FWI’s Satellite Journey: The End of Our Remote Sensing Project (for now)
- Paul Monaghan
- 9 hours ago
- 11 min read
Summary
Since early 2024, FWI has been exploring if using satellite imagery to remotely monitor water quality is a viable option to incorporate into our programs. Over the past two years we have employed a number of different approaches in an attempt to develop models suitable for our purposes. Unfortunately, none of these approaches has led to the type of models we need.
We recognize that cracking this problem may simply not be achievable with current satellite technology, so we have decided to cease work on remote sensing for the time being. That said, we remain willing to reconsider it in the future if we believe there are opportunities that could lead to the successful development of models suitable for our purposes.
Background
In early 2024, FWI began exploring if satellite-based remote sensing is a viable option to allow us to remotely monitor water quality. What drew us to this approach is its potential to increase the scalability, impact, and cost-effectiveness of our farm program, the Alliance for Responsible Aquaculture (ARA). The scalability of this program is limited by its dependency on in-field measurements, which caps the number of farms FWI staff can visit each day. We therefore want to determine ways to assess water quality without intensive in-field presence. Leveraging satellites to detect water quality issues remotely would lessen our current reliance on visiting farms individually, and would dramatically increase the scale of our work by facilitating data collection at many more farms.
Remotely monitoring water quality via analysis of satellite imagery presents significant challenges. One major challenge for us is that we seek to monitor many—ideally thousands!—water bodies (i.e., fish farms), rather than a single water body. Variations in fish farms (e.g., size, depth, shape, water source, proximity to vegetation, the angle of sunlight, etc.) present a challenge to training models suitable for our purposes. Another major challenge is that some of the key water quality parameters of most interest to us—notably dissolved oxygen and ammonia—are “non-optical” in nature (i.e., there are no visual differences in water as the parameter changes). Satellite-based remote sensing relies on assessing satellite imagery for differences, so the lack of visual differences presents a significant challenge.
Despite the challenges, we believe that satellite-based remote sensing is potentially transformative for our ability to scale our programming and positively impact the lives of millions of fishes. Since the beginning of 2024, we have employed a number of different approaches in an attempt to develop models suitable for our purposes. This blog post outlines these approaches, where we’re currently at, and our plans for remote sensing going forward.
Our First Efforts at Model Development: Collaboration with a Private Partner
When we first set out on our remote sensing work, FWI had no prior experience in satellite-based remote sensing. We began our work to test the potential for remote sensing in February 2024 by partnering with a private partner in India to leverage their experience with accessing and analyzing satellite imagery. Together, we conducted a proof-of-concept study which compared water quality data determined by analysis of satellite images (provided by our partner) with empirical water quality data determined by direct measurements at the same 20 ARA member farms in the Eluru region of Andhra Pradesh (provided by FWI).
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 key water quality parameters, suggesting that this technology had potential for us to integrate into the ARA. However, concerns about the findings came to light when we engaged independent consultants to probe into the data. Efforts by the consultants to develop predictive models of their own—using the same empirical dataset we collected as part of the proof-of-concept—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 a 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 drawing conclusions (when we initially circulated the positive findings, we relied exclusively on the validation conducted by our partner).
Persevering: An Open Call for Models via Innovation Challenges
Despite the disappointment of the revised findings from the initial proof-of-concept study, we continued to be optimistic about the potential for satellite-based remote sensing. To boost our chances of successfully developing useful models, we launched an Innovation Challenge in December 2024. This approach opened up the challenge of developing models to any interested party willing to get involved, rather than us working exclusively with a selected collaborator. We offered a financial reward to any party who could successfully show that they have models matching our needs. Learning lessons from the initial proof-of-concept study—for which we relied on the validation of models by the developers themselves—we conducted a stringent internal validation process to assess the accuracy of the models submitted through the Innovation Challenge.
In June 2025, we published a blog post summarizing the outcomes of the Innovation Challenge. Unfortunately, none of the submitted models met our minimum criteria. Despite the Innovation Challenge not resulting in useful models, we believed this strategy was worth one more try. In June 2025, we relaunched the Innovation Challenge with some changes, most notably: we significantly increased the financial rewards on offer, and we offered parties historical data from the ARA to help train their models (this dataset was not offered as part of the original Innovation Challenge as we were reluctant to have models constrained by potential limitations in the dataset).
In November 2025, we published a blog post summarizing the outcomes of the second Innovation Challenge. Unfortunately, the outcomes of this second round were similar to the first: none of the models met our minimum criteria.
