🤖 machine learning

Event-Based Early Warning of Vineyard Disease Risk from Environmental Time Series

This paper proposes an event-based approach for predicting vineyard disease risk within a 3-7 day window by reformulating the task to detect transitions into risk periods rather than daily presence, demonstrating that this formulation better captures environmental precursors and supports actionable early warnings compared to traditional daily classification methods.

Original authors: Ivica Dimitrovski, Ivan Kitanovski, Danco Davcev, Slobodan Kalajdziski, Kosta Mitreski

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ivica Dimitrovski, Ivan Kitanovski, Danco Davcev, Slobodan Kalajdziski, Kosta Mitreski

Original paper licensed under CC BY 4.0 (http://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 grape farmer. Your biggest worry isn't just that your grapes might get sick today, but that they are about to get sick tomorrow. If you wait until you see the spots on the leaves, it's often too late to stop the spread. You need a warning system that tells you, "Hey, the weather is about to get dangerous for your vines in the next few days, so spray your crops now."

This paper is about building a smarter, more practical "weather alarm" for vineyards. Here is how they did it, explained simply:

1. The Old Way vs. The New Way

The Old Way (Daily Check):
Most previous studies acted like a daily health check-up. They looked at the weather and asked, "Is the vine sick today?"

The Problem: Diseases don't just appear and disappear instantly. Once a disease starts, it stays for a while. If a model just learns to say "Yes, sick" every day the disease is present, it's like a broken clock that is right twice a day. It's just repeating what it already knows (persistence) rather than predicting what is coming next.

The New Way (The Event Alarm):
The authors changed the question. Instead of asking "Is it sick today?", they asked: "Is a new disease outbreak going to start in the next 3 to 7 days?"

The Analogy: Think of it like a fire alarm. You don't want the alarm to just scream "Fire!" while the fire is already burning. You want it to scream "Fire coming!" before the smoke gets thick.
The "Gap" Rule: To make sure the alarm isn't going off constantly for the same old fire, they added a rule: A "new event" only counts if the disease has been completely gone for at least 5 days first. This stops the system from getting confused by short breaks in the disease.

2. The Ingredients (The Data)

To make this prediction, the team fed the computer a massive history of weather data from a vineyard in North Macedonia. They didn't just give it raw numbers; they cooked up special "flavor enhancers" (features) to help the computer understand the story the weather was telling:

Humidity & Rain: They looked at how much rain fell over the last 3, 5, or 7 days (like checking if the soil has been soaking wet for a week).
Temperature Swings: They tracked how much the temperature changed from day to night.
Seasonal Rhythm: They taught the computer that "July" feels different from "January" using a special code (cyclic encoding) so the model understands the seasons naturally.

3. The Three "Detectives" (The Models)

The researchers tested three different types of AI "detectives" to see which one was best at spotting the upcoming danger:

Detective X (XGBoost): This is a classic, very smart statistician. It looks at all the past 30 days of weather as a giant list of numbers and finds patterns. It's great at spotting complex rules but doesn't "remember" the order of events as well as the others.
Detective L (LSTM): This is a memory expert. It reads the weather data day-by-day, like reading a story, remembering how yesterday's rain affects today's humidity. It's designed to understand sequences.
Detective T (TCN): This is a pattern scanner. It looks at chunks of time simultaneously to find specific shapes in the data (like a specific pattern of rain followed by heat) without reading day-by-day.

4. The Big Test

They trained these detectives on data from 2020 and 2021, let them practice on 2022, and then gave them a final exam using data from 2023 (a year they had never seen before).

The Results:

The Winner: Detective L (LSTM) was the most reliable. It successfully predicted 7 out of 8 upcoming disease outbreaks. Crucially, it gave the farmers an average of 6 days of warning before the trouble started. It balanced being accurate with not crying "Wolf!" too often.
The Runner-up: Detective T (TCN) was very good at spotting the events but tended to raise false alarms (crying "Wolf!" when nothing was coming) more often than the winner.
The Underperformer: Detective X (XGBoost) was the most cautious. It rarely cried "Wolf!" falsely, but it also missed more actual outbreaks than the others. It was too conservative.

5. Why This Matters

The paper concludes that changing how we ask the question (from "Is it sick now?" to "Will it get sick soon?") makes a huge difference.

The Takeaway: In the real world, a farmer doesn't need a perfect daily diagnosis; they need a reliable heads-up. The study shows that using AI models that understand the flow of time (like the LSTM) is better at giving that early warning than models that just look at a static list of numbers.

In short: They built a system that listens to the weather, understands the rhythm of the seasons, and gives farmers a 6-day head start on disease outbreaks, helping them protect their grapes more efficiently and with less guesswork.

Technical Summary: Event-Based Early Warning of Vineyard Disease Risk from Environmental Time Series

1. Problem Formulation

The paper addresses the challenge of predicting vineyard disease risk (specifically downy and powdery mildew) using environmental time-series data. The authors argue that existing literature predominantly formulates this task as daily disease-state classification (predicting whether a disease is present on a specific day). The authors identify a critical limitation in this approach: daily labels often exhibit temporal persistence, meaning high predictive performance may simply reflect the model's ability to reproduce the continuity of an existing state rather than identifying the environmental precursors that signal an upcoming risk period.

To address this, the paper reformulates the problem as an event-based early warning task. Instead of predicting daily status, the objective is to predict whether a new disease-risk episode will begin within a future actionable window of 3–7 days.

Event Definition: A new event is defined as the first positive day following a minimum disease-free gap (set to 5 days in the study). This rule reduces fragmentation caused by short interruptions in binary labels.
Target: Given a historical window where disease risk is currently inactive, the model must predict if at least one disease-risk onset event will occur in the interval $[t+3, t+7]$ .

