Inferring seagrass meadow resilience from self-organized spatial patterns

This archive contains the code, trained models, and derived datasets associated with the study “Inferring seagrass meadow resilience from self-organized spatial patterns”.

The archive provides a self-contained companion resource to the manuscript. It includes the computational workflow used to generate synthetic seascapes from a mechanistic model of seagrass self-organization, preprocess the simulated patterns, train convolutional neural network (CNN) models, and apply the trained models to empirical habitat cartography from the Balearic Islands.

The archive contains:

• the full code repository used in the study;

• trained deep-learning models for pattern classification and effective mortality estimation;

• geospatial outputs of predicted meadow states and inferred effective mortality for each region;

• summary tables of predicted pattern classes;

• a pixel-level table of inferred effective mortality and predicted pattern type;

• a patch-level dataset of structural descriptors, inferred mortality, and predicted pattern type;

• a text file mapping the numerical pattern labels to their ecological class names.

The central aim of the study is to infer the condition of Posidonia oceanica meadows from self-organized spatial patterns. To do so, the workflow combines ecological theory, machine learning and empirical benthic cartography. The resulting models are used to predict discrete meadow states and a continuous effective mortality parameter from spatial pattern alone.

This archive is intended to facilitate reproducibility, reuse of the trained models, and inspection of the derived spatial products.

Notes on archive contents:

• models/ contains the trained models in HDF5 (.h5) format together with metadata files.

• patterns_summary.csv summarizes the number of predicted patterns of each type in each region.

• patterns_mortalities_all.parquet contains pixel-level predictions, including inferred effective mortality and predicted pattern type.

• measures_dataset.csv contains patch-level structural descriptors together with inferred mortality and predicted pattern type.

• patterns/ contains GeoTIFF and GeoJSON files of predicted pattern classes for each region.

• mortality/ contains GeoTIFF files of inferred effective mortality for each region.

• pattern_label_mapping.txt provides the translation between numerical class labels and ecological pattern names.

The empirical habitat cartography used in the study was obtained from regional data sources from the Balearic Islands. Access to the original raw habitat data depends on the corresponding providers.