From Gujarat’s monsoon to hydrological forecasting

Hiren Solanki’s Saurashtra flood case study documents the intense rainfall of 16–17 June 2025 across Botad, Amreli, and Bhavnagar. It reports 48-hour totals of 358 mm in Sihor, 386 mm in Umrala, 418 mm in Botad, and 387 mm in Gadhada.

The same account follows Shetrunji Dam as storage approached 94% of capacity by the evening of 17 June. Sentinel-1 radar imagery adds a spatial view of the flooding. Rainfall, reservoir history, and satellite observations each tell a different part of the event; bringing them together makes the sequence easier to understand.

The website puts taluka-level rainfall and reservoir storage on interactive maps. You can inspect daily rainfall, season totals, rainfall as a percentage of the annual average, and the history of individual talukas and reservoirs.

Its data comes from official state reports: the State Emergency Operation Centre’s rainfall reports through GSDMA, and dam reports from Gujarat’s Narmada, Water Resources, Water Supply & Kalpsar Department. A Python ingestion pipeline downloads and parses the reports, stores dated records in MongoDB, and makes them available through a Next.js application with Leaflet maps and history charts.

A practical engineering lesson here is that data freshness and data coverage are separate concerns. A scheduled update can run successfully while an older report remains unavailable. The pipeline checks that a report contains a plausible number of records before replacing a day’s data, and retries missing reports. Gaps in the source should never be interpreted as zero rainfall.

Our paper, One to Three-Day Lead Streamflow Forecast Using Multi-Head Attention With Long Short-Term Memory in Reservoir Regulated Catchment, was published in Water Resources Research in August 2026. I co-authored it with Hiren Solanki, Sayuj Gupta, and Vimal Mishra.

The study focuses on the reservoir-regulated Narmada basin. Reservoirs complicate the relationship between rainfall and downstream flow: water can be stored and released, changing the timing of the river’s response. This is a separate study area from the Saurashtra event above, but it addresses a related question: how can past observations give us useful warning of what comes next?

Table 1 compares SeqLSTM-MHA with multiple linear regression, vanilla LSTM, CatBoost, and bidirectional LSTM for the 2011–2019 evaluation period. The metric is Nash–Sutcliffe efficiency (NSE): 1 indicates a perfect match, while 0 corresponds to predicting the observed mean.

SeqLSTM-MHA records streamflow NSE values of 0.74, 0.73, and 0.71 at one-, two-, and three-day leads. For water level, the corresponding values are 0.85, 0.85, and 0.81. The comparison is useful because it shows how performance changes as the forecast moves further into the future.

These are retrospective evaluation results. Their value is in demonstrating the method under the study’s conditions; deploying a forecast would also require timely inputs, monitoring, and evaluation under changing conditions. A strong aggregate score is a reason to inspect individual flood peaks and their timing, not a substitute for doing so.

For me, the connection between the two projects is practical. The monitor makes observations inspectable. The research asks how those kinds of observations can support useful lead time. Both depend on careful treatment of where the data came from and what it can tell us.