Just Keep Swimming…in Deadly Conditions
Gems in STEM: How Machine Learning and Remote Sensing Can Help Save Coastal Dead Zones
You’re a goldfish in a special fishbowl. It’s not a bad life, except that if you swim into certain spots of this bowl, you get suffocated. Whoops. You’re dead. It’s like those messed up “Would You Rather?” games: would you rather stay in the same place for eternity or die of suffocation? Not going to lie, I’m not a fan of either choice.
Luckily, you’re not actually a goldfish in a doomsday fishbowl! Phew, talk about a near-death experience. But, the real fish in our world’s lakes and oceans aren’t so lucky.
Introducing: dead zones.
What are dead zones?
Dead zones are exactly what they sound like: areas in bodies of water with near impossible living conditions. Specifically, dead zones are what we call hypoxic, meaning low-oxygen. Since most organisms need oxygen, marine life can’t survive in these dead zones.
Okay, here’s a solution: just don’t live in dead zones. Easy, right?
Not so fast. While some dead zones occur naturally, most are created by, you guessed it, humans. Since the 1960s, the number of documented dead zones has increased from 10 to 415 worldwide because of human activity and interference [1].
How, what, why, when?
In one word: eutrophication.
The Dangers of Eutrophication
Eutrophication is when a body of water gets an excess of nutrients, usually phosphorus and nitrogen, which then leads to dangerous algal blooms and hypoxic conditions.
Algal blooms, also called “red tides,” basically suffocate the sea life beneath them, thereby creating dead zones. These blooms stop light from penetrating the water, keeping oxygen from reaching the organisms underneath them. As expected, this leads to the mass death of marine mammals, fish, and even shore birds. The excess nutrients lead to rapid and unsustainable growth of algae and aquatic plants, which eventually use up all the oxygen and die off. But even in death, the colonies of algae are nefarious. As the dead plants and algae sink to the bottom, their bacterial decomposition uses up whatever oxygen is left in the water.
This graphic breaks down the process of eutrophication.
Eutrophication typically occurs near inhabited coast lines because the excess of nutrients comes from things like fertilizers in lawns and agricultural fields, sewage, nitrogen produced by power plants and cars, urbanization, and other human activities. The levels of the nutrients vary around the world, and in some regions, atmospheric nitrogen can also worsen eutrophication via the water cycle.
Eutrophication’s most dangerous effect? Dead zones.
So while us goldfish can try to avoid the death spots in our fishbowl, aquatic life is having their homes turned into dead zones because of massively inefficient agricultural practices. In fact, over half of the phosphorus and almost two-thirds of the nitrogen we use for crops become a pollutant [2].
I’ve got another “Would You Rather?” for you: would you rather have more crops or aquatic life? Before you answer, let’s talk about the big picture of dead zones.
What do dead zones mean for the future?
Literally, it’s a bigger picture.
Because of human activities like heavy commercial fertilizer use and factories and sewage facilities, dead zones have grown to cover 95,000 square miles — the size of the UK [3].
Even if these dead zones don’t manage to kill off all the fish in the area, the hypoxic conditions can cause several reproductive problems for them, like low viable egg production, decreasing the size of their reproductive organs, perceived emasculanization (i.e., less female fish), and decreased spawning [4].
But dead zones don’t stop at hurting our marine ecosystems, they also worsen climate change. Since 1955, over 90% of excess heat trapped in the atmosphere has been absorbed by the ocean. As global warming heats our ocean and sea temperatures rise (by 0.13°C every decade), dead zones are expanding [5].
It’s a vicious cycle that researchers call the feedback effect: as oxygen levels decrease in areas of coastal waters and approach 0, bacteria turn to nitrogen for energy. It then produces nitrous oxide, which eventually makes its way to the water’s surface and into the atmosphere. But this “laughing gas” is no joke. Though it makes up 6% of emissions, nitrous oxide is a greenhouse gas that is 300 times more potent than carbon dioxide. It doesn’t stop at that, these expanding dead zones also limit the ocean’s ability to absorb carbon from the atmosphere.
So, global warming leads to bigger dead zones which then produce more greenhouse gases which worsens global warming. Wow, it’s like the opposite of the circle of life.
