Resources for Remote Sensing
- Haven King-Nobles
- 12 hours ago
- 9 min read
Summary
This post provides an overview of everything we’ve done so far in our effort to develop satellite-based water quality monitoring. In the hope that they might be useful to other interested parties, we also provide a list of resources, including:
The forecasting and Chl-a remote sensing models our Upwork contractors have completed
Water quality data we haven’t previously shared
Some of our team’s unfiltered thoughts on this project
Lastly, we invite those interested in this project to sign up to our Mini Remote Sensing Newsletter.
Our Quest for Water Quality Remote Sensing
As we have frequently discussed, FWI is interested in developing satellite-based remote sensing in order to assess the water quality of fish farms in India at scale. This would significantly improve the cost-effectiveness of our farm program, enabling us to meet our 2026 Goal.
We have sought this technology for over a year now, and thus far to no avail. This post lays out everything we have attempted so far, including the code for some of the models we have had developed, in order to maximize the chance that someone is able to crack it.
Our Previous Approaches For Remote Sensing
The following are the three approaches we have utilized so far:
Previous Approach #1: Collaboration with an Aquaculture Company
We began our work to test the potential for remote sensing in February 2024 when we partnered with Captain Fresh—an Indian seafood company experienced in using satellite imagery in the aquaculture industry—to conduct a proof-of-concept study.
This study—conducted between February and March 2024—involved 20 fish farms, all part of the Alliance for Responsible Aquaculture (ARA) in the Eluru region of Andhra Pradesh. The study focused on six key water quality parameters: dissolved oxygen, ammonia, pH, chlorophyll-a, phycocyanin, and temperature. Data for these parameters were collected both empirically (through direct measurements collected by FWI ground teams) and remotely (via satellite imagery analysis provided by Captain Fresh). Data collection occurred every five days, aligning with the Sentinel-2 satellite's flyover schedule, with five rounds of data collection for each farm (once every five days).
In June 2024, we circulated a report summarizing the key findings from this proof-of-concept study as part of a blog post published on our website. At the time, we were excited, as the results indicated a strong correlation between the satellite-predicted and the empirical data for four out of six key water quality parameters, suggesting that this technology had potential for us to integrate into the ARA.
However, concerns about the findings soon came to light when we subsequently engaged independent consultants to probe into the data. These consultants believed the results were “too good to be true,” and their efforts to develop predictive models of their own using the same empirical dataset used by Captain Fresh were unsuccessful. Additionally, when we attempted to replicate the findings on our own by using the same model that had given us the initial results in the study, we were unable to do so.
We published another blog post in January 2025 to convey this updated assessment of the proof-of-concept study. This experience provided us with valuable lessons about the importance of additional validation before concluding.
Previous Approach #2: Innovation Challenge (Version 1)
Despite the setbacks from our initial proof-of-concept study, we remained optimistic about the potential of satellite-based remote sensing. To boost our chances of successfully developing useful models, we launched an Innovation Challenge in December 2024. Unlike our earlier approach—where we worked exclusively with a single collaborator—this challenge invited any interested party to submit models, with financial rewards offered for those that met our specifications.
A key lesson from our earlier study was the need for a stringent internal validation. Previously, we had relied on the developers themselves to validate their models—a mistake we aimed to correct. For the Innovation Challenge, we conducted our internal validation process. Parties were asked to use their models to estimate values for key water quality parameters—dissolved oxygen, ammonia, pH, and/or chlorophyll-a—at 20 farms in the Eluru region of Andhra Pradesh, at which we had collected ground-truthed data. We evaluated each submitted model by comparing its predictions to our ground-truthed data. Financial awards—up to USD 10,000—were based on predetermined performance thresholds for each water quality parameter.
In June, we published a blog post sharing the results of this Innovation Challenge. Ten different groups participated, submitting a total of 33 models:
9 models for dissolved oxygen
6 models for ammonia
8 models for pH
10 models for chlorophyll-a
Unfortunately, none of the 33 models met our minimum criteria. Even the very few models that performed relatively well (compared to the other models) on our criteria had very little predictive value in real-world terms.
