How a CNN‑Transfer Learning Model Boosts Mumbai Monsoon Forecast Accuracy by 400% Using 36 Stations

Recent increases in extreme rainfall frequency and intensity during Mumbai's monsoon season have exposed the limitations of global forecast systems, whose 25 km grid resolution cannot capture local weather features. To address this gap, researchers from IIT Bombay and the University of Maryland developed a hyperlocal prediction model that combines convolutional neural networks (CNN) with transfer learning (CNN‑TL).

Data Sources and Station Selection

The study uses two data streams: (1) model data from the Global Forecast System (GFS) covering June 2015 to September 2023 at 0.25° × 0.25° resolution, and (2) observational data from 36 automatically weather stations (AWS) operated by the Mumbai Municipal Corporation, recording 15‑minute intervals from 2006 to 2023 (excluding 2014). Stations with extensive missing data were excluded, retaining only those with at least five consecutive years of complete records during the monsoon months (June‑September).

After quality control—removing gaps, selecting correlated variables (precipitable water, precipitation, relative humidity, temperature, pressure), and spatial matching—the GFS grid data were aligned with station observations to form input‑output pairs for model training.

Model Architecture

The CNN downscales coarse GFS forecasts to station‑scale by ingesting four‑dimensional inputs (selected meteorological variables) and extracting spatial‑temporal features through convolutional layers, followed by flattening layers that align predictions with observed rainfall.

Transfer learning fine‑tunes the pre‑trained CNN: the base convolutional layers remain frozen while the upper layers are updated using extreme‑rainfall samples (above the 95th and 99th percentiles). Early‑stopping and regularization are employed to avoid over‑fitting on the limited extreme‑event data.

Event Synchronization and Louvain Clustering

To capture the spatial synchrony of extreme rain events, the study applies the Event Synchronization method, which measures temporal coincidences between station rainfall time series. This approach outperforms traditional linear correlation for detecting abrupt, nonlinear extreme events.

From the resulting synchronization matrix, the Louvain algorithm clusters stations exhibiting similar synchrony patterns, revealing coherent regional rain‑fall zones within the city.

Performance Evaluation

Multiple statistical metrics—correlation coefficient (CC), root‑mean‑square error (RMSE), and false‑alarm rate (FA)—were used to compare the CNN‑TL model against raw GFS forecasts and a baseline CNN without transfer learning.

For overall rainfall prediction, the downscaled CNN improves spatial accuracy, raising CC values and lowering RMSE across the city. Incorporating transfer learning further stabilises predictions for moderate‑to‑high intensity rain and reduces systematic under‑estimation.

In extreme‑rainfall tests (95th and 99th percentiles), the transfer‑learned model increases forecast accuracy by 60 % to 400 % relative to GFS, delivering reliable 1‑ to 3‑day lead‑time predictions and markedly reducing false alarms.

Spatial analysis shows that the model not only predicts individual station rainfall more accurately but also preserves the clustered synchrony patterns identified by the Louvain analysis, confirming its ability to reflect city‑wide rain dynamics.

Broader Context

The paper notes that, alongside academic interest, the Indian government is accelerating AI development through the IndiaAI Mission, procuring thousands of GPUs and partnering with domestic AI firms to build sovereign large‑language models. This policy backdrop underscores the strategic importance of AI‑driven climate forecasting for national resilience.