DeepMet 1
A deep learning model for high-resolution rainfall prediction using INSAT satellite data. Fine-scale precipitation forecasts for the entire Indian subcontinent.
DeepMet 1 is a deep learning-based rainfall prediction model by Fyllo that uses INSAT satellite data to generate precipitation forecasts at 10 km spatial resolution with 6-hour lead times and 1-hour temporal granularity across the Indian subcontinent.
Fyllo's Nowcast Model: DeepMet 1
Real-time precipitation forecasts powered by advanced AI
Why DeepMet 1?
Short-term rainfall forecasting remains a challenge due to the complexity of atmospheric processes and the limitations of conventional numerical weather prediction (NWP) models in resolving fine-scale precipitation events. The Indian subcontinent, with its high rainfall variability and dense population, requires localized and timely rainfall forecasts for disaster preparedness and operational planning.
Traditional NWP models often lack the spatial granularity to predict hyper-local precipitation patterns over India. DeepMet 1 addresses this gap by applying spatio-temporal deep learning to INSAT satellite data, delivering fine-scale rainfall predictions at 10 km resolution — consistently outperforming traditional NWP-based services in nowcasting precision and recall.
Satellite Data Ingestion
Thermal & water vapor channels from INSAT 3DS at half-hourly intervals, with 3DR and EUMETSAT backup
Data Cleaning & Homogenization
Spatial artifact removal, temporal harmonization to hourly resolution, robust feature transformation
Deep Learning Inference
CNN + ConvLSTM / attention-based layers process 6 hours of prior data to generate forecasts
10 km Precipitation Forecast
Hourly granularity precipitation fields for the next 6 hours across the Indian subcontinent
How DeepMet 1 Works
An integrated deep learning architecture combining spatial feature extraction with temporal sequence modeling for accurate precipitation nowcasting.
Convolutional Neural Networks
Spatial feature extraction from satellite imagery captures localized precipitation patterns, cloud formations, and terrain-influenced dynamics across the 10 km grid.
Spatio-Temporal Sequence Models
ConvLSTM and attention-based layers capture rainfall evolution over time, understanding precipitation flow patterns, onset, duration, and termination dynamics.
Attention-Based Post-Processing
Spatial-aware attention heads refine output forecasts, improving prediction accuracy for precipitation cell tracking, intensity estimation, and boundary delineation.
Multi-Source Data Fusion
Integrates INSAT 3DS, 3DR, and EUMETSAT/METEOSAT 9, ensuring reliability during satellite blackouts, maintenance windows, and equinox-related solar intrusions.
Temporal & Static Features
Augmented with hour-of-day, day-of-month, month-of-year temporal features, plus elevation, latitude, and longitude static features for improved convergence.
Real-Time Deployment
Optimized for scalable, real-time inference with API-based integration. Available globally via Fyllo web platform and robust REST API endpoints.
Outperforming Industry Benchmarks
Precision and recall scores computed across 25 farm plots across India from Jan–Jul 2025, validated against Fyllo's in-house Kairo rainfall measurements.
*Rainfall events defined as ≥ 0.2 mm accumulated rainfall. Ground truth validated using Fyllo Kairo device measurements installed at each farm plot. DeepMet 1 consistently leads in recall across the nowcasting horizon.
Applications & Industries
DeepMet 1 provides mission-critical precipitation intelligence across diverse sectors.
Disaster Management
Early warning systems for localized flooding and extreme weather events.
Urban Infrastructure
City drainage management, construction planning, and smart city operations.
Energy Systems
Hydropower reservoir management and renewable energy grid stability.
Logistics & Transport
Route optimization avoiding severe weather blocks, reducing transit delays.
Insurance & Finance
Parametric insurance models and hyper-local agricultural risk assessment.
Agriculture
Precision irrigation scheduling, crop protection planning, and harvest timing.
Future Work & Improvements
We are actively developing the next iteration of the DeepMet architecture, aiming to increase resolution, horizon length, and multi-parameter capabilities.
24–48 Hour Forecast Horizons
Implementing Vision Transformers (ViTs) to extend the predictive horizon.
Multi-parameter Forecasting
Expanding outputs to include temperature, humidity, and wind speed.
NWP Hybrid Integration
Fusing deep learning predictions with physical NWP boundary conditions.
Global 4 km Resolution
Scaling via transfer learning for high-resolution global coverage.
IoT Sensor Integration
Assimilating ultra-dense ground data from Fyllo Kairo/Nero network.
Frequently Asked Questions
DeepMet 1 is a deep learning-based rainfall prediction model developed by Fyllo (Agrihawk Technologies Pvt. Ltd.). It uses INSAT satellite data to generate precipitation forecasts at 10 km spatial resolution with 6-hour lead times and 1-hour temporal granularity across the entire Indian subcontinent.
In benchmarks from Jan–Jul 2025 across 25 farm plots, DeepMet 1 achieved 0.50 average recall and 0.18 average precision, outperforming ECMWF IFS (0.37 recall, 0.15 precision), AccuWeather (0.26 recall, 0.13 precision), and Climacell (0.42 recall, 0.15 precision). Ground truth was validated using Fyllo's Kairo rainfall sensors.
Primarily INSAT 3DS thermal and water vapor channels at half-hourly intervals. INSAT 3DR serves as backup, and European EUMETSAT/METEOSAT 9 provides fallback for blackouts due to maintenance or equinox solar intrusion.
10 km × 10 km spatial resolution with 6-hour forecast lead time and 1-hour temporal granularity. Uses prior 6 hours of satellite data as input. Future revisions target 4 km resolution, 30-minute granularity, and 12–48 hour horizons.
Visualizations are available at fyllo.in/weather-forecast. For API integration, email contact@fyllo.in.
Disaster management, urban infrastructure, energy systems, logistics & transportation, insurance & finance, agriculture, and research & academia.
CNNs for spatial feature extraction, ConvLSTM and attention-based layers for spatio-temporal modeling, and attention heads for post-processing. Augmented with temporal (hour, day, month) and static (elevation, lat, lon) features.
24–48 hour forecast horizons via Transformers, multi-parameter forecasting (temperature, humidity, wind), NWP hybrid integration, global 4 km resolution via transfer learning, IoT integration with Fyllo Kairo/Nero sensors, and open datasets on Hugging Face.
Ready to Integrate DeepMet 1?
Access high-resolution rainfall forecasts via API or explore live predictions on our weather platform.