Over the course of the two rounds of Innovation Challenges, we put 202 models (61 dissolved oxygen, 49 ammonia, 47 pH, and 45 chlorophyll-a) from 47 different parties through our validation process. 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. This underscores how difficult a technical challenge it is to develop accurate remote sensing models.
Building Our Own Capacity: In-House Development of Models
Pursuing the second round of the Innovation Challenge was one of two parallel tracks we simultaneously pursued. We purposely pursued two parallel tracks as we recognized that developing the models we seek is a huge challenge, but the rewards—should we be successful—are potentially transformative for our ability to scale our programming and help millions of fishes. The other track was an in-house development process. Up until this point, our efforts to develop models relied on leveraging external expertise. We decided that it was time to bring in-house expertise in remote sensing and develop machine learning models. In August 2025, we recruited a Remote Sensing Lead—Sol White—on a six-month contract to lead in-house development of models.
Between November 2025 and February 2026, Sol worked exclusively on developing models for our key water quality parameters. As part of his efforts, he evaluated the suitability of our existing ARA dataset for training models. Although the dataset is large—spanning more than 200 farms since 2021—Sol concluded that two limitations reduce its usefulness for remote sensing:
Sampling location: Most measurements were taken from pond edges (for logistical reasons). However, edge sampling introduces land–water mixing effects, complicating satellite analysis.
Timing mismatch: Sentinel-2 satellites pass over Andhra Pradesh at roughly 10:30 a.m., whereas ARA sampling typically occurs early morning or evening. This is especially problematic for modeling dissolved oxygen, which fluctuates substantially throughout the day and requires close time alignment.
To address the limitations of our historical dataset, and to provide Sol with the best data we could for developing models for our purposes, we launched a dedicated data campaign designed specifically to generate high-quality, time-aligned data that satellite-based modeling requires (we published the final dataset as a blog post in March 2026). This was a big undertaking for FWI, involving support from across FWI’s departments, but we believed this undertaking was worth the resources.
Sol utilized this data for training his models, while also incorporating external datasets to increase diversity and sample size. Additionally, he explored integrating weather and environmental covariates (temperature, humidity, wind, precipitation, solar radiation/sunshine, cloud history, and seasonality proxies). Unfortunately, even with the high-quality, time-aligned data specific to our ponds, none of the paths that Sol pursued resulted in models that met our minimum criteria. Here we include a summary of Sol’s work, as well as links to his models.
An Ongoing Approach: Leveraging External Consultants
Throughout our remote sensing work, we have engaged a number of external consultants. Initially, we engaged three separate independent consultants who kindly provided their services pro bono to review the data from our initial proof-of-concept study with our private partner. This involved probing into the data as well as attempting to develop predictive models of their own using the same empirical dataset we collected as part of the proof-of-concept.
An independent consultant helped us with setting the metrics and assessing the final outputs for our Innovation Challenges (we are grateful to Ren and Animal Ask). We also engaged two paid consultants on a short-term basis through Upwork.com specifically to develop models for determining chlorophyll-a, as the only optically active water quality parameter of focus, which is considered the most feasible to determine through analysis of satellite images. Unfortunately, when validated in the same way as models submitted through the Innovation Challenge, neither model met our minimum criteria (for reference, see one of the contractor’s app and report).
Weather/Prediction Models
External consultants have also helped us with a slightly different approach: rather than trying to read water quality off satellite imagery directly, we asked whether we could use weather forecasts as a proxy for pond dissolved oxygen risk—since temperature, wind, humidity, and cloud cover all influence overnight oxygen levels.
In March 2026, we put out an open call for volunteers to attempt this, with several people contributing work over the following months. One volunteer (thanks Skye!) built a regularized regression model trained on two years of our historical ARA water quality data, using only Open-Meteo weather variables (no in-field measurements). The deployed version produces a daily "alert tier" for the Eluru region—Normal, Elevated, or High—based on tomorrow's forecast, with the intended use being light: on Alert days, staff would add a small number of extra visits. The live model and our analysis are available here (note that we may take this link down in the future—contact us if you’d like to access it and the link isn’t working).
There are several limitations worth flagging:
The model mostly predicts regional, not per-pond, risk. It primarily tells us whether tomorrow looks risky for the Eluru region as a whole—not which specific pond is bad. The practical use is therefore adding visits on Alert days, not changing which ponds we visit on Normal days.
