Global Fishing Watch &
Google Earth Engine
Grant R. McDermott | AERE 2017 Workshop
1 / 25
About me
- Environmental economist interested in big data and spatial analysis (among other things).
- Incoming faculty at the University of Oregon for AY2017/18.
- Currently finishing up a postdoc at UCSB under Chris Costello.
- Member of the Sustainable Fisheries Group.
www.grantmcdermott.com
github.com/grantmcdermott
gmcdermott@ucsb.edu
@grant_mcdermott
2 / 25
Goals for today
- Introduce you to Global Fishing Watch
- Introduce you to Google Earth Engine
- Show you an application that combines 1. and 2.
3 / 25
Goals for today
- Introduce you to Global Fishing Watch
- Introduce you to Google Earth Engine
- Show you an application that combines 1. and 2.
Meta goal: Encourage you to use these amazing tools yourself.
3 / 25
Global Fishing Watch
4 / 25
Introduction
- Global Fishing Watch (GFW) is a joint initiative between Google, Skytruth and Oceana.
- GFW offers unprecented insight into global fishing activity.
- Includes locations that were previously inaccesible to outside observers.
- Includes a suite of tools for accessing and analysing the data. (More on this later.)
- Economists are under-represented among the set of current GFW research partners.
- We hope that this will change when the data are made public in the next 2-3 months.
- AERE workshop participants are getting a head start!
5 / 25
- Low-hanging fruit.
Where does GFW get its data?
- Ships use AIS (Automatic Identification System) for martime safety.
- Broadcast location to navigate and avoid collisions.
- Includes info on vessel ID, flag, speed, tonnage, length, destination, etc.
- Problems: Location "spoofing". Some fields are incorrectly entered by owner. Raw data contains many mistakes.
- Satellite (and terrestrial systems) can receive and record AIS messages too.
- AIS is big data...
- 22 billion messages from 250k industrial fishing vessels over 2012-2016.
- 20 million messages being added per day (and growing).
6 / 25
- AIS is required on all vessels >300 tons on international voyages. Many countries require smaller vessels to use AIS within their EEZs as well.
- A moving vessel broadcasts a position message every 2 to 20 seconds
- An anchored vessel every 3 to 6 minutes.
- GFW average is 50 satellite positions per vessel per day.
Density of AIS
7 / 25
Processing and classfying the data
- Short version:[1] Cloud computing and machine learning (convolutional neural network).
- Probability that a vessel is fishing,
\(P \in [0,1]\). - What type of fishing is most likely (trawl, long line, purse sein, etc.)
- Probability that a vessel is fishing,
- Training data: 240k hours’ worth of AIS data from 624 vessels (569k AIS positions) that have been hand labelled by fisheries experts and/or validated with logbook data.
[1] Lucius Fox: "I just wanted you to know how hard it was."
8 / 25
- There are actually two CNNs:
- One to identify vessel characteristics, including length, engine power, and vessel type. (Training data: 73,994 vessels matched to official fleet registries, including about 13,500 fishing vessels.)
- A second to detect which AIS positions were indicative of fishing activity. (Training data: 240,000 hours’ worth of AIS data, from 624 vessels, with over 569,000 AIS positions labelled by fisheries experts.)
Processing and classfying the data
9 / 25
GFW dataset
- The cleaned GFW dataset contains 70k likely fishing vessels.
- AIS-equipped vessels account for 75% of offshore fisheries (>100 nm from land).
- Individual vessel tracts.
- Where, when and for how long a vessel was fishing.
- What type of fishing were they doing.
- Other covariates of interest: Flag, tonnage, length, speed, etc.
10 / 25
- Small fraction of the world's ~2.5 million motorized fishing vessels... but it contains a majority of active vessels over 24 metres.
Let's explore the data
- Live session.
- The GFW website is great for (visual) exploration.
- It is meant to serve as tool for policy engagement and citizen science.
- However, it offers limited scope for in-depth analysis and research.
- No API for accessing and downloading the data en masse, or for performing statistical analysis.
- Enter Google Earth Engine...
11 / 25
Google Earth Engine
12 / 25
Planetary-scale analysis
- Google has done the hard work of ingesting (much of) the world's public satellite and remote sensing data.
- Landsat, Sentinel, MODIS, NOAA/NCEP/NCAR, etc.
- Paired with vast, colocated computational power.
- "Bring your code to the data."
- In-browser console editor (Javascript) or Python API.
- Seemlessly interacts with many other products within the Google ecosystem.
