I'm excited to announce that I have a new publication out in the Journal of Environmental Quality titled Enhancing predictions of nitrous oxide emissions from agricultural soils using a classification-swap machine learning approach! For this paper I got to partner with my good friend Facundo on an idea I had with its origins in, interestingly enough, dermatology and medical science. In this blog post, I'd like to share some background of how this paper came to be, and why I think its findings are important for the future of ag+climate modeling.
There are a couple of huge challenges with modeling nitrous oxide (N2O) emissions from ag soils. First, N2O emissions are pretty hard to measure- it typically requires a huge amount of manual labor to collect gas samples from the field, and then expensive lab equipment to analyze those samples. As a result, there are relatively few high-quality datasets of N2O emissions from ag soils, which makes it hard to put together enough data to build reliable and generalizable models. Second, N2O emissions are influenced by a complex interplay of factors, including soil properties, weather conditions, and- crucially to this paper- biogeochemical processes occurring in the soil. You see, N2O is a byproduct of microbial metabolic activity in the soil. There are two main groups of bacteria which use nitrogen as a key part of their metabolism- nitrifying bacteria, which convert ammonium (NH4+) to nitrate (NO3-), and denitrifying bacteria, which convert nitrate (NO3-) to nitrogen gas (N2). Both of these groups of bacteria produce N2O, but under very different conditions. For example, nitrifiers are aerobic, meaning they are most active when soils are well-aurated and dry, while denitrification is an anaerobic process, and thus overwhelmingly occurs when soils are wet and poorly-aerated. All of this means that there are at least two very different "modes" of N2O production in ag soils, each with its own set of drivers and relationships to the environment.
So as you can see, these two challenges compound one another to make accurate N2O modeling a famously elusive goal. As a general rule, the more complex the data patterns and relationships you want to model, the more data you need to train and validate that model. Because soil N2O emissions data is both complex and scarce, there generally hasn't been a whole lot of success in building accurate models for it- machine learning or otherwise. The goal of my PhD work has been to tackle these challenges, and this paper represents one attempt aimed at simplification.
As we discussed, a key reason why N2O emissions are so complex is because it is actually the net result of two very different processes- nitrification and denitrification- which respond to environmental changes in very different ways. Often N2O emissions are modeled as a single process, which means that the model is trying to learn two different and often contradictory sets of relationships. What's worse, nitrification is the dominant process for the large majority of the time, with brief pulses of denitrification-driven N2O emissions often being drowned out as a small minority of the data.
This is where the dermatology connection comes in. Medical and racial justice researchers and advocates (who I am proud to say include my significant other, who is a dermatological PA and has taught me much!) have long pointed out that medical textbooks and training materials have a strong bias towards lighter skin tones and white populations. This has real consequences- for example, a large-scale Nature Medicine paper evaluating over 1000 dermatologists and primary care physicians found that both groups perform significantly worse at accurately diagnosing the same skin conditions on dark skin compared to lighter skin tones. Poor health outcomes resulting from this bias are well-documented, but can be hidden in aggregate health data by the overwhelming majority of lighter-skinned patients. To address this issue, medical researchers have advocated against using one-size-fits-all models and performance metrics, and instead recommend explicit representation and evaluation of different skin tones and populations.
Image credit: Designed by Freepik
Inspired by this work, I wondered if a similar approach could be applied to N2O emissions modeling. What if, instead of trying to build one model to capture all types and modes of N2O emissions, we explicitly separated out different "classes" of emissions, and built separate models for each class? This way, each model could focus on learning the specific and more consistent relationships and drivers relevant to its class. Theoretically, this would lead to a lower data requirement for each model, since each model would be simpler and more focused. Additionally, by separating the emissions into distinct classes, we could prevent the dominant class from drowning out the minority class in the aggregate data.
This is precisely the approach we took in this paper. To build individual models for nitrification-dominant background emissions and denitrification-dominant pulse 'hot moment' emissions, we first segmented all of our N2O emissions data into these two classes based on their rate of emissions. We then trained separate random forest models for each class, using the same set of environmental predictor variables. Finally, we combined the models into a type of ensemble model we called a "classification-swap" model- for each new data point, we first used a classifier to determine which model was most appropriate, and then used that model to make the final prediction.
The results were very promising! We found that the classification-swap model significantly outperformed a traditional single-model approach, achieving ~400% improvement in R2 and ~15% reduction in RMSE on an independent holdout dataset. This suggests that explicitly accounting for the different modes of N2O emissions can lead to substantial improvements in model accuracy and reliability. These improvements were largely driven my our classification-swap model's superior representation of the rare but highly impactful denitrification-driven N2O pulses, which were poorly captured by the traditional machine learning model. Ordinarily, you would expect such improvements on a minority class to require a significant increase in data- however, our approach allowed us to achieve these gains from a relatively modest dataset from only a single experimental site with multiple years of observations.
Overall, I believe this paper demonstrates a valuable case-study in a strategy to improve agricultural GHG modeling given the significant challenges and constraints we face. Perhaps more than that, even, is how interdisciplinary inspiration and thinking can lead to novel solutions to long-standing problems. I hope that this work encourages others to be curious about each other and the world. Each of us has wisdom to offer from our own unique experiences and perspectives as long as we are open to seeing it.
