BlackTide

Model performance

1,867 training / 443 held-out test samples (80/20 random).

Water — research comparison

BlackTide's own marine models (per-pass Sentinel-1 radar, local darkness, wind, incidence angle), trained on SkyTruth Cerulean's reviewed slicks. Not used for the map: at sea scale they raise too many false alarms, so marine events come from Cerulean.

Random Forest

0.91
Accuracy
0.82
Kappa
0.92
Precision
0.93
Recall
0.93
F1

Catches 93% of the 195 human-reviewed oil labels (Cerulean) in the test set.

Wrongly flags 14% of calm-water look-alikes and 9% of clean water as oil.

Confusion matrix (test set)
Predicted oilPredicted cleanActual oil
243
true positive
19
missed spill
Actual clean
20
false alarm
161
true negative

Neural Network (MLP)

0.90
Accuracy
0.80
Kappa
0.92
Precision
0.92
Recall
0.92
F1

Catches 93% of the 195 human-reviewed oil labels (Cerulean) in the test set.

Wrongly flags 11% of calm-water look-alikes and 13% of clean water as oil.

Confusion matrix (test set)
Predicted oilPredicted cleanActual oil
241
true positive
21
missed spill
Actual clean
22
false alarm
159
true negative

Precision–recall

How precision holds up as the model is pushed to catch more spills

Random ForestNeural Network (MLP)

What the Random Forest relies on

Mean decrease in impurity across trees

The land model appears here once enough reviewed land-spill labels exist.