Previous Approach #3: External Consultants
We have also engaged several external consultants who kindly provided their services pro bono at various stages in our process (more on those in the “Note of Gratitude” section below).
In addition to these consultants, we hired three separate contractors on Upwork: Two to develop remote sensing models for chlorophyll-a (chl-a), and one to develop a forecasting model (more on that below). The idea behind focusing on the Chl-a model was that a) it is the only optically active parameter and thus the easiest one to develop, and b) a dissolved oxygen model will likely require effective Chl-a prediction as an input.
Unfortunately, when validated in the same way as models submitted through the Innovation Challenge, neither model met our minimum criteria for accuracy. You can see the code for both of these models here:
Upwork Chl-a Model 1 (Andrey)
Upwork Chl-a Model 2 (Uli)
Our Current Approaches
We are currently pursuing the following two approaches in parallel:
Current Approach #1: In-House Development
We are currently hiring for a Water Quality Remote Sensing Lead to develop this technology in-house. As of the time of this publication, the initial application deadline (June 29) has now passed, though we are still accepting applicants on a rolling basis.
We expect this person to begin work around September 1, and for them to spend some time at our field site in India.
Current Approach #2: Innovation Challenge Version 2
Separately, we have relaunched our Innovation Challenge, this time with two key changes:
We have increased the top prize from $10K to $100K, to be paid upon successful validation of a model that predicts dissolved oxygen, ammonia, pH, and chlorophyll-a.
We are sharing our previously collected ARA historical data with applicants who request it.
Note that we have always published this data; however, the key difference here is that we haven’t previously included GPS coordinates with it. Also, note that there are some concerns with the accuracy of this historical data. See “Additional Resources” below for more information.
At the time of writing, the Challenge is ongoing. Interested parties are invited to notify FWI that they have a model or models ready for validation by August 20th, 2025. Thereafter, they will be asked to use their model(s) to determine values for the water quality parameter(s) relevant to their model(s) at 20 farms.
Forecasting Models: An Overlooked Approach?
Thus far, most of our effort has focused on satellite-based remote sensing—trying to infer today’s dissolved oxygen, pH, ammonia, and chlorophyll-a directly from Sentinel-2 imagery. However, it’s possible that we have overlooked short-horizon forecasting—models that use our ground time-series plus open weather and air-quality feeds to predict future water quality values.
To this end, we had previously engaged two data scientists and recently completed a project with another Upwork contractor to build a forecasting model. Unfortunately, neither of these models yielded sufficiently accurate results. You can find further information on them here:
Additional Resources
For those working on our Innovation Challenge, or who are otherwise interested in this technology, the following might be of interest:
Farm Program Historical Data: We currently have a dataset (also see data tool) of about 10,000 historical farm measurements. For accessing this water quality data, please contact us. However, do note some caveats with this data:
Methodologies and equipment have changed (and generally improved) over time as the ARA evolved, giving rise to questions over the reliability and accuracy of some, particularly earlier, data.
There is limited overlap with the flyover schedules of Sentinel-2 satellites, which in practice means that most values in our dataset will probably not be useful for training a remote-sensing model, as water quality changes regularly (though some parameters much more than others).
For each farm, the data collected is from a single location, always at the side of the pond. There are some questions about how predictive such a datapoint is for the broader water quality of the pond. If such a datapoint is not predictive, that would likely mean that our historical farm dataset is less helpful here (and would also have broader programmatic implications for our work). We are evaluating the data from our recent Sampling Point Study to determine how much variance there is between different areas of the pond, and expect to have our final results this month. We will share them in our Remote Sensing Mini Newsletter (see below), if not also on our blog.
Previous R&D Study Data:
Sampling Point Study Data (blogpost pending)
Note:
If you’d like to access either of these, please request access and provide your reasoning (we expect to grant access to all who request).
These datasets should be considered more reliable than our farm program historical data.
