Locating and measuring marine aquaculture production from space: a computer vision approach in the French Mediterranean

TL;DR

This study trains a computer vision model to identify marine aquaculture cages from aerial and satellite imagery, and generates a spatially explicit dataset of finfish production locations in the French Mediterranean from 2000-2021 that includes 4,010 cages (69m2 average cage area).

Abstract

Aquaculture production -- the cultivation of aquatic plants and animals -- has grown rapidly since the 1990s, but sparse, self-reported and aggregate production data limits the effective understanding and monitoring of the industry's trends and potential risks. Building on a manual survey of aquaculture production from remote sensing imagery, we train a computer vision model to identify marine aquaculture cages from aerial and satellite imagery, and generate a spatially explicit dataset of finfish production locations in the French Mediterranean from 2000-2021 that includes 4,010 cages (69m2 average cage area). We demonstrate the value of our method as an easily adaptable, cost-effective approach that can improve the speed and reliability of aquaculture surveys, and enables downstream analyses relevant to researchers and regulators. We illustrate its use to compute independent estimates of production, and develop a flexible framework to quantify uncertainty in these estimates. Overall, our study presents an efficient, scalable and highly adaptable method for monitoring aquaculture production from remote sensing imagery.

Figures (13)

• Figure 1: Example detections of aquaculture cages. Detections of surface square (pink) and circular (red) aquaculture cages identified by our object detection model on aerial imagery of the French Mediterranean. Imagery: Institut national de l'information géographique et forestière (IGN)
• Figure 2: Performance on the French Mediterranean coast. We measure our method's performance in terms of precision and recall at the cage level. The dark line reflects performance of the overall methodology (i.e., the object detection model in addition to the removal of land-based detections, and cage clustering), whereas lighter lines reflect the standalone performance of the object detection model without these sequential post-processing steps. Recall is virtually unchanged across the detection model and post-processing steps.
• Figure 3: Marine finfish aquaculture production locations in the French Mediterranean. Red points indicate the known locations found by trujillo_fish_2012 in their manual survey of 2002-2010 Google Earth. Blue points indicate cage clusters detected by our model during 2000-2021 that are at least one kilometer away from these known locations.
• Figure 4: Predictions for a cluster of cages in the French Mediterranean over time. Imagery: Institut national de l'information géographique et forestière (IGN)
• Figure 5: Marine finfish aquaculture tonnage in the French Mediterranean over time. Model estimates reflect annualized tonnage computed from the area of our predicted cages, with error bars reflecting bootstrap standard errors that incorporate uncertainty in the aerial imagery, in model performance and in the tonnage production factors. Human-in-the-loop (HITL) estimates reflect annualized tonnage computed from the area of annotated cages, with error bars reflecting bootstrap standard errors that incorporate uncertainty in the aerial imagery and in the tonnage production factors. Food and Agriculture Organization (FAO) data reflects average annual production reported to FAO during the period, with error bars reflecting the standard deviation of these values. We include estimates of annualized production that account for the fact that aerial imagery is at times unavailable for some locations.