Okay, this is a little depressing. But don’t lose hope! Here comes the good part: we can do something about it.
The Solution? Goldilocks Fertilization.
First, I think we need to do a little Law & Ordering and clear some names. Yes, dead zones are, well, deadly, but the real culprit here is what creates dead zones: overfertilization.
Fertilizers increase crop yields and thereby use less land for agriculture — great! But fertilizer is often overapplied, which is what leads to this nutrient runoff that creates dead zones.
Okay, but it’s literally killing aquatic ecosystems, so let’s just use less fertilizer, right?
Unfortunately, our “Would You Rather?” question isn’t that simple. It turns out that half of our global population is dependent on synthetic fertilizers for food production [6]. Farmers seem to be facing a trade-off dilemma with little room for error: use too little fertilizer and their crop yields suffer (and people don’t get fed), use too much and cause drastic environmental damage.
To solve this problem, we need to channel our inner Goldilocks to find the exact amount of fertilizer required — not too much, not too little, but just right. Efficient fertilizer use means that our population is still fed while minimizing the excess nutrients that poison our aquatic ecosystems.
Robert Jackson, professor of Earth System Science at Stanford University, agrees, “The most important change we can make is to improve the nitrogen use efficiency of our crops by wasting less nitrogen fertilizer and timing its application more closely to when crops need it.” [7]
And, huzzah, this approach to optimizing fertilizer efficiency has worked before!
In 2005–2015, researchers worked with 20.9 million smallholder farmers across China to attempt to increase crop yields while decreasing their environmental impact [8]. There was no magical technology or life-changing policy implemented, all they did was teach farmers about efficient and environmentally-friendly agriculture practices. The result?
While the amount of nitrogen fertilizer used went down by 16%, the average yields of maize, rice, and wheat went UP by 11%. The increased crop output and decrease of fertilizer used were equivalent to an economic return of US$12.2 billion. That’s a lot of zeros.
Turns out the trade-off isn’t actually as drastic as we believed — we can increase crop yield and decrease the fertilizers’ environmental harm at the same time. Ha, take that “Would You Rather” — I found a win-win!
Before we continue, it’s important to note that not all countries overapply fertilizer; in fact, some need to use more. For example, many countries across Sub-Saharan Africa barely use fertilizer, and their crop yields suffer as a result. In fact, if they used more fertilizers, they could close large gaps in crop yield, prevent habitat loss, increase food security, and enjoy a host of socioeconomic and environmental benefits [9]. That’s why it’s so dangerous for organizations to push the notion that the “less fertilizer used, the better,” an oversimplification of what’s really going on. This kind of message hurts farmers, people, and the environment. Instead, we need to get the fertilizer balance right.
With agricultural activity steadily increasing, it’s time to go all in on what’s called precision agriculture to find exactly where we need fertilizers the most and to figure out the best way to achieve this ideal fertilization point. Enter: machine learning.
What Machine Learning Can Do
Data is so important because the more data we collect, the smarter we can make machines–which is exactly what machine learning (ML) does. Machines are “trained” with data sets and use this knowledge to respond to situations they’ve never seen before, letting them automatically do things like classification, detection, and pattern recognition.
Now, the whole idea of precision agriculture is that you can’t manage what you can’t measure. But it’s not easy to predict the exact amount of fertilizer needed by particular crops, let alone in real-time, so a low-cost and effective monitoring method for nutrients in crops and nearby coastal areas is urgently needed to implement sustainable agricultural practices. This is where remote sensing comes in!
Lately, in remote sensing, which uses satellites and other airborne instruments to collect environmental data, ML has become pretty popular because it can manage these massively complex datasets and provide valuable information without needing significant human intervention. This is in part due to recent advances in earth observation technology that let us obtain images with unprecedented high spatial, spectral, and temporal resolutions — it’s like the iPhone 14 Pro Max camera but for Earth!
A machine learning approach with remote sensing can improve predictions about how natural systems behave, improve data analysis automation, and use these insights to better manage our resources. So, AI and ML have taken precision farming to the next level, but, of course, challenges still remain (which we will talk about later).
