📄 agriculture

An Explainable AI-Driven Probabilistic Rainfall Forecasting and Harvest Optimization Framework for Precision Agriculture

This paper proposes XAI-PRFHO, an explainable AI framework that integrates a TCN–BiLSTM model with conformal prediction and multi-layer interpretability to generate calibrated probabilistic rainfall forecasts, which are then optimized via NSGA-III to significantly reduce weather-induced crop losses and increase yields in precision agriculture.

Original authors: Mohammad Zahangir Alam

Published 2026-08-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Zahangir Alam

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a farmer trying to decide the perfect day to harvest your crops. You know that if it rains too hard right when you're cutting the wheat, your yield could be ruined, but if you wait too long, the crop might get old or rot. For centuries, farmers have looked at the sky and guessed, or relied on simple rules like "harvest when the grain is dry." But today, we have computers that can predict the weather. The problem is, most of these computer predictions are like a single, confident voice saying, "It will rain 5 millimeters tomorrow." That sounds helpful, but it hides a secret: the computer isn't actually sure. It doesn't tell you if there's a tiny chance of a flood or a massive chance of a drizzle. It just gives one number.

This is where a new field of science called "Explainable AI" comes in. Think of it as a translator that doesn't just give you the answer, but also explains why the computer thinks that answer is true. It's like asking a weather forecaster, "Why do you think it will rain?" and having them point to the humidity, the wind, and the soil moisture, saying, "These three things are shouting the loudest." When you combine a weather forecaster that admits what it doesn't know (probabilistic forecasting) with one that explains its reasoning (Explainable AI), you get a tool that doesn't just predict the weather; it helps you make life-or-death decisions for your farm with much more confidence.

The Paper's Big Idea: A Weather Crystal Ball with a "Why" Button

In this research, Mohammad Zahangir Alam from the University of Turku built a super-smart system called XAI-PRFHO. You can think of this system as a high-tech farm manager that acts like a detective, a fortune teller, and a chess grandmaster all rolled into one. Its job is to solve a very tricky puzzle: How do you harvest crops when the weather is unpredictable?

The Detective: Finding the Clues
First, the system looks at the past. It doesn't just look at yesterday's rain; it looks at a whole month of weather history, including temperature, wind, soil moisture, and humidity. It uses a special kind of computer brain (a mix of "Temporal Convolutional Networks" and "Bidirectional Long Short-Term Memory") that is really good at spotting patterns in time, kind of like how you might notice a pattern in your friend's mood swings over a week.

The Fortune Teller: Guessing with a Safety Net
Most weather apps give you a single number, like "3 mm of rain." But Alam's system is different. It uses a technique called Conformal Prediction to draw a "safety net" around its guess. Instead of saying "It will rain 3 mm," it says, "We are 95% sure the rain will be between 2 mm and 4 mm." This is huge because it tells the farmer, "Hey, there's a small chance it could be a lot wetter, so maybe don't harvest right this second." In tests, this safety net was incredibly accurate, catching the real rainfall 94.7% of the time when it aimed for 95%.

The "Why" Button: Explaining the Magic
Here is the most fun part. Usually, these super-smart computers are "black boxes"—you put data in, and an answer comes out, but you have no idea how it got there. Alam's system has a "Why" button. It uses two tools, SHAP and Integrated Gradients, to show the farmer exactly which clues mattered most.

The Result: The system revealed that the most important clues were the rain that happened before (antecedent rainfall), the humidity in the air, and how wet the soil was.
The Metaphor: Imagine the computer is a detective solving a mystery. Instead of just saying "The butler did it," it hands you a report showing that the butler was the only one with muddy shoes (humidity) and was seen near the scene (soil moisture) right before the crime (the harvest). This makes farmers trust the system because they can see the logic behind the advice.

The Chess Grandmaster: Making the Move
Finally, the system takes all this information—the rain guess, the safety net, and the explanation—and plays a game of chess against the weather. It uses an algorithm called NSGA-III to find the perfect harvest date. It's trying to balance three things at once:

Don't get caught in the rain.
Don't lose your crop to weather damage.
Get the highest possible yield.