Remote Sensing Mini Newsletter: We’re experimenting here with a project-specific newsletter, where we’ll post brief (and probably not particularly formatted) updates on our work and evolving thinking here. We expect to send out updates every 4–8 weeks here. We think this newsletter will be particularly useful for parties interested in developing this technology (e.g., Innovation Challenge participants). Those interested can sign up here.
Our Own Hunches
In this section, we’re including the unfiltered thoughts of a few of our staff who have worked on this project:
Staff #1 (Haven)
Is this even possible? We’ve now spent over a year, and at least tens of thousands of dollars developing this technology. We’ve tried three different approaches, some of which involved multiple independent parties trying to determine a solution. We’ve talked with many companies and have started to get the sense that, despite the promise that some satellite/aquaculture companies make, there are not actually good solutions on the market right now. We even spoke with an aquaculture professional the other day who told us that he spent ten years trying to make this technology work, only to eventually give up on it. After all this, one does start to get the sense that we might be pursuing some holy grail—a huge breakthrough, surely, but one that may not exist, at least with current technology levels.
I do think there’s a good chance that what we’re looking for is simply not technically feasible. If I had to bet, I’d say there’s a 50/50 chance that this is the case. I’m not sure.
What is clear to me is that this is not a simple problem. It’s also probably not a problem that can be cracked without very extensive on-ground data.
If we are going to crack this problem, I think it will take a model with a variety of types of input, including:
Satellite imagery data
Weather data
Historical water quality data about the region
In particular, I think we will need to collect significantly more on-the-ground training data, and of an improved quality than what we have now (for instance, right now much of our data collection doesn’t line up with the Sentinel-2 flyover). Fortunately, our team is experienced at collecting this kind of data.
On another note of optimism, it’s also worth noting that the model doesn’t have to be perfectly accurate for it to add tremendous value to our work. Rather, it just needs to be a few times more likely than pure chance (which is what we’re currently working with) in estimating the likelihood that a given farm currently has poor water quality.
Staff #2 (Jennifer)
Regarding the prediction model to forecast water quality, I feel a little more optimistic because we are not relying on remotely assessing water quality. Instead, the model is built on the correlations of various parameters (including weather, water quality, farm size, etc.), which could determine a water quality outcome. The model is then trained using historical data to develop understanding of these correlations and predict similar future situations with some degree of accuracy. This to me, seems, in theory, simpler than remote-sensing water quality, and I’d thus intuit a 60-70% chance of us finding a model that’s more accurate than guessing water quality.
Staff #3
There is ample research on satellite remote sensing of inland water bodies (e.g.) demonstrating its ability to estimate water-quality parameters—particularly chlorophyll-a, temperature, and CDOM (dissolved organic matter). However, these studies also note that the models require rigorous calibration and validation with in-situ data for accurate results, which is exactly what our Innovation Challenge data collection process provides. I remain optimistic about our approach, though I agree that we need to feed a substantial amount of ground-truth data into the models to train and calibrate them effectively.
My thoughts: If our models achieve roughly 50–60% confidence, we could start integrating them into our current program. By continuously incorporating new data from our ongoing ARA collection efforts, we’ll refine the models further, and within a few years, we should reach a confidence level that allows us to pivot to fully satellite-based remote monitoring.
Potential hyperspectral/multispectral drone imaging: If Sentinel-2 imagery proves limited (e.g., due to resolution or cloud cover or any other challenges that other companies observed), we could apply our calibrated models to drone-captured images. There are still questions to resolve—optimal flight range, the number of ponds per sortie, and the cost of hyperspectral cameras, etc, but according to GPT’s response, our existing models should extend to drone platforms with minimal modification and calibration methods.
Forecasting models: Incorporating meteorological forecasts could help predict short-term water-quality changes, since weather heavily influences these parameters. I do, however, anticipate challenges on this.
A Note of Gratitude
Many people have given their pro bono support for this project. We particularly want to thank:
Ren Springlea and Animal Ask
Sebastian Quaade
Josiah Chamberlain
Nadja Flechner
Dan Wahl
Philip Popien
Rachel Mason
Steven Rouk
Thank you all for contributing in your own ways to maximize our chances of success here :)