---
license: mit
tags:
  - oil-spill-detection
  - maritime
  - sar
  - ais
---

# POSEatSea — model weights

Weights for the POSEatSea maritime oil-spill detection and vessel attribution
console (SIH Problem Statement 26143, NTRO).

| File | Architecture | Notes |
|---|---|---|
| `best_sar_model.pth` | U-Net + MiT-B2 encoder, 5 classes | Expects 512×512 input. **See caveat below.** |
| `trajectory_lstm_baseline.pth` | LSTM(6→128)×2 + Linear(128→2) | Mauritius AOI, ~60 s ping cadence |
| `ais_phase1_autoencoder.pth` | Autoencoder 11→16→8→4→8→16→11 | Threshold 1.104481 |
| `ais_phase1_scaler.joblib` | StandardScaler, 11 features | **Mandatory** — the model returns nonsense on unscaled input |

Code: https://github.com/23f2003521/SIH2026

## SAR checkpoint

Expects **512x512** input. Running inference at another resolution does not
degrade gracefully -- it returns incoherent noise.

An earlier export of this file carried a trained encoder with an *untrained*
decoder and head, and emitted noise on every input. The current file is
trained through (BatchNorm `num_batches_tracked` = 7530) and reproduces the
reference outputs at 0.81 oil IoU.