Dataset: Mask R-CNN and U-Net model and output of coral reef halo measurements based on global multispectral satellite imagery

The Mask R-CNN model was trained using the training data set, thereby automatizing the identification of reef halos from the test set of satellite images and extracting the shape of the reef halos from the imagery background (i.e., extracting both the patch reef and its surrounding halo). The Mask RCNN was trained with 3322 reef halos from 13 AOIs, while the remaining 805 halos from seven AOIs were kept to evaluate the model’s performance.

After object extraction, we automated halo measurement using a U-Net pixel classification model which discriminated the halo (sand ring) from the interior reef patch. A total of 6428 annotations from the training set were used for model training. Annotations consisted of pixel areas manually classified by users, divided into “patch reef” and “halo” classes, respectively.

The total of 6428 annotations used for the classification model was derived from splitting some of the 4127 original annotations (see related dataset). This process was implemented as part of the model training to improve pixel classification by creating diverse and multiple samples.These shapefiles were created as part of the training process and can be recreated from the original annotations already provided by processing them within the ArcGIS using the .dlpk model (as explained also in the github tensorflow).