It doesn't just pick one "best" day; it gives the farmer a menu of "Pareto-optimal" choices. Think of it like a menu at a restaurant where you can choose between a "Super Safe" meal (harvest early, lower risk, maybe slightly less food) or a "High Reward" meal (wait longer, risk a bit more rain, but get more food). The farmer can then pick the option that fits their own risk tolerance.

What Did They Find?
When the researchers tested this system in Bangladesh, a place where rain can be very unpredictable, the results were impressive.

Better Guesses: The system predicted rainfall 7 days in advance with an average error of only 3.41 mm/day. This was much better than other models, beating the next best one by about 9% and beating the old-school methods by a huge 33%.
Less Loss: By using this smart system to decide when to harvest, the farmers could have reduced crop losses caused by rain by 31.7%.
More Food: They also projected a 11% increase in the total amount of crop harvested compared to using old-fashioned rules.

What It's Not
It's important to know what this paper didn't do. The researchers didn't go out to a real farm and watch farmers use this system for a whole season to see if their bank accounts got bigger. Instead, they used historical data (past weather and crop records) to simulate what would have happened. So, while the numbers look great, they are estimates based on a computer simulation, not a real-world field trial. The paper also admits that it didn't account for things like whether the farmer has a tractor available, if they can afford to hire extra workers, or if the market price for crops is high that day. It's a brilliant weather and math tool, but it's not a complete business plan yet.

The Takeaway
This paper shows that we can build AI that doesn't just guess the future but explains its reasoning and gives us a safety net for our guesses. By combining a smart weather predictor with a clear explanation of why it's making a prediction, and then using that to plan the perfect harvest, we can help farmers grow more food and lose less to the rain. It's a step toward farming that is not just hard work, but smart work.

Technical Summary: XAI-PRFHO Framework

1. Problem Statement

Rainfall variability poses a critical source of uncertainty for agricultural systems, particularly in climate-vulnerable regions like South Asia. Existing rainfall prediction models predominantly generate deterministic point estimates, failing to quantify the inherent uncertainty of meteorological processes. This limitation hinders high-stakes agronomic decision-making, as farmers cannot assess the risk of extreme events. Furthermore, deep learning models often lack interpretability, creating a trust gap between technology providers and practitioners. Finally, there is a lack of systematic integration between probabilistic rainfall forecasts and operational harvest decision-making; current approaches often rely on threshold-based heuristics that do not fully exploit uncertainty information to optimize harvest timing.

The proposed XAI-PRFHO (Explainable AI-Driven Probabilistic Rainfall Forecasting and Harvest Optimization) framework addresses three specific gaps:

The absence of meaningful uncertainty quantification in rainfall forecasts.
The lack of actionable interpretability for complex deep learning models.
The missing bridge between probabilistic forecasts and multi-objective harvest optimization.

2. Methodology

The framework operates as a unified pipeline integrating five interconnected modules:

A. Hybrid Forecasting Architecture (TCN–BiLSTM)

The core forecasting model combines a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory (BiLSTM) network.

TCN Layer: Utilizes dilated causal convolutions to extract short-term and multiscale temporal patterns. It employs three residual blocks with increasing filter counts (64, 128, 256) and dilation rates (1, 2, 4, 8) to expand the receptive field while preserving causal ordering.
BiLSTM Layer: Processes the TCN output through two layers (128 and 64 units) to capture long-range dependencies in the 30-day input window.
Input: Multivariate agro-meteorological sequences including rainfall, temperature, humidity, pressure, wind speed, soil moisture, and crop-stage information.
Output: 7-day ahead rainfall forecasts.

B. Conformal Uncertainty Calibration

To move beyond point estimates, the framework applies Split Conformal Prediction to the TCN–BiLSTM outputs.

Mechanism: Nonconformity scores are calculated as the absolute difference between observed and predicted rainfall on a calibration set.
Result: The model generates calibrated prediction intervals $[L_{t+h}, U_{t+h}]$ with finite-sample marginal coverage guarantees under the exchangeability assumption. This ensures that the uncertainty communicated is statistically valid (e.g., 95% nominal coverage).

C. Dual-Layer Explainability

To enhance transparency, the framework employs a two-tier interpretability mechanism:

Global Interpretability (SHAP): Uses SHapley Additive exPlanations to identify the global importance of meteorological drivers (e.g., antecedent rainfall, humidity) across the dataset.
Temporal Interpretability (Integrated Gradients): Applies instance-level Integrated Gradients to reveal when specific features influenced a forecast within the historical input sequence.

