Reviving Bengaluru’s First Line of Flood Defence

Bengaluru is paving its way to becoming a concrete jungle, choking its lakes and flooding its streets.

Team: Shashank V., Sakshi N., Akshit V.

01. Abstract

Once a city of lakes long before it became a city of tech parks, Bengaluru is now under the magnifying glass for its gross mismanagement of water bodies. As a result, the city faces recurring flood events exacerbated by weed encroachment and predatory urbanization.

To dissect this story, this study analyzed 162 lakes (> 0.5 ha) over a 2020-25 year period using multisource satellite data including Sentinel-1, SAR, GEE Dynamic World, NASA GPM, SRTM, and MERIT Hydro. The goal was to understand the specific factors driving the ecological murder of these lakes.

A Machine Learning algorithm trained on engineered features achieved a cross-validation score of 50.7% with a Mean Absolute Error (MAE) less than 5%. This explains approximately 50% of flood frequency variance, identifying biological factors (30% importance) and imperviousness (22%) as the primary drivers of risk.

To communicate these results, a Streamlit decision dashboard was developed to enable pre-emptive actions and achieve flood resilience through evidence-based geospatial intelligence.

02. Executive Summary

Bengaluru flood crisis gets blamed on cloudbursts and extreme rain events. The real tragedy, however, is that we have been treating our lakes like isolated heritage parks or selfie spots and not like the connected drainage network engineered by Bengaluru's natural geography.

To make matters worse, we are always reactive –– waiting for water to flood our streets and then taking to Instagram to post memes like: “Hello, how are you? I’m under water.” History often repeats, and what was once a Lake City morphs into Venice of potholes come the monsoons. And of course, politics promises and BBMP de-silts once every five years, courtesy the election season, and yet, we swim to navigate the city, as if it's a mandatory fitness exercise.

To overcome this complex challenge, we have built a hybrid satellite diagnostic dashboard employing tools like Google Earth Engine (GEE), USGS topography, Copernicus Sentinel 2, and Merit hydrography datasets to name a few.

We did not just look at water pixels from satellite but measured the prominent physical properties of the basin like the lake surface area capacity, surrounding imperviousness, encroachment by invasive weeds, rainfall intensity, slope-elevation, and urban stress indices.

We then fused GradientBoosting with Bayesian Regression to create a model that doesn't just predict but gives a 100% safety coverage interval –– all lakes in the sample were within the predicted safety margins. Our model identifies vulnerable areas with 3.74% Mean Absolute Error (MAE) –– tight enough for a BBMP officer to not miss the target.

Key Performance Metrics

  • R² of 0.64 We can explain ~ 65% of why our neighborhood floods just using land cover and topography data based on GEE data sets.
  • 100% Prediction Interval The model ensures that we don't miss the actual floods since all lakes fall within our safety margins.
  • MAE of 3.74% Very small error, hence a BBMP municipal officer can confidently approve de-silting budget without fearing consequences.

Additionally, we are working on a dashboard that can turn this geeky satellite data into simple and effective communication tool so that stakeholders can finally go from disaster response to predictive maintenance.

Therefore, our model proves that Bengaluru's floods are predictable consequences of bad urban design suffering from predatory encroachment, supplemented by administrative neglect. Hence, by placing actionable satellite intelligence into stakeholders’ hands, we aim to achieve flood resilience in Bengaluru.

03. Methodology

01

Lake Identification & Spatial Filtering

  • We first had to identify lakes within Bengaluru, for which we used Open Street Map (OSM) data via the osmnx library to extract lakes and their boundary coordinates as polygons.
  • We then filtered out smaller lakes using a threshold >0.5 ha to focus on lakes with considerable hydrological impact on the city.
$$ \text{Area}_{\text{Lake}} > 0.5 \text{ ha} $$
02

Rainfall Intensity & Saturation Modeling

  • We followed a multi-year analysis to evaluate how rainfall intensity and urban concrete density (Imperviousness) interact at lake level and its impact on floods.
  • For each layer, we created a 200-meter circular buffer to capture the immediate inflow zone where impervious surfaces have the most direct impact on runoff surge into lake basin.
  • We use NASA GPM (Global Precipitation Measurement) IMERG V07 dataset (0.1° resolution) to extract flood triggers instead of total annual rainfall.
  • We integrated 30-minute precipitation rates into 24-hour totals (peak_30min_intensity_mm).