2. Methodology

2.1 Dataset and Study Setting

The study utilizes a multi-year agro-meteorological dataset collected from a single vineyard site (originally introduced by Arvanitis et al.).

Inputs: Daily environmental observations including humidity, temperature, and rainfall.
Labels: Management-oriented disease-risk annotations derived from fungicide application records, serving as proxies for disease occurrence rather than direct pathological severity measurements.
Temporal Split: To ensure chronological realism and avoid data leakage, the data is split by year: 2020–2021 for training, 2022 for validation, and 2023 for final testing.

2.2 Feature Engineering and Preprocessing

The pipeline constructs input representations that capture short-term dynamics and seasonal structure:

Cyclic Temporal Encoding: Month and day-of-year are encoded using sine and cosine functions to preserve seasonal continuity.
Agro-Meteorological Descriptors: Derived features include cumulative rainfall (3, 5, 7 days), rolling means of humidity and temperature, daily temperature ranges, and counts of humid/rainy days over specific windows.
Sequence Construction: Fixed-length 30-day historical windows are extracted.
- For XGBoost, these windows are flattened into a tabular feature vector.
- For LSTM and TCN, the windows are preserved as ordered multivariate sequences.
Constraints: Samples are retained only if the current day is disease-free (to predict new episodes), and windows do not cross year boundaries.

2.3 Model Architectures

Three representative model families are compared under the unified event-based formulation:

XGBoost: A gradient-boosted tree classifier operating on flattened tabular data. It serves as a strong baseline for structured, non-sequential data.
Long Short-Term Memory (LSTM): A recurrent neural network that processes the 30-day sequence step-by-step to capture temporal dependencies and accumulation effects.
Temporal Convolutional Networks (TCN): A convolutional model using causal and dilated convolutions to capture local and medium-range temporal patterns with parallelizable computation.

2.4 Evaluation Protocol

The authors propose a dual-evaluation strategy to assess both discrimination quality and operational utility:

Standard Metrics: F1-score, Precision, Recall, and AUROC at the sample level.
Event-Oriented Metrics:
- Event Recall: The fraction of true disease-risk onset events for which at least one valid alert is generated within the 3–7 day warning window.
- Mean Lead Time: The average number of days between the earliest successful alert and the actual event.
- Alert Precision & Episode Precision: Measures of false alarms, distinguishing between "near-miss" alerts (close to the valid window) and "strict false" alerts (outside any event interval). Alerts are aggregated into contiguous episodes to avoid penalizing clustered warnings.

3. Key Contributions

The paper outlines five primary contributions:

Event-Based Formulation: Introducing a short-horizon event prediction framework that targets the onset of disease-risk episodes rather than daily states.
Pipeline Development: Constructing a prediction pipeline that integrates multi-year agro-meteorological data, derived temporal descriptors, and seasonal encoding.
Fragmentation Reduction: Implementing a minimum disease-free gap rule to create consistent event definitions for evaluation.
Event-Oriented Evaluation: Proposing a protocol that measures lead time, event recall, and false-alert behavior, tailored to practical early warning needs.
Comparative Assessment: Providing a controlled comparison of XGBoost, LSTM, and TCN, highlighting trade-offs between model complexity, recall, and alert reliability in imbalanced agricultural settings.

4. Results

The study was evaluated on the 2023 test year (unseen data):

LSTM Performance: The LSTM model demonstrated the most favorable balance for operational warning. It achieved the highest F1-score (0.6190) and AUROC (0.9104) on the test set. Crucially, it detected 7 out of 8 disease-risk onset events (Event Recall: 0.875) with a mean lead time of 6.0 days. It also maintained the highest episode precision (0.5833) among the models.
TCN Performance: The TCN showed strong performance on the validation set (2022) but exhibited less robust generalization on the test set. While it detected 6 of 8 events, it produced a higher burden of strict false alerts (17 strict false alerts vs. 9 for LSTM) and lower precision.
XGBoost Performance: XGBoost was the most conservative model, generating the fewest total alerts and strict false alerts. However, it suffered from lower event recall (detecting only 6 of 8 events) and a significantly lower AUROC (0.7849), suggesting it struggled to rank positive instances effectively compared to the neural models.

Key Observation: The shift from daily classification to event-based evaluation revealed differences not captured by standard metrics. Specifically, the LSTM's ability to generalize to the unseen year and provide timely, reliable alerts made it the superior choice for this specific task, whereas the TCN's validation performance did not translate to the test set.

5. Significance and Claims

The paper claims that reframing vineyard disease forecasting as an event-based early warning problem yields predictions that are more meaningful for operational decision support than conventional daily classification.

Practical Utility: The event-based formulation encourages models to identify environmental precursors for upcoming risk periods rather than merely reproducing state persistence.
Evaluation Necessity: The study demonstrates that standard point-wise metrics (like F1-score) are insufficient for assessing warning systems; event-oriented metrics (lead time, event recall) are essential for understanding the practical usefulness of alerts.
Model Selection: In the context of limited, imbalanced, multi-year agro-meteorological data, sequence-aware modeling (specifically LSTM) appears to offer a better trade-off between detection ability and alert reliability than tabular methods (XGBoost) or convolutional approaches (TCN) for this specific formulation.

The authors maintain a modest scope, acknowledging that the study is a single-site temporal case study and that the labels are management-oriented proxies rather than strict biological onset measurements. They suggest future work should focus on multi-site validation, pathogen-specific targets, and multimodal extensions.