Physics-Informed Neural Networks for Jamuna River Sandbar Dynamics

The Jamuna river in Bangladesh is one of the most braided, hydrodynamically violent fluvial networks on earth. Sandbars (চর) emerge, shift, and dissolve within seasonal flood cycles, directly impacting millions of riverside families in Sirajganj, Tangail, and Bogura.

The Blindspot of Pure Satellite Modeling

Standard deep learning computer vision models applied to multispectral satellite feeds suffer from recurring cloud-cover occlusions during peak monsoon season (June to September). When flood risks are highest, optical imagery is frequently obscured.

To resolve this, Logicdock Studio constructed a hybrid methodology:

1. SAR Phase Decomposition: Utilizing Sentinel-1 C-band SAR backscatter data to penetrate monsoonal cloud cover.
2. Physics-Informed Loss Constraints: Enforcing Saint-Venant hydraulic conservation equations into the latent spatial loss function.
3. Hyper-Local Ground Truth Calibration: Calibrating flow rates using observations recorded by local boatmen and river gauge staff.

# Conceptual Hydrological Loss Constraint
def hydraulic_continuity_loss(pred_depth, pred_velocity, topography):
    flux_grad = compute_spatial_divergence(pred_depth * pred_velocity)
    temporal_diff = compute_temporal_rate(pred_depth)
    return mse(flux_grad + temporal_diff, 0.0)

Grounded Predictions for Community Early Warning

Unlike centralized models that output abstract flood risk indices, our open toolkit generates route-passability maps specifically tailored for village schools, rural healthcare boats, and emergency grain storage.

“The river does not arrive as a disaster. It arrives as a change in the route to school.”

We believe climate AI must be built in direct communion with the delta’s lived realities.