Now, let’s cha cha real smooth back to the challenge of finding the fertilizer balance. We can accurately measure crop nitrogen through destructive leaf-tissue sampling and wet-laboratory experiments, but this clearly isn’t scalable nor cost-friendly to do for millions of acres in the long-term. However, we know that some crop traits are strongly correlated to a collection of spectral wavelengths, so we can indeed leverage remote sensing to measure crop nitrogen!
How do we do this? Simply put, remote sensing can detect the energy reflected from ground surfaces. The chemical composition of leaves, including their nitrogen levels, changes how much energy is reflected, but we need high sensitivity to monitor this minute change.
Airborne Hyperspectral Sensors to Measure Crop Traits
Introducing: hyperspectral sensors! Operating on the nanoscale, hyperspectral sensors can detect differences as small as 3–5 nanometers across their entire range and offer hundreds of wavelengths across the full range of visible, near-infrared, and shortwave infrared with high spatial resolution (<1 m). For comparison, other airborne remote sensing technologies can only pick up the visible spectrum and potentially near-infrared, i.e., some small number of spectral bands.
Earlier this year, a research team from the University of Illinois Urbana-Champaign put powerful hyperspectral sensors on a plane and flew it over an Illinois corn field three times [10]. These sensors let them scan fields incredibly quickly, taking only a few seconds per acre. Since the sensors obtain much higher spectral and spatial resolution compared to satellites, the team was able to detect the crops’ nitrogen status efficiency with up to 85% accuracy, close to “ground-truth quality.” So these airborne sensors are not only a powerfully precise tool for remote sensing, they also allow us to monitor larger areas quickly and at low cost, without sacrificing too much accuracy.
This study by UIUC was the first attempt ever to use full-range (400–2400 nanometers) optical airborne hyperspectral sensors to measure a bunch of important crop traits, like photosynthetic capacity and nitrogen content/concentration, at both leaf and canopy scales. This is science, baby!
For their research, the UIUC team ALSO developed the current best algorithm for detecting nitrogen reflectance data from the hyperspectral sensors, which they expect will be used in upcoming detection technologies.
Sheng Wang, assistant professor in the Agroecosystem Sustainability Center of UIUC, says, “Our approach fills a gap between field measurements and satellites and provides a cost-effective and highly accurate approach to crop nitrogen management in sustainable precision agriculture.”
Their end goal is to equip satellites with this technology of hyperspectral sensors, thereby allowing farmers to monitor their fields’ nutrient status early on in the growing season to make better-informed decisions about how to use fertilizers sustainably and efficiently.
Okay, so we’ve talked a lot about crops, crops, and more crops. But we started out this article talking about dead zones, i.e., areas in water. Turns out we can tag team this problem by monitoring nutrients in both crops and coastal waters.
Spatiotemporal Deep Learning to Monitor Coastal Waters
Two important limiting nutrients in coastal waters are dissolved inorganic nitrogen (DIN) and dissolved inorganic phosphorus (DIP). High DIP and DIN can trigger eutrophication and all its gruesome effects that we talked about before, leading to water quality deterioration. To make sure this doesn’t happen, we need to monitor DIN and DIP in coastal waters, just like we monitored the nitrogen levels in crops.
Unfortunately, most field survey techniques for monitoring these nutrients in water are expensive and time-consuming. Not only that, current site-based monitoring techniques only really give us the regional water quality status. Considering how much water we would need to monitor constantly, both of these are probably not the Goldilocks monitoring method we want. :( The search continues!
What we do want is a big picture understanding of water quality in the context of both time and space, or what we call its spatiotemporal patterns, at a large scale, but we want to obtain this understanding at a low cost. If we can do this, we can understand our coastal waters and how to effectively treat them.
Wait, why do we want spatiotemporal patterns? Seems like more trouble than it’s worth, no? Well, we can intuitively see that water (especially off coasts) is a pretty complex environment. Say we’re trying to model a wave and how each of its water droplets behave. We can imagine that this behavior changes depending on the specific location and coast, not to mention the changing seasons. If we only took into account one of these factors, we probably wouldn’t get an accurate picture! Indeed, coastal waters usually have highly seasonal changes in short time intervals, and even more complex factors depending on its coast. So, it makes sense that the relationship between nutrients in water and remote sensing varies with space and time. That’s why we care about spatiotemporal patterns — they enable us to make better predictions.
