Inspiration
When a flood hits, roads, bridges and phone lines are often the first things to go. Farming communities are hit twice: the water destroys the crop, and the damp, waterlogged weeks that follow bring fungal and bacterial outbreaks that can wipe out what survived. Our original project was an AI crop disease predictor. We realised the same satellite imagery that shows where a flood hit can also show which fields are about to get sick. So we extended it into one pipeline: see the damage from space, then predict the disease risk that follows.
What it does
- Maps flood extent on farmland from satellite imagery, separating flooded from non-flooded pixels.
- Estimates crop damage by overlaying flood masks on cropland and vegetation indices.
- Predicts post-flood disease risk per field, flagging where fungal or bacterial outbreaks are most likely.
- Highlights isolated areas (farms and villages cut off from roads) so responders and agricultural officers can prioritise.
How we built it
Flood segmentation. We used [Sentinel-1 SAR / Sentinel-2 optical / dataset name] imagery, since SAR sees through cloud cover during storms. We trained [U-Net / DeepLab / other] to segment water. We also computed spectral indices:
$$NDVI = \frac{NIR - Red}{NIR + Red}, \qquad NDWI = \frac{Green - NIR}{Green + NIR}$$
A sharp NDVI drop combined with a high NDWI marks inundated, stressed cropland.
Disease risk model. Our existing disease predictor takes [inputs: vegetation indices, crop type, temperature, humidity, rainfall, days since flooding] and outputs a risk score:
$$R_{field} = f\big(\Delta NDVI,\; t_{flood},\; T,\; H,\; crop\big)$$
Interface. [Dashboard / web map] shows flood extent, damaged fields and risk-ranked areas.
Challenges we ran into
- Cloud cover during floods blocks optical imagery, which is why we leaned on SAR.
- Limited labelled data for flood-plus-disease outcomes. [How you handled it: transfer learning, augmentation, proxy labels, etc.]
- Class imbalance, since flooded pixels are a small share of the scene.
- [Add one or two real challenges you actually faced.]
Accomplishments we're proud of
- A single pipeline linking flood detection to disease prediction
- [Your metric, e.g. IoU / F1 = X on the test set]
- [Anything working end to end that you're proud of]
What we learned
How to work with multispectral and SAR data, why evaluation metrics like IoU matter more than accuracy on imbalanced masks, and how much value comes from connecting two models instead of building them in isolation.
What's next
- Near-real-time alerts using fresh satellite passes
- Crop-specific disease models for more regions
- SMS/WhatsApp advisories for farmers with poor connectivity
- Validation with local agricultural extension officers
Log in or sign up for Devpost to join the conversation.