weecology/deepforest-marine-biodiversity

DeepForest marine biodiversity detector

A single-class ("Object") RetinaNet detector for animals in high-resolution aerial survey imagery of coastal and offshore waters: seabirds, waterfowl, marine mammals, sea turtles, and large fish. It finds animals; it does not identify species. It is the detection stage of the BOEM aerial survey pipeline, where a separate classifier labels each box.

Usage

from deepforest import main

model = main.deepforest()
model.load_model("weecology/deepforest-marine-biodiversity")
model.config.score_thresh = 0.70  # see "Choosing a score threshold"
boxes = model.predict_tile("frame.jpg", patch_size=1000, patch_overlap=0)

Choosing a score threshold

Use 0.70. That is what the BOEM pipeline runs at. The score_thresh of 0.4 in config.json is a training-time default, not a recommendation; on open water it flags far more empty frames.

Do not reuse a threshold tuned on an earlier version of this model. Score distributions differ between checkpoints: this version's scores reach 0.993, the June 2026 version's top out at 0.974. Re-derive the threshold against your own labelled frames whenever the weights change.

Training

Checkpoint	1968e962113f4c57bc6bf5f07fae1b25 epoch 29
Trained	2026-09-03
Initialised from	weecology/deepforest-tree
Data	196,272 patches of 400 px (404,517 boxes, 19,629 empty patches) from 76 flights. USGS and BOEM aerial surveys, including the July 2026 survey review pass
Training	30 epochs, SGD, lr 0.001, batch 64, half of each batch positive patches; rotation, flip and scale augmentation

Evaluation

Measured on two survey flights withheld from training (JPG_20260201_134000 and JPG_20260202_141900): 648 fully reviewed frames, 1,229 annotated animals. A predicted box counts as a match at IoU ≥ 0.4. These are zero-shot results on unseen flights, not validation-split scores.

Score threshold 0.70	This version (Sept 2026)	Previous version (June 2026)
Animals detected (box recall)	57.9%	52.0%
Frames with animals in which at least one was detected	94.2%	87.2%
Box precision	86.6%	89.0%
Highest box recall at any threshold	71.4%	70.2%
Empty-water frames flagged, per 1,000	1.0	1.0

At equal precision (about 85%, threshold 0.60 for both), this version detects 61.0% of animals against 56.9% for the previous one.

The previous version probably trained on these two flights; they were only held out from July 2026. That makes the comparison favour the older model, so the gain shown here is likely an underestimate.

Limitations

• About 29% of animals are never detected at any threshold. Small birds in dense flocks are missed the most: at 0.70 the model finds fewer than six in ten animals overall.
• Breaking waves and foam cause false positives, and so does beach sand. The BOEM pipeline removes land frames with a separate filter.
• Trained on downward-looking RGB frames at survey altitude. Results on other sensors, altitudes, or regions are untested.

Versions

Date	Checkpoint	Notes
2026-06-18	19022856f861432485a1d3911d85e967	First release. Trained 2026-05-28.
2026-11	1968e962113f4c57bc6bf5f07fae1b25 epoch 29	Retrained on the corpus updated with the July 2026 survey review.

To keep using the previous version, pin revision="eb360dc06c2978e390df7e51046039cf36007ce6" in load_model.