Note: These are post-hoc attribution mechanisms and do not alter the forecasting or optimization objectives.

D. Crop-Response Modeling (XGBoost-Pheno)

The probabilistic forecasts and conformal intervals are fed into an XGBoost-based crop phenology model.

Inputs: Forecasted rainfall, prediction interval bounds, meteorological variables, soil conditions, and crop maturity stages.
Outputs: Estimates of expected yield, weather-induced crop loss, and crop maturity levels for candidate harvest dates.

E. Multi-Objective Harvest Optimization (NSGA-III)

The final stage uses the Non-dominated Sorting Genetic Algorithm III (NSGA-III) to derive Pareto-optimal harvest schedules.

Objectives: The algorithm minimizes a vector of conflicting objectives:
1. Rainfall-induced risk (probability of exceeding crop-specific damage thresholds).
2. Estimated weather-induced crop loss.
3. Negative expected yield (converted to a minimization problem).
Constraint: Harvest dates must satisfy a minimum crop maturity threshold.
Output: A set of non-dominated solutions representing trade-offs between risk, loss, and yield.

3. Key Contributions

The paper claims four primary contributions:

Probabilistic Forecasting with Guarantees: Development of a TCN–BiLSTM architecture integrated with split conformal prediction, providing calibrated uncertainty estimation with finite-sample marginal coverage guarantees, moving beyond deterministic point forecasts.
Dual-Layer Explainability: Introduction of a pipeline combining global SHAP analysis and instance-level Integrated Gradients to provide both global and temporal insights into forecast drivers.
Integrated Optimization: An XGBoost-based crop-response model coupled with NSGA-III to convert probabilistic rainfall scenarios into Pareto-optimal harvest strategies that balance risk and yield.
Unified Decision-Support Pipeline: A complete framework linking forecasting, calibration, explainability, and optimization, demonstrating how these components function together to support risk-aware harvest planning.

4. Experimental Results

The framework was evaluated using data from 34 meteorological stations in Bangladesh (2005–2023), covering floodplain, coastal, and drought-prone zones.

Forecasting Accuracy: At a 7-day horizon, XAI-PRFHO achieved a Mean Absolute Error (MAE) of 3.41 mm/day and an RMSE of 4.87 mm/day. This outperformed baselines (ARIMA, Random Forest, LSTM, GRU, TCN, Informer) by 9.3% to 33.4% in MAE reduction.
Uncertainty Calibration: The model achieved an empirical Prediction Interval Coverage Probability (PICP) of 94.7% at the 95% nominal level, with a calibration error of only 0.3 percentage points. This significantly outperformed Monte Carlo Dropout (88.2% coverage) and Quantile Regression Forest (91.5% coverage).
Explainability: SHAP analysis identified antecedent rainfall (28.4%), relative humidity (21.7%), and soil moisture (17.9%) as the dominant drivers. Integrated Gradients confirmed these variables exerted the strongest influence in recent historical lags.
Harvest Optimization: Compared to conventional maturity-based heuristics, the full XAI-PRFHO framework reduced model-estimated weather-induced crop losses by 31.7% and increased projected seasonal yield by 11.0%. Ablation studies confirmed that these gains resulted from the coordinated contribution of all framework components (forecasting, calibration, response modeling, and optimization).

5. Significance and Claims

The paper positions XAI-PRFHO as a significant advancement in precision agriculture by addressing the "black box" nature of deep learning and the disconnect between prediction and action.

Scientific Value: It demonstrates that integrating predictive accuracy, calibrated uncertainty, transparent interpretation, and multi-objective optimization creates a robust decision-support system.
Practical Impact: The framework offers a method to reduce climate-induced risks for smallholder farmers, potentially improving revenue stability and food security.
Modesty and Limitations: The author explicitly states that the reported crop-loss reduction and yield gains are model-estimated benefits derived from historical data and optimization, not realized farm-level impacts from prospective field trials. They acknowledge limitations regarding the study's focus on Bangladesh, the temporal gap in soil moisture data (SMAP available only 2015–2023), and the exclusion of socio-economic constraints (labor, machinery, market prices) in the optimization model. The framework is presented as an analytical decision-support tool for agencies and planners rather than a deployed farmer-facing application.