We present a deep learning framework based on Convolutional Neural Networks (CNN) to automate the segmentation of marine benthic habitats in the Mediterranean Sea from satellite imagery. The final predictions are based on a consensus from 10 different deep learning models. Hereafter we refere to our model as CAMELE, which stands for Consensus for Automated Marine Ecosystem Labelling and Evaluation.
The models were trained using data from the habitats along the coast of the Balearic Islands. The robustness and generalization capability of CAMELE was assessed by training only with data from one island (Mallorca) and testing the model in the other geographically separate regions (Menorca, Ibiza, Formentera and Cabrera), following a roughly 50/50 data split.
If you use the model to obtain the benthic habitats in other regions of the Mediterranean Sea other than the Balearic Islands we recommend the following considerations:
- Assume an IoU score of the predicitons of about a 60%
- In general, assume a relative error of 25% for measures of Posidonia oceanica extension (surface area). However, for extensions greater than 20 km² you can assume 10% relative error with 75% confidence and 13% error with 95% confidence.