| ---
|
| license: mit
|
| tags:
|
| - image-classification
|
| - aquaculture
|
| - shrimp-disease
|
| - tabular-classification
|
| library_name: pytorch
|
| ---
|
|
|
| # AquaSense — Trained Artifacts
|
|
|
| Model weights for **AquaSense**, a multimodal early-warning system for shrimp
|
| (udang) disease detection and pond risk scoring. Built for RISTEK Datathon 2026
|
| by Tim 3 Plenger.
|
|
|
| Application code: https://github.com/baihaqidawanis/dataton-semifinal
|
|
|
| ## Contents
|
|
|
| | File | Module | Description |
|
| |---|---|---|
|
| | `manifest.json` | all | Single source of truth for metrics, thresholds, and the M2 form schema. The backend serves this via `GET /meta`. |
|
| | `m1_vision.pt` | M1 | Image classifier over `{healthy, black_gill, wssv}`. Contains `state_dict`, `classes`, `backbone`, calibration temperature, and the cost-tuned decision threshold. |
|
| | `m2_risk.pkl` | M2 | Tabular outbreak-risk model. Bundles two variants: `model`/`fitur` (39 survey features) and `model_web`/`fitur_web` (23 features the web form collects). **The application must load the `_web` variant.** |
|
| | `m3_meta.pkl` | M3 | Fusion metadata. |
|
|
|
| ## Intended use
|
|
|
| Decision support for shrimp farmers — **not** clinical confirmation. Outputs
|
| should be combined with direct inspection and local farm SOPs.
|
|
|
| ## Training data
|
|
|
| - **M1**: ShrimpDiseaseBD + BD Fish & Shrimp Disease (public). Split *per shrimp
|
| individual*, not per photo, to prevent leakage across train/val/test.
|
| - **M2**: WSD Affected Shrimp Farmers (233 ponds, Bangladesh).
|
|
|
| ## Known limitations
|
|
|
| - Black Gill recall is 54.5% — the model misses roughly half of Black Gill cases.
|
| - M2 ROC-AUC is moderate (0.70 on the 23-feature web subset); treat it as a
|
| screening signal, not a diagnosis.
|
| - Multimodal fusion is a transparent noisy-OR rule, not a trained meta-learner:
|
| no public dataset pairs photos with pond surveys for the same pond, so a
|
| learned fusion could not be validated.
|
| - All training data is from Bangladesh; generalisation to Indonesian ponds is
|
| untested.
|
|
|
| Decision thresholds are tuned by **economic cost** (false negative ≈ 223× the
|
| cost of a false positive), not accuracy — the models deliberately over-flag.
|
|
|