  • To account for soil saturation (antecedent moisture) we calculated sum of rainfall over 3 consecutive days, and then annual maxima of this value was extracted as the primary indicator for flood events.
  • The max 30-minute rate (max_3day_rain_mm) recorded in a year was extracted to account for cloudburst events, which can trigger urban flash floods.
$$ \text{Saturation} = \sum_{d=1}^{3} \text{Rainfall}_d $$
03

Urban Stress (Imperviousness)

  • We quantified the degree of urbanization surrounding each lake using the Google Earth Engine (GEE) Dynamic World (V1) dataset.
  • This dataset identifies land cover types by applying deep learning models to Sentinel-2 imagery.
  • We isolated the built label (Index-6) and classified these specific pixels as impervious (impervious_fraction).
  • We extracted the mean of this mask within a 200-meter buffer to estimate the percentage of land preventing natural seepage and accelerating runoff.
$$ I_{\text{fraction}} = \frac{\sum \text{Pixels}_{\text{built}}}{\sum \text{Pixels}_{\text{total}}} $$
04

Hydrological Flow & Drainage Network

  • We also need to quantify the upstream catchment area that drains into a lake to accurately model the risk. Hence, we use MERIT Hydro dataset (Multi Error Removed Improved Terrain) which has a resolution of ~90 meters at equator.
  • We define upstream accumulation area as the total surface area in km^2 that drains into a specific 90-meter pixel (flow_accumulation).
  • But since these accumulation values can have a large range, we make use of log-10, thus mapping both major drains and smaller gullies.
  • We also identify flow direction by assigning a code (1,2,4,8,16,32,64,128), each representing a compass direction (flow_direction).
  • This data was calculated for each lake by using their centroids and drawing a 200-meter buffer around it.
$$ \text{Normalized Flow} = \log_{10}(\text{Flow}_{\text{acc}}) $$
05

Topographic Slope and Elevation

  • Here, we use NASA SRTM (Shuttle Radar Topography Machine) GL1 data set (~ resolution 30 meters), which is the global standard for topographic analysis.
  • Also, unlike other analysis where we use circular buffers around lakes, this data extraction uses exact lake boundaries to ensure data accuracy.
  • We made use of two topographic parameters to characterize our lake basins:
a. Elevation

We extracted this directly from the SRTM Elevation Band:

  • Used to estimate the lake's position (valley, plateau, or sea level) within Bengaluru's integrated drainage network.
  • Lakes at lower elevations act as sinks for urban runoff, directly increasing flood frequency.
b. Slope

We calculated the max rate of change in elevation between pixels:

  • Steeper slopes accelerate water flow velocity, heightening flash flood risks during cloudburst events.
  • Applied Reduce Regions function to localize and extract mean slope and elevation at the lake level.
$$ \text{Slope} = \max\left(\frac{\Delta \text{Elevation}}{\Delta \text{Distance}}\right) $$
06

SAR-Based Flood Frequency Analysis

1. SAR Methodology & Data Acquisition
  • We make use of SAR (Synthetic Aperture Radar) instead of optical sensors since it is cloud penetrating, allowing monitoring during monsoons.
  • We used the Sentinel-1 Ground Range Detected (GRD) collection, operating in the C-band for high-resolution radar backscatter.
  • Vertical-Horizon (VH) Polarization was selected to distinguish between smooth water surfaces and rough terrestrial terrain.
  • The model operates over a five-year period (Jan 2020 – Dec 2025) to capture climatic variations and urban changes.
  • The Region of Interest (RoI) was defined by a 200-meter buffer around lake centroids.
2. Inundation Detection & Baseline
  • Water acts as a smooth reflector, resulting in low backscatter returns.
  • A fixed threshold of -20 dB was used; pixels lower than this were classified as water (All Water Mask).
  • To isolate temporary floods from permanent water, we filtered observations for dry months (Jan–Mar).
  • Areas identified as water in ≥ 70% of dry season observations were classified as permanent.
$$ \text{Flood Pixels} = (\text{All Water Mask}) \cap \neg(\text{Baseline Mask}) $$
3. Quantifying Frequency
  • Individual flood masks were aggregated to represent probability of flooding for every 10-meter pixel.
  • Calculated as the ratio of flood events to total satellite observations.
4. Lake-wise Extraction
  • Data was aggregated to lake levels to make it actionable for BBMP officers.
  • The mean flood frequency was extracted within the 200-meter buffer as the target variable for machine learning models.
07

Data Fusion & Predictive Modeling

1. Transformation & Digital ID Creation

We transform raw geospatial data into clean metrics to create a digital ID for each lake, forming the foundation for our Machine Learning (ML) model. The fused datasets include:

  • Meteorology: Rainfall data.
  • Biological Health: Water/weed ratios and surface aerial capacity.
  • Topography: Mean elevation and slope.
  • Hydrology: Hydrological flow and urban imperviousness.
  • Target Variable: SAR-derived flood frequency.
2. Engineered Physical Indices

Indices were engineered to improve the precision of estimating the physical reality of lakes:

Catchment to Surface Area Ratio (CSR)

Identifies lakes overwhelmed by their drainage network compared to their size.

$$ \text{CSR} = \frac{\text{Flow Accumulation Area (km}^2)}{\text{Potential Lake Area (ha)}} $$
Urban Stress & Bio-Clogging
  • Urban Stress: Product of impervious fraction and flow accumulation to estimate surface water retention due to concrete cover.
  • Biological Clogging: Area occupied by weeds relative to surface area, identifying lakes where choking is highest.
3. Predictive Modeling (GBR)
  • We utilized a Gradient Boosting Regressor (GBR), ideal for capturing non-linear relationships and feature interactions by sequentially correcting errors.
  • We configured the model with 300 estimators, a learning rate of 0.05, and a max depth of 4 to prevent overfitting.
  • Input features were categorized into Meteorological (Max 3-day rain), Urbanization (Imperviousness), and Bio-physical factors (CSR, Slope, Clogging).
Training & Validation
75:25 Split — Evaluated via R-squared (R^2) and Mean Absolute Error (MAE) with Cross-Validation for reproducibility.
4. Vulnerability Ranking & Stakeholder Action
  • We then ranked variables to identify and communicate to decision makers which factors most influence urban floods.
  • Further, we generated an action plan by identifying the top 15 most vulnerable lakes.
  • Finally, we compared predicted vs. actual flood frequency to provide actionable intelligence for preventive measures.

04. Findings

3.1

The Hydrological Drought

While on paper, the city's water coverage may appear to increase from 36.4% in 2020 to 45.7% in 2025, this is very shallow and not surprising considering the rare “triple-dip” La-Nina episodes that marked the 2020-2023 period, followed by a neutral El-Nino in 2024 and a La-Nina again in 2025.

This prolonged cooling of the central and eastern Pacific Ocean significantly altered the regional weather patterns in India, contributing to higher-than-normal monsoon rainfalls. This can be corroborated with evidence gathered from Landsat and Sentinel satellites.

Almost 20% of Bengaluru's large lakes (>1ha area) are dry with water occupying less than 10% of the lake’s area. We see a strong correlation (0.94) between a lake’s aerial extent and its weed cover. While this correlation is relatively weak (0.68) with its water cover as shown below. Therefore, as the lake’s size increases, encroachment by weeds increases almost proportionately, effectively nullifying the lake’s ecological role as a reservoir.

3.2

Biological Choking & Invasive Dominance

The most severe threat to Bengaluru's lakes is not just the lack of water, but Biological Choking. The average lake in the city is over 53% covered by weeds. Furthermore, 52 out of 162 lakes are classified as "Severely Choked" with greater than 75% weed cover.

Specific Dead Zones have emerged where major water bodies, including Chikka Beguru Lake and Doddabidirakallu Kere, have reached 100% weed dominance with 0% detectable open water. These lakes have completely lost their ecological and flood-mitigation functions.

Top 5 Most Choked Lakes (Average 2020-2025)

Lake Name Dominance Status
Chikka Beguru Lake 1.00 Dead Zone
Doddabidirakallu 1.00 Dead Zone
Pattandur Agrahara 0.99 Severely Choked
Vidyaranyapura Kere 0.99 Severely Choked
Mallasandra Lake 0.99 Severely Choked

Top 5 Most Dried Lakes

Lake Name Fill Ratio Status
Chikkalasandra Lake 0.00 Dry
Konanakunte Lake 0.00 Dry
Nagarbhavi Thorai 0.0007 Dry
BEL Lake 0.0011 Dry
Kengeri Lake 0.0013 Dry
3.3

Exposed Beds

Across all the lakes, approximately 25% to 34% of the lake area is neither water, nor weeds. This represents exposed, dry lake beds.

These are the most critical zones for intervention, as they are hydrologically inactive and serve as the primary entry points for illegal activity and the permanent loss of lake area.

3.4

The Inverse Relationship of Water and Weeds

The most significant statistical finding is the strong negative correlation (-0.71) between water trends and weed trends. This establishes a "Switch" Mechanism: for every hectare of open water lost per year, there is a corresponding and rapid expansion of weed cover.

