance on the very low class, which has a significantly larger number of instances.
We also used a technique called K-means clustering to create a more general flood index, using information from the Principal Component Analysis (PCA) indices. The key differences between the K-means clustering and Random Forest methods is the purpose and data we used to build the model: In the case of Random Forest, it’s focused on predicting the damage index by other index from the PCA result. On the other hand, the K-means clustering model is designed to create one flood index that encompasses various aspects related to flooding, such as flood hazard, flood vulnerability, flood resilience. We incorporated all PCA indices, including the damage index, when generating such flood index using K-means clustering.
Instead of ranking entities, K-means clustering groups them based on their similarities. To give this clustering a ranking aspect, we combined important indices like the government investment index, damage index, inundation index, and rain index into a single score. We then used this combined score to rank the K-means groups. This classification approach is valid because our flood index accurately reflects the level of damage, inundation, and rainfall in specific areas, and it helps identify inconsistencies in previous resource allocation.
The upper left plot demonstrates the general flood risk of Assam state at June, 2022. A darker color (red) suggests a higher flood risk, vulnerability, and damage, indicating a need for increased fund allocation in such areas. Dark blue areas signify locations that received disaster relief funds, while light blue areas indicate no fund allocation. We decided not to use exact allocation values based on the consideration that each area has a different population, and using precise allocated fund values may not accurately reflect the actual situation. The lower left plot demonstrates the real damage scenario as of June 2022. Yellow signifies high damage, while blue indicates low damage, with the intensity of yellow corresponding to greater damage. Similarly, the lower right plot is the actual inundation situation at June 2022, yellow represents high inundation, blue represents low inundation, and the intensity of yellow signifies increased inundation. The plot reveals that the flood index generated by K-means clustering accurately reflects both the flood damage and inundation intensity in the area, suggesting an increase in fund allocation. The actual allocation map further highlights that certain damaged or inundated areas did not receive disaster relief fund allocation. However, It’s worth noting that we used an unsupervised K-means algorithm to create this flood index, and while it’s promising, we plan to conduct statistical verification in the future. For now, we’re evaluating our results through visual comparisons.
In conclusion, the frequent and devastating floods in Assam, India, call for a shift away from relief allocations influenced by politics towards a more data-driven approach. Our Random Forest classifier has shown high accuracy and provided a reasonably good flood vulnerability map. However, it’s important to note that its accuracy is somewhat influenced by its performance on the dominant “very low” class. The K-means results have effectively mirrored the real-world impact of floods, including damage, inundation, and rainfall, while also revealing disparities in past resource allocations.
Nevertheless, our next crucial step is to establish a method for validating these results to ensure their reliability. Looking ahead, in our next phase of work, we can develop a flood prediction model that estimates the likelihood of a flood occurring. We can then combine this prediction with the flood vulnerability information from the Random Forest model to create a more comprehensive flood index. This new index would provide a broader perspective on the flood risk. We can then use this enhanced flood index to cross-verify the flood index generated from the K-means clustering model, ensuring the accuracy and reliability of our assessments.
We hope that our research offers valuable insights, and we hope that other researchers can use these datasets and models for a wide range of purposes in a more comprehensive way. Our goal is to contribute to the well-being of the people of Assam and beyond by making this data available for various applications and initiatives.
This article is part of CivicDataLab and New York University’s Capstone project.
References:
