Aquafarm Segmentation for Climate Smart Shrimp

Mapping aquaculture ponds, which is critical for restoration and climate adaptation efforts, is labor and time-intensive. Most existing automated techniques are focused only on spatial location and do not classify these areas based on production intensification (extensive vs intensive). We trained deep-learning computer vision models to predict pond locations and classifications in the Philippines and Indonesia using openly available PlanetScope multispectral satellite data. Our best-performing model achieves an F1 score of 71.8% and sufficiently identifies intensive ponds from other classes despite the inherent class imbalance in the training data. To support model development, we also present our process of creating a human-annotated, expert-validated dataset of pond classifications over the Philippines and Indonesia, which is made readily available and usable for other researchers doing similar work on aquaculture mapping. We've uploaded the model, the training data, and the rollout data. model.zip contains the Pytorch Lightning model checkpoint of the trained model training_data.zip contains TIFF files of the training images and raster masks. It also contains a geopackage of the training polygons that were rasterized into masks rollout_data.zip contains TIFF files of images across Indonesia and the Philippines for aquaculture segmentation. It also contains a geopackage of the predicted aquaculture polygons. We've also uploaded the underlying data behind the site suitability criteria for Climate Smart Shrimp idn_site_scoring_datasets for Indonesian criteria phl_site_scoring_datasets for Philippine criteria Link to Github repo: https://github.com/thinkingmachines/aquafarm-segmentation Link to the project web app: https://ci-aquafarm-mapping.web.app/