Lakes like Hulimavu Lake and Kalkere Lake are currently in a state of "Hydrological Collapse," losing water at rates of 3–4 hectares per year while simultaneously being consumed by invasive weeds.

Top 5 Fastest Drying Lakes

Lake Name Water Cover (Ha) Status
Hulimavu Lake -4.40 Rapid Collapse
Kalkere Lake -3.14 Rapid Collapse
Yellamallappa Chetty -2.23 Critical Loss
Hoodi Lake -1.22 High Stress
Madiwala Lake -0.83 High Stress

Fastest Weed Expansion

Lake Name Weed Cover (Ha) Status
Jakkur Lake +4.10 Invasive Surge
Kalkere Lake +3.43 Invasive Surge
Rachenahalli Lake +2.85 Aggressive Growth
Attur Lake +2.41 Aggressive Growth
Dore Kere +1.45 Active Choking
3.5

Slope as a Double-Edged Sword

The relationship between topography and lake health in Bengaluru reveals a fascinating paradox. While specific cases like Vasanthapura Lake support the "Flushing Effect" theory—where steep slopes act as temporary conduits leading to "all or nothing" hydrological profiles—the overall statistical average across the 162-lake dataset reveals a more complex reality.

In these Siltation Hotspots, steep surrounding terrain facilitates high sediment loads that settle, shallowing the basin and providing a permanent substrate for weed expansion. This explains the negative correlation observed in outliers like Dubasi Palya Kere, where steepness leads to seasonal drying.

However, the Statistical Counter-Narrative shows that on average, as the slope increases, health metrics actually improve. Lakes in the High Slope (>5°) category maintain the highest average Water Ratio (0.57), significantly higher than Mid-Slope (0.32) or Low-Slope (0.41) lakes.

This Correlation Flip (+0.14) suggests that steeper terrain is often associated with deeper, better-maintained water bodies. High-slope standards like Sankey Tank and Yediyur Lake utilize their deep basins to capture runoff effectively, while their steep geometry limits the "littoral zone," resulting in the city's lowest Weed Ratio (0.23).

The true "danger zone" is the Low-Slope (<2°) category. These basins are most prone to "clogging" (0.38 weed ratio). Without the gravitational "flush" of steeper terrain, they become stagnant nutrient sinks where water movement is insufficient to prevent massive weed proliferation.

Therefore, while steep slopes create localized risks of siltation in degraded outliers, the "Flushing Effect" generally acts as a cleaning mechanism. In deep, steep-sloped basins, the higher velocity of water inflow combined with limited shallow-water margins creates a "Goldilocks" environment: enough energy to move water, but enough depth to store it—effectively resisting the weed-induced clogging seen in flatter landscapes.

3.6

The Flow Accumulation Paradox & Elevation Gradients

Contrary to expectations, high flow accumulation does not always equate to better lake health. Low Flow Choking occurs in catchments where stagnant nutrients lead to a total takeover by vegetation, resulting in an average of 59.8% weed dominance. Conversely, High Flow Resilience is observed in lakes with steady water throughput, showing the highest water-to-potential ratios (~40%).

This is further complicated by the Elevation Gradient. Lakes at higher elevations (900m+) show significantly better Water Retention Ratios (~38%) compared to those in lower valleys (~24% at <840m). Lower-elevation lakes bear the brunt of the city's cumulative pollution and siltation, causing them to shallow out and lose storage capacity.

3.7

The Directional Health Gradient: South-West vs. North-East

There is a clear directional bias in lake health across the city. The South-West Stressor is particularly evident, as lakes draining towards the South-West and South-East are significantly more stressed. These water bodies show the lowest average water ratios (~29%) and the highest weed ratios (~62%), suggesting these corridors bear heavier nutrient or sewage loads and are more prone to stagnation.

Conversely, North-East Resilience is observed in lakes with a North-East flow direction. These are the healthiest in the city, maintaining an average water ratio of 48% and the lowest weed dominance at 42%.

3.8

The "Starved Giants" (Hydrological Mismatch)

Analysis has identified 37 large lakes (>5 ha) that are effectively "Hydrologically Starved." These lakes possess massive potential footprints but suffer from extremely low flow accumulation (<0.1 km²). Consequently, they receive almost no natural drainage and depend almost entirely on direct rainfall.

Large examples include Yelahanka Lake (96 ha) and Madiwala Lake (82 ha). When these show high water ratios, it often indicates they are being artificially fed by treated sewage. Conversely, Kengeri Lake and Hulimavu Lake are actively drying out as their natural "flow-veins" have been severed by intense urban development.

