How to detect and classify aquaculture ponds with Artificial Intelligence?

Photo of author

By Milthon Lujan

sample model predictions 1

Researchers Joshua Cortez and JC Nacpil, members of Climate Change AI (CCAI), have published a tutorial on using deep learning, a type of artificial intelligence, for the detection and classification of ponds used in aquaculture based on their production intensity.

The tutorial will help you understand the relevance of aquaculture ponds in climate adaptation, map and classify aquaculture ponds based on their production intensity, and use PyTorch Lightning to train and deploy a semantic segmentation model.

The guide is aimed at data scientists, researchers, and developers at all levels within computer vision, Earth observation, and related geospatial industries.

Importance of pond mapping

The researchers emphasize that mapping ponds used in aquaculture is essential for improving environmental monitoring, conservation efforts, and climate adaptation.

Monitoring changes in the production intensity of aquaculture systems can provide insights into how areas are impacted by the aquaculture industry.

In this way, the tutorial can benefit “those working in aquaculture, conservation, climate adaptation, and land use planning by learning more about the climate impact of aquaculture ponds and how we can reliably map and classify them using satellite images.”

Climate Smart Shrimp Program

The tutorial authors collaborated with Conservation International (CI) to support their Climate Smart Shrimp (CSS) program, which aims to intelligently design aquaculture ponds to effectively restore mangroves and mitigate environmental risks.

In the tutorial, the researchers share some of the methods they developed to quickly identify extensive and intensive aquaculture ponds using deep learning. “This ability to quickly identify these viable areas was key to helping CI in their site prioritization efforts for CSS and accelerating their efforts to implement green-gray aquaculture at pilot sites,” they highlighted.

Basic requirements

The tutorial is designed for a broad audience as it covers basic concepts. However, the authors emphasize that users should have a basic knowledge of Geographic Information Systems (GIS) and machine learning in processing satellite imagery, computer vision concepts, and PyTorch or equivalent machine learning frameworks.

The researchers used freely accessible images from Planetscope provided by Norway’s International Climate and Forest Initiative (NICFI) and the “Aquaculture Classification” database for the Philippines and Indonesia developed by Thinking Machines Data Science.

Semantic segmentation for pond classification

Semantic segmentation is the process of classifying each pixel in an image. For this study, the authors employed three relevant classes: intensive ponds, extensive ponds, and background (if not an intensive or extensive pond).

The researchers used the following classification:

  • Extensive ponds rely on the natural productivity of the water body and are, therefore, less densely populated and often do not have machinery present.
  • Intensive ponds are characterized by higher yields per unit area due to capital investments in technology such as aerators.


The researchers detail the methodology step by step and recommend breaking down the code into small fragments, using text cells, and avoiding duplicate code through modularization.

They also caution that due to the use of a semantic segmentation model, the model does not explicitly distinguish between adjacent ponds of the same class, especially if the boundaries between ponds are relatively thin.


“We demonstrated how we can use deep learning with PyTorch and Lightning to detect and classify aquaculture ponds in satellite images. We hope the methods described here are useful, especially for conservation and environmental monitoring efforts,” they concluded.

The researchers highlight several future work opportunities to leverage the model and enhance contributions:

  • The U-Net model is only designed for semantic segmentation, which does not explicitly delineate individual ponds.
  • A similar approach can be applied to produce aquaculture classifications in other Southeast Asian countries and worldwide, where mangrove deforestation due to aquaculture remains an issue.
  • Researchers can explore time series of satellite images to monitor the number of intensive ponds over time.

Reference (open access)
Cortez J. and JC Nacpil. 2023. Aquaculture Mapping: Detecting and Classifying Aquaculture Ponds using Deep Learning. CCAI Community Authors.

Leave a Comment