- GoogleDrive, Cloud storage, TensorFlow ML library, etc.
- I highly recommend applying for the annual GEE User Summit next year.
13 / 25
- 200 datasets, 5 million images (4k being added every day), 5 petabytes of data...
- Automatic version control plus shareable scripts make for good research practice and reproducible science.
From my inbox last week...
Hey Grant,
It took me 2 days to download the data[...] So I guess the work is reduced from about 6-10 Months to max 7 days.
Thanks a ton. This is amazing.
Best,
M
14 / 25
- Don't just take my word for it: Here's a recent email that I received after suggesting someone switch to GEE.
- It's hard to overstate how valuable that time difference can be.
- For a graduate student, that's the difference between finishing you PhD on time and having to find funding (student debt?) for another year.
- Or, having the time to write a great job market paper versus to an average one.
Let's explore GEE
- Live session.
- GEE is an amazing tool for remote sensing data and spatial analysis (in general).
- For most researchers, it will also be the primary platform for accessing GFW data.
- Caveat: GFW rasters within GEE will lose some information.
- E.g. No individual vessel IDs.
15 / 25
Let's take stock before continuing
- What can we use GFW data for?
- Policy effectiveness and spillovers.
- Interaction between labour/energy/etc. markets and fishing activity.
- Fisheries supply and demand elasticities relative to broader economic cycles.
- Validation of fisheries catch and effort data.
- Continuous data source ready to be paired with natural experiments.
- Etcetera, etcetera.
16 / 25
Let's take stock before continuing
- What can we use GFW data for?
- Policy effectiveness and spillovers.
- Interaction between labour/energy/etc. markets and fishing activity.
- Fisheries supply and demand elasticities relative to broader economic cycles.
- Validation of fisheries catch and effort data.
- Continuous data source ready to be paired with natural experiments.
- Etcetera, etcetera.
- I'll show you an example application of type 1.
16 / 25
Application:
"The Blue Paradox"
W.I.P. with Chris Costello, Kyle Meng and Gavin McDonald
17 / 25
Introduction
- Marine Protected Areas (MPAs) and no-take zones (NTZs) are important tools in marine conservation efforts.
- However, their efficacy remains in question.
- "Paper parks" (e.g. Edgar at al., 2014; Halpern, 2014)
- Data limitations, questionable modelling?
- Most studies ignore incentives and behavourial changes:
- Simple before/after (inside/outside) comparisons of abundance.
- Blue paradox: Delay between MPA announcement and enforcement creates incentive to mine down the stock.
18 / 25
- Edgar et al.'s five features: No take, well enforced, old (>10 years), large (>100km2), and isolated by deep water or sand.
- Current MPA literature does not account for the fact that the announcement itself may trigger a response in fishing effort.
- Blue paradox idea draws inspiration from the Green Paradox (Sinn, 2012) and related literature on preemptive habitat destruction under the E.S.A. (Lueck and Michaels, 2003)
Approach
- Research question: Does an MPA announcement trigger increased fishing activity?
- Focus on the Phoenix Island Protected Area (PIPA).
- Identification: Spatial DiD.
- Fishing effort per km in neighbouring "treated" (MPA) vs "control" (non-MPA) areas.
- NB: Fishers have no incentive to lie about their position pre-enforcement.
19 / 25
- I'm using a draft version of the GFW raster data that hasn't been made public yet.
- Unfortunately, I can't share my full GEE script yet until that's done.
Results (early)
21 / 25
Conclusions (early)
- MPA announcement appears to trigger a preemptive fishing response.
- Could even lead to a long-term decline in fish stocks under certain conditions
- E.g. Surge in fishing pushes a stock below some min. population threshold.
- More work still to be done, but at a minimum suggests that announcement and enforcement gap should be closed.
- Granularity of the data should allow us to test different causal hypotheses (e.g. spatial knowledge vs. quasi-property rights).
- This kind of analysis is only being made possible now thanks to data innovations like GFW and GEE.
22 / 25
Wrapping up
23 / 25
Wrapping up
- GFW provides an unprecedented window into global fishing behaviour.
- GEE dramatically lowers the barriers to entry for working with remote sensing and geospatial data.
- Provides a way to operationalise many of the ideas that we've been talking about today.
- Economists can bring added value to remote sensing data.
- Old questions, new insights.
- Undoubtedly helps to work with remote sensing experts and non-economists.
- Rich possibilities for collaboration.
24 / 25
Thank you!
Questions?
25 / 25