3.9

The "True Flood" Intensity & Negative Health Correlation

Flood frequency is measured as the percentage of time water is detected outside the normal dry-season baseline. While the average lake experiences flooding in 10.3% of observations, the "Very High Risk" group peaks at nearly 50%.

A significant negative correlation (-0.27) exists between flood frequency and water ratio. This reveals a critical paradox: drier lakes actually flood more frequently. Having lost the ability to absorb water gradually, these degraded basins act like concrete sumps that flash-flood instantly during rain.

3.10

Flash Flood Vulnerability Index (FFVI)

By identifying "Flash Flood Sinks"—lakes with low baseline storage but high inundation rates—we have pinpointed the city's most dangerous hydrological pressure points. These lakes are typically dry but inundate rapidly and violently.

Flash Flood Sink Priority List

Lake Name Flood Freq % Status
Doddabommasandra Kere 49.6% Highest City Risk
Kogilu Lake 36.0% Extreme Inundation
Bellandur Lake ~31.0% Terminal Failure
Ramapura Kere 28.9% Catchment Overload
Varthur Lake ~24.0% Terminal Failure
3.11

The Urbanization-Storage Trade-off

There is a stark inverse relationship between surrounding urbanization and hydrological health. Lakes in low urbanization areas maintain high water levels (avg 49% water ratio). However, in extreme urbanization zones (>80% imperviousness), water ratios plummet to 22%. These lakes are effectively being "paved over," isolated from natural groundwater recharge zones. Urbanization doesn't create more floods, it creates deadlier ones. Low-urban lakes flood often but gently. Whereas, extreme-urban lakes flood rarely but catastrophically.

Urban Intensity vs. Hydrological Health

Category Water Ratio Weed Area (ha) Status Label
Low Urban 49.2% 13.67 Natural Wetland
Moderate Urban 39.7% 2.72 Fragile Buffer
High Urban 34.5% 1.46 Stressed Sink
Extreme Urban 22.2% 1.05 Critical Failure
3.12

Rainfall Intensity: The "Stress Tests"

The city's infrastructure faced a 2022 Peak with a 3-day rainfall average of 108mm, explaining the massive flood events of that year. Current trends show a 2025 Surge, with averages reaching 100.4mm, ensuring that vulnerable lakes remain under extreme pressure.

While there is a moderate correlation (0.41) between rainfall and lake water area, the "Creeping Concrete" trend—with city-wide impervious fractions rising from 48.6% to 51.2%—means that even moderate rain now triggers disproportionate flood risks.

3.13

Five-Year Urbanization Trend (2020–2025)

The city-wide Average Impervious Fraction has seen a steady rise, peaking at 51.2% in 2024. This reflects a persistent "Creeping Concrete" effect where natural absorbent soil is being replaced by impermeable surfaces at a rate that the cascading lake system can no longer sustain. Therefore, Bengaluru paved its way to flash floods.

05. Decision Dashboard

Explore the Bengaluru Lakes Revival: Decision Support Dashboard → here

06. Model Evaluation

6.1

Performance Analysis

The model demonstrates high predictive reliability, successfully capturing the complex non-linear interactions between rapid urbanization and extreme weather events across Bengaluru’s hydrological network.

R² = 0.6464
Predictive Accuracy
MAE = 3.74%
Error Margin
R² = 0.5071
Cross-Validation

The model explains approximately 65% of the variance in flood frequency. In urban hydrology, where local blockages introduce significant noise, an R^2 above 0.6 is considered a robust indicator of systemic trends. The dip in cross-validated performance suggests that certain "outlier" lakes possess unique engineering interventions that deviate from city-wide trends.

6.2

Feature Importance

A. The "Sponge" is Clogged (Basin Health: 40.3% Total)
The most significant predictor is Biological Clogging (30.2%). Lakes have lost their buffering capacity because they are physically choked with invasive weeds and silt. Even large lakes suffer reduced effective storage, leading to premature overflowing.

B. Surface Transformation (Urban Stress: 37.3% Total)
Imperviousness and Urban Runoff Stress statistically confirm the "Creeping Concrete" theory. As catchments are paved, the speed and volume of water exceed natural drainage thresholds.

C. Rainfall is No Longer the Only Trigger (Rainfall: 9.3% Total)
Surprisingly, Rainfall Totals represent less than 10% of model importance. Infrastructure is now so sensitive that moderate rainfall triggers floods if the lake is clogged and the area is heavily paved.