Dark Fleet Detection — CNN + LSTM + RNN Fusion

A multimodal deep learning framework for detecting dark fleet vessels in the Indian Ocean using Sentinel-1 SAR imagery and Global Fishing Watch vessel detection data.

Model Description

This repository contains five trained model files for a three-branch dark fleet detection pipeline:

File	Branch	Description
`cnn_branch.keras`	CNN	MobileNetV2 SAR vessel density classifier
`lstm_best.keras`	LSTM	Bidirectional LSTM dark vessel detector
`rnn_best.keras`	RNN	SimpleRNN positional anomaly detector
`lstm_scaler.pkl`	LSTM/RNN	StandardScaler for feature normalisation
`fusion_meta_model.pkl`	Fusion	Logistic regression meta-learner

Performance

Branch	AUC	Accuracy	Precision	Recall
CNN	0.941	87.4%	91.0%	81.4%
LSTM	0.911	85.1%	68.1%	80.0%
RNN	0.927	86.0%	69.8%	82.2%
Fusion	0.926	88.0%	78.0%	72.0%

Training Data

CNN branch: SARscope dataset (6,735 Sentinel-1 SAR image patches, COCO format)
LSTM/RNN branches: Global Fishing Watch Sentinel-1 SAR vessel detections, filtered to Indian Ocean (lat −30° to 30°N, lon 40° to 100°E), March 2026 — 16,795 records, 1,617 vessel sequences

How to Use

Install dependencies

pip install tensorflow scikit-learn joblib huggingface_hub

Load models

from huggingface_hub import hf_hub_download
import tensorflow as tf
import joblib
import numpy as np

REPO = "neural-shubh/dark-fleet-models"

# Load all models
cnn_model    = tf.keras.models.load_model(hf_hub_download(REPO, "cnn_branch.keras"))
lstm_model   = tf.keras.models.load_model(hf_hub_download(REPO, "lstm_best.keras"))
rnn_model    = tf.keras.models.load_model(hf_hub_download(REPO, "rnn_best.keras"))
scaler       = joblib.load(hf_hub_download(REPO, "lstm_scaler.pkl"))
fusion_model = joblib.load(hf_hub_download(REPO, "fusion_meta_model.pkl"))

print("All models loaded ✅")

Run inference on a vessel sequence

# Vessel sequence — shape (1, 2, 12)
# Features: lat, lon, presence_score, length_m, fishing_score,
#           matching_score, hour, dayofweek,
#           lat_diff, lon_diff, speed_proxy, impossible_speed

sequence = np.array([[
    [12.5, 72.3, 0.99, 180.0, 0.02, 0.0, 2.0, 5.0, 0.0, 0.0, 0.0, 0.0],
    [12.8, 74.1, 0.97, 180.0, 0.02, 0.0, 2.1, 5.0, 0.3, 1.8, 0.002, 1.0]
]], dtype=np.float32)

# Scale features
sequence_scaled = scaler.transform(
    sequence.reshape(-1, 12)
).reshape(1, 2, 12)

# Predict
lstm_score   = lstm_model.predict(sequence_scaled, verbose=0)[0][0]
rnn_score    = rnn_model.predict(sequence_scaled, verbose=0)[0][0]
fusion_score = fusion_model.predict_proba([[lstm_score, rnn_score]])[0][1]

print(f"LSTM score:   {lstm_score:.4f}")
print(f"RNN score:    {rnn_score:.4f}")
print(f"Fusion score: {fusion_score:.4f}")
print(f"Verdict: {'DARK VESSEL ⚠️' if fusion_score > 0.5 else 'Normal ✅'}")

Architecture

SAR Imagery ──► CNN Branch (MobileNetV2) ──► Spatial Triage Score

GFW Vessel  ──► LSTM Branch (BiLSTM 64→32)   ──► Dark Score ──► Logistic
Sequences   ──► RNN Branch (SimpleRNN 64→32) ──► Dark Score ──► Regression ──► Risk Score

Limitations

Labels derived from AIS matching status, not independently verified dark fleet behaviour
Sophisticated spoofing (AIS transmitting but position falsified) would evade detection
Trained on single-month GFW snapshot — most vessels appear only once, limiting sequence length to 2
Intended as analyst support tool, not autonomous enforcement mechanism

Citation

If you use these models in your research, please cite:

@misc{darkfleet2026,
  author = {Shubh},
  title = {Monitoring Dark Fleets and Maritime Domain Awareness: 
           A Multimodal CNN-LSTM-RNN Framework for the Indian Ocean Region},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/neural-shubh/dark-fleet-models}
}

neural-shubh
/

dark-fleet-models