Explainable AI for Agricultural Climate-Risk Prediction: A Systematic Review of Progress and Strategic Roadmap

The global food supply hangs in a delicate balance, constantly tested by the shifting moods of the climate. Droughts can wither fields, salty soils can poison roots, and unpredictable weather can turn a promising harvest into a loss. For decades, scientists have turned to computers to predict these dangers, using complex algorithms that can spot patterns in weather and soil data far faster than any human. These artificial intelligence systems have become incredibly good at guessing what will happen next, often with near-perfect accuracy. However, a significant problem has emerged: while these computer models can predict the future with great precision, they cannot explain how they reached their conclusions. They operate like a sealed box, offering a result without showing the work inside. For the farmers and agricultural experts who must make life-or-death decisions based on these predictions, this lack of transparency creates a wall of distrust. If a computer says a field will fail, but cannot say why, the farmer cannot know whether to trust the warning or ignore it.

A new systematic review, published in 2026 by researchers from Majan University College, Lakehead University, and Ontario Tech University, takes a hard look at this growing divide. The authors gathered and analyzed dozens of recent studies that attempted to fix this "black box" problem by adding a layer of explainability to agricultural predictions. Their goal was to see if the technology has matured enough to be trusted in the real world. The review covers a specific window of time from 2021 to 2026, a period where the focus of research shifted sharply toward making these models understandable. The researchers examined how these new systems handle five critical areas: predicting crop yields, detecting soil salinity, forecasting drought, identifying plant diseases, and assessing general climate risks. They found that while the underlying prediction models have reached impressive levels of accuracy, the methods used to explain them are often unstable, incomplete, or disconnected from the actual biology of the plants.

The review reveals that the most successful prediction models today are no longer simple statistical tools but deep, complex networks that mimic the way human brains process information. In the realm of crop yield, these advanced systems have achieved accuracy scores as high as 0.975, meaning they can forecast harvest sizes with remarkable precision. Similarly, for detecting plant diseases, some models have reached 99.63% accuracy, correctly identifying sick leaves from healthy ones almost every time. These systems rely on data such as temperature, humidity, soil quality, and satellite images of the fields. However, the study highlights a troubling trend: the explanations provided for these high scores are often an afterthought. Most researchers build a powerful prediction model first and then attach a separate tool to try to explain it later. This approach is like building a car engine and then trying to draw a map of how the fuel flows only after the car is already running. The review found that these "post-hoc" explanations often fail to capture the true logic of the model, especially when dealing with the messy, interconnected reality of weather and soil.

One of the most common tools used to explain these models is a method called SHAP, which tries to break down a prediction to show how much each factor, like rainfall or soil pH, contributed to the final result. The researchers found that SHAP is mathematically robust and consistent, making it a strong choice for understanding complex interactions in soil salinity and yield. However, it is also very slow and computationally heavy, requiring significant power to run on large datasets. Another popular tool, known as LIME, is faster and easier to use, but the review found it to be dangerously unstable. Because LIME creates a simplified, local approximation of the model, small changes in the data can lead to completely different explanations for the same situation. This inconsistency makes it unreliable for farmers who need to know exactly why a specific field is at risk. The study also noted that older methods, which look at how one variable changes in isolation, often fail in agriculture because weather factors are deeply linked. For instance, you cannot change the wind speed without also changing the temperature or humidity, yet some explanation tools treat them as if they are independent, leading to confusing or impossible scenarios.

The review also uncovered a significant gap in how these systems handle uncertainty. In the real world, no prediction is ever 100% certain, yet most agricultural models present their results as absolute facts. A farmer might be told a field will have a low yield, but the model does not say how confident it is in that number. The researchers point out that without a clear measure of confidence, such as a range of possible outcomes, farmers might make risky decisions based on a guess that the computer treats as a certainty. Only a few studies, such as one focusing on soil salinity, have begun to include these confidence intervals, showing the difference between a high-risk, high-certainty warning and a speculative one. The absence of this information in most current research means that the technology is not yet ready for widespread, high-stakes use in the field.

Another major issue identified is the "silo effect," where researchers study drought, disease, and salinity as if they are separate problems. In reality, these risks feed into one another. A drought might force a farmer to use poor-quality water for irrigation, which then increases soil salinity, which in turn makes the crops more susceptible to disease. The current generation of explainable AI models generally fails to show these cascading connections. They might explain why a plant is dry or why the soil is salty, but they rarely explain how one problem led to the other. This fragmentation limits the usefulness of the technology, as it cannot provide a complete picture of the farm's health. Furthermore, the review noted a strong bias in the research, with most studies focusing on regions in the Global North or specific, well-documented areas like the Nile Delta. The models developed for these places often do not work well when applied to different crops, soils, or climates in other parts of the world, and the explanations they provide may not make sense to local farmers.

Looking toward the future, the authors propose a strategic roadmap to move the field from experimental labs to real-world application. They argue that the solution is not just to add an explanation tool after the fact, but to design the models from the very beginning with transparency in mind. This concept, called "Explainability-by-Design," would mean building the AI so that it naturally produces clear, logical reasons for its predictions, rather than trying to reverse-engineer them later. The researchers also call for a unified approach that respects data privacy, allowing different regions to share knowledge without sharing sensitive raw data. Crucially, they insist that any explanation must be validated by human experts, such as agronomists and plant pathologists, to ensure that the reasons the computer gives actually match the biological and physical realities of the farm. If a model says a plant is sick because of a specific pattern of light, a human expert must confirm that this pattern is a real sign of disease and not just a camera artifact.

The ultimate goal of this research is to bridge the gap between high-tech prediction and human trust. The review concludes that while artificial intelligence has the power to safeguard global food security against climate change, it cannot do so until it becomes a partner that farmers can understand. The technology must move beyond simply being accurate to being trustworthy. This means providing not just a prediction, but a clear, confident, and scientifically sound explanation of why that prediction was made. Until the field adopts these new standards of design, validation, and uncertainty communication, these powerful tools will remain locked in the laboratory, unable to help the farmers who need them most. The path forward requires a shift in mindset, treating explainability not as a technical add-on, but as the foundation of any system intended to guide the future of agriculture.