Model card

Overview

• Task: multi-label climate-hazard classification plus continuous severity regression per forecast unit, per horizon.
• Hazard classes (8): cold wave, drought, fire, flash flood, flood, heat wave, severe local storm, tropical cyclone.
• Input: 15-channel spatio-temporal tensor (Sentinel-1 VV/VH, Sentinel-2 indices, ERA5-Land surface and soil state, precipitation and wind from Open-Meteo).
• Architecture: 3D depthwise-separable convolutional network with a shared trunk and two heads — 8-way softmax classification and severity regression.
• Bundle: FP32 TFLite, roughly 0.75 MB, sha256-manifested in Models/VERSION.json and refreshed by the pipeline.
• Execution: server-side for scheduled runs, and in the visitor's browser through TensorFlow.js WASM for interactive and offline use.

Intended use

HazardNet is built for agricultural extension officers, NGOs and disaster-management practitioners in Bangladesh who need a consistent national picture at district and sub-district level, and for researchers who need a transparent, citation-friendly baseline. It is a prioritisation and situational-awareness tool: it helps you decide where to look next and which district to call.

Out of scope — do not use it for this

• Issuing or replacing official warnings. Cyclone signals, flood bulletins and evacuation orders come from BMD, FFWC and the local administration.
• Any decision where a false negative costs lives without a human check — the model is a screening layer, not an authority.
• Property, insurance or compensation determinations; the severity index is not a damage assessment.
• Sub-district micro-siting. The outlook is published per administrative unit; it does not resolve individual fields, roads or embankments.

Known failure modes

• Cloud and scan gaps: Sentinel-2 optical input is missing under persistent cloud, and Sentinel-1 revisit is not instantaneous, so a fast-onset event between passes can be under-observed.
• Reanalysis lag: ERA5-Land is a reanalysis product with its own latency; the most recent hours are estimated, not measured.
• Proxy physics: the independent severity track uses formula proxies. A proxy can be wrong in the same direction as the model, which weakens the cross-check as an error detector.
• Class imbalance: rare hazards (for example, severe local storms) have fewer training examples than floods, so their calibration is weaker.
• Long horizons: the 15-day outlook inherits the uncertainty of a deterministic weather window and should be read as an outlook, not a prediction.
• No INT8 edge bundle exists: a 3D-convolution quantisation path is closed by the runtime at FP32-only precision (ADR 0007). Older documents that mention an INT8 sub-2 MB bundle are out of date.

Fairness, privacy and data protection

The model consumes environmental data only — no personal data, no farm-level records and no household information enter training or inference. Account data used by the dashboard (profile, saved districts, connectors) is separate from the model and is not used as model input.

Reporting a problem with a prediction

If a district card looks wrong — a missed event, an over-stated severity, a stale date — send it to the team with the district, hazard, horizon and date. Ground-truth reports are the only route by which the platform can eventually publish skill metrics.