It targets only dissolved oxygen. Historically, most of our "fishes helped" instances involve dissolved oxygen plus another parameter (such as ammonia or pH) being out of range at the same time. So catching more dissolved oxygen-only events only partially translates to additional welfare impact.
Despite these caveats, we believe the model is worth a field test. From June through August 2026 we are running a two-month field test with our ARA program: each morning we'll check the model for the day ahead, and on Alert days we'll ask each field staff to visit 1–3 more farms than they would normally. We'll then compare the out-of-range detection rate on those extra visits to the rate we see on Normal days.
We don't expect this model to be transformative, but we do hope that it will allow us to better optimize our rate of detecting bad water quality.
Path Forward
From the outset of our work on satellite-based remote sensing, we recognized that developing accurate models for the specific water systems and welfare-relevant parameters we monitor is an exceptionally difficult technical challenge. This difficulty has been underscored by the fact that our efforts to date—involving working with a private partner, 47 independent parties via two rounds of Innovation Challenges, an in-house remote sensing expert, and numerous external consultants—have failed to lead to models that suit our purposes.
After multiple attempts to crack this problem over the past two years, including dedicating significant resources to build a high-quality data set, we are very cognizant of the possibility that this simply isn’t achievable with current satellite technology.
Why have we failed to develop models? There’s no easy answer to that question, but there are various possible reasons:
Developing a model for monitoring a specific water quality parameter that is applicable to many different ponds (as opposed to being used for monitoring the same water body over time) is a challenge. Variations in the size, shape and depth of fish ponds present a challenge. Importantly, the angle at which sunlight reflects off the water surface, which will differ from pond to pond as the satellite flies over and captures data, presents a significant challenge to developing accurate and reliable models.
The non-optical nature of our key water quality parameters of focus—dissolved oxygen, ammonia and pH—was recognized from the outset as being a challenge. It’s possible that we simply didn’t have enough suitable training data to develop models for these parameters. Or, it’s possible that, regardless of quantity and quality of the training data, these non-optical parameters simply can’t be reliably measured using the satellites—and their onboard sensors—that we leveraged (primarily Sentinel-2).
We know that water quality is dynamic and complicated. It’s possible that varying water conditions (i.e. changes in water quality parameters other than the one a specific model is trying to monitor) could affect a model for a specific parameter. Also, water quality varies across a pond, and it’s possible that the spatial resolution of the satellite data may be insufficient to account for intra-pond variations in water quality.
It’s possible that there is insufficient signal and/or too much noise in the satellite data. This could be exacerbated by errors in the ground-truthed data (e.g. natural error or drifts in the readings from our handheld water quality sensors used for collecting our ground-truthed data, or by potential operator errors at the time of collection). Noise in both measurements (ground-truthed and satellite-derived data) presents problems for a model trying to learn a subtle correlation.
Water quality varies across a pond, and it’s possible that the spatial resolution of the satellite data may be insufficient to account for intra-pond variations in water quality. Similarly, there could be a “depth mismatch” between the ground-truthed and satellite data. The satellites capture data from the surface of the water, but our handheld water quality sensors capture data from below the surface (the probes need to be submerged). While the difference in depth may be minimal, it could impact models.
Other unknown unknowns.
We have invested significant time and resources on trying to crack this problem over the past 2+ years. As part of our commitment to evidence-based programming, our process of validating submitted models (those submitted through the two rounds of Innovation Challenge, as well as the models developed by our in-house and external consultants), served as a form of field-based study to assess if the models are sufficient to take forward to a program (models were assessed against real world data collected from the type of ponds that a program would ultimately focus on). It’s clear that we have not yet identified a model that meets our requirements.
Are we giving up? Yes, and no. Satellite-based remote sensing has the potential to be transformative for our ability to scale our programming and positively impact the lives of millions of fishes, but our efforts to date suggest that the technology is not yet viable for our purposes. We will put this on the back burner for now, but we are very willing to pick this up again if we believe there are opportunities that could lead to success.
Acknowledgements
We’d like to express our sincere gratitude to all those who have provided support for our efforts to develop remote sensing models. We especially recognize the support and technical knowledge provided by our private partner at the outset, all the individuals/groups who participated in our Innovation Challenges, and external consultants (many of whom provided their support pro bono). Building models for this setting is a significant technical challenge, and we deeply appreciate the time, expertise, and care you invested. We may have not achieved the outcomes we were hoping for, but your contributions have helped us better understand the current feasibility of remote sensing for aquaculture. The lessons learned from working with all of you will be invaluable for future work on remote sensing. Thank you.
Comments