Journal article clarifies how Global Fishing Watch fishing data is generated
- By Rocío Joo
- Published
Response published to a study that claimed our algorithms overestimate fishing activity
At Global Fishing Watch, we strive to enable scientific research and drive a transformation in how we manage our ocean by increasing transparency of human activity at sea. Sharing accurate and useful fishing data is key for those goals, so we took it seriously when a scientific paper suggested our algorithms overestimate fishing activity.
The recent study by Hintzen et al. compared fishing events from our API with gear deployment observations from self-sampling programs on board two pelagic trawling fleets in the Northeast Atlantic Ocean. The authors concluded that our fishing events overpredict effort for these fleets after finding strong agreement in fishing locations and differences in the number of events and their durations. However, the number of events and durations from our data are not directly comparable to the number and duration of hauls. The journal has published our response to this study, clarifying how our fishing data is generated as well as its intended and non-intended uses.
The vessel behaviors representing fishing activity (e.g. searching, setting, hauling or soaking gear) can vary substantially by fleet and region, making it challenging to accurately differentiate the behaviors based on tracking data only. For this reason, our general fishing model was not designed to distinguish between these different fishing behaviors. Instead, the model aims at providing a global dataset of fishing activity that accurately represents its spatial distribution across fisheries.
Our response also explains how the two fishing datasets in our APIs, currently labeled apparent fishing events and fishing effort, are computed. Fishing events, the dataset used in Hintzen et al., was devised to create a visually logical representation of related fishing activity on the Global Fishing Watch platform by grouping consecutive positions of predicted fishing. Therefore, the fishing events dataset should not be expected to match single hauls of gear deployment.
Another claim by Hintzen et al. was that relying on Global Fishing Watch fishing datasets may lead to “incorrect analyses of the impact of fishing on numerous ecosystem aspects,” citing several studies that used Global Fishing Watch data. We reviewed the studies that were cited and found that most of them used our data to analyze the spatial distribution of fishing, which was shown to be accurate.
To mitigate the misuse of our data and increase its usefulness, we are implementing the following technical and documentation improvements:
- We have recently published documentation to better guide the appropriate use of our fishing events and effort datasets, and we have initiated user research to explore alternative names for these two datasets.
- We are changing the architecture of our model from a convolutional neural network to a transformer-based neural network. The transformer architecture can capture long-range dependencies effectively through attention mechanisms, machine learning techniques that enable each position in a sequence to attend to all other positions. This allows the model to capture relationships between different parts of the sequence in a way that is not practical with convolutional neural networks.
- The data used to train this new version of the general fishing model covers wider geographic distributions and different fleets around the world.
In addition to the general fishing model, we are working on gear-specific models, whose predictions would be better proxies of gear deployment and could be more easily translated into effort metrics. A static dataset of predicted events for drifting longlines, distinguishing between setting and hauling behaviors, is publicly available for free to registered users through our data download portal and an updated dataset will be available in 2026. A trawler model, currently in development, aims to differentiate bottom and midwater trawling. A provisional static dataset of predictions is available upon request and will be published on our website next year.
We believe that closer collaboration with fishing experts and comparisons with high-quality logbook or observer data are essential for improving the fishing activity data that we share with the world. We aim to create spaces to discuss better ways to estimate and represent fishing activity in the future and will do our best to reach out to fishing experts from around the world. Please contact us if you are interested in collaborating with data or expertise to improve the quality of these models and the datasets, or if you are leading any fishing research project that you would want us to contribute to.