To find this big picture understanding, researchers Wu et al. (2022) from Zhejiang University and UIUC developed a spatiotemporal deep-learning model (ST-DBN) to estimate large-scale nutrients [11]. Using remote sensing, they achieved very strong predictions, established relationships between measured environmental factors and satellite maps, and reduced estimation errors by over 40% compared to non-spatiotemporal models!
How?
For their study, these researchers used satellite data to explore the spatiotemporal distributions of DIN and DIP over the region of Zhejiang Coastal Sea (ZCS) from 2010–2018. Using the spatiotemporal patterns of nutrients based on the annual, seasonal, monthly, and 8-day average distributions, they wanted to answer three main questions:
- Can the long-term and large-scale DIN and DIP distribution be accurately estimated by a nonlinear ST-DBN?
- How did the spatiotemporal distributions of DIN, DIP, and water quality in ZCS change in the period 2010–2018?
- How can the government control nutrients and improve water quality in the future?
To answer these questions, here is the breakdown of their study:
They found that:
- It’s a yes to Question #1! We can accurately estimate the long-term and large-scale DIN and DIP distribution with a spatiotemporal model.
- The water quality was better in spring and summer and poorer in fall and winter.
- The concentration of DIN and DIP decreased by 24% and 19% in the period 2010–2018, respectively. But, the water quality didn’t significantly improve. Even though the DIN concentration was lower, it still greatly exceeded the worst quality level’s critical value.
- DIN contributed 93.9% to the worst quality, while DIP only accounted for 37.8%. This goes to show that the eutrophication of DIP in the ZCS has gotten much better compared to that of DIN.
The researchers concluded that the Zhejiang Province government should monitor the runoff, velocity, and DIN concentration more frequently in their waters. They should especially try to control DIN more effectively, particularly in fall and winter when the water quality is worse.
This study proves that using spatiotemporal-incorporated deep learning models with remote sensing technology works to monitor nutrients and water quality in coastal areas! But, there’s a reason we haven’t hit the ground running yet — there’s a few challenges left to conquer.
Challenges in the Long-Term and Large-Scale
In order to scale and implement these machine learning models to all our coastal waters, we need data. Unfortunately, we can’t do much if we don’t have accessible and current data sets to train our model. In particular, we would need Analysis Ready Data (ARD) to be readily available, which requires a lot of time and computational power (not to mention smart people) to prepare. We also have technical mountains to climb in preprocessing, extracting, synthesizing, analyzing, storing, transferring, and basically just wrangling these large data sets more efficiently to end up with an accurate and well-developed training data set, which can get hard when we’re talking about complex environments like coastal waters and multi-scale, multi-sensor and multi-platform, and multi-temporal earth observation. Phew, that’s a lot of multi’s!
As our climate changes, accurate projections are increasingly important. With machine learning and close to real-time data from satellites and ocean exploration from remote sensing, we can dive head-first into monitoring dead zones and fertilizer pollutant levels to predict and respond to eutrophication quickly. We can stop the expansion of dead zones, and protect our aquatic life and ecosystems from the doomsday fishbowl.
Before you go, I have one last question (a pick up line, if you will) to end things on a more positive note.
Do you like spatiotemporal patterns? Because I love & appreciate you across space and time. ❤
Until the next one! If you found this interesting, make sure to follow me to stay updated on future articles.
In the meantime, check out other articles in my column here! If you have any questions or comments, email me at apoorvapwrites@gmail.com.
To be the first one to hear about all my new articles, recent events, and latest projects, subscribe to my newsletter: Letter? I Hardly Know Her!
As a reminder: this column, Gems in STEM, is a place to learn about various STEM topics that I find exciting, and that I hope will excite you too! It will always be written to be fairly accessible, so you don’t have to worry about not having background knowledge. However, it does occasionally get more advanced towards the end. Thanks for reading!
