🤖 machine learning

Enhancing Strawberry Yield Forecasting with Backcasted IoT Sensor Data and Machine Learning

This study demonstrates that integrating AI-generated backcasted IoT sensor data with historical records significantly enhances the accuracy of strawberry yield forecasting models compared to using only real sensor and weather data.

Original authors: Tewodros Alemu Ayall, Andy Li, Matthew Beddows, Milan Markovic, Georgios Leontidis

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tewodros Alemu Ayall, Andy Li, Matthew Beddows, Milan Markovic, Georgios Leontidis

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 trying to teach a chef how to predict exactly how many strawberries a farm will harvest next year. To do this, the chef needs a recipe book filled with past experiences: how much water the plants drank, how hot or cold it was inside the greenhouse, how much sunlight they got, and what the actual harvest was.

The problem? The farm only started keeping this detailed recipe book (using smart sensors) two years ago. But the farm has been growing strawberries for four years. The chef is stuck with a recipe book that has two years of blank pages. Without those missing pages, the chef can't learn the full story, and the predictions aren't very good.

This paper is about a clever trick the researchers used to fill in those blank pages so the chef could learn better.

The Problem: Missing Pages in the Recipe Book

The researchers set up smart sensors in strawberry greenhouses (called polytunnels) in Scotland. These sensors acted like a high-tech diary, recording things like temperature, humidity, soil moisture, and water usage every hour.

However, they only had this "diary" for the last two growing seasons. They also had the harvest numbers (the yield) for four seasons. They wanted to use all four years of harvest data to train an AI to predict future yields, but they were missing the sensor "diary" for the first two years. It's like trying to solve a puzzle with half the pieces missing.

The Solution: "Backcasting" (Rewinding the Tape)

Instead of waiting another two years to get more data, the researchers invented a method called backcasting.

Think of it like this: Imagine you have a video of a storm happening outside your house (from a national weather station), but you don't have a camera inside your house to see how the wind rattled your windows. However, you do have a video of the storm from the last two years where you did have an inside camera.

The researchers used the "inside camera" footage from the last two years to learn the relationship between the outside storm and what happened inside the greenhouse. Once they understood that relationship, they went back in time. They took the "outside storm" videos from the two years where they didn't have an inside camera and used their learned relationship to guess (or synthesize) what the inside conditions must have been.

They called this "synthetic data." It wasn't a real sensor reading, but a very educated guess based on real weather data and the patterns they had already observed.

The Experiment: Testing the New Recipe

The researchers then taught three different types of AI "chefs" (machine learning models) to predict strawberry yields. They tested them in two ways:

The Old Way: Teaching the AI only with the two years of real sensor data.
The New Way: Teaching the AI with the two years of real data plus the two years of "guessed" (synthetic) data.

The Results: More Data Means Better Guesses

The results were clear:

The "Old Way" struggled: When the AI only had the limited real data, adding more environmental details (like temperature or humidity) actually confused it. It didn't have enough examples to learn how those details affected the strawberries.
The "New Way" succeeded: When the AI was fed the combined dataset (real + synthetic), its predictions got much better. The models that used the "guessed" sensor data were significantly more accurate than those that didn't.

Interestingly, the AI learned that the specific conditions inside the greenhouse (measured by the sensors) were more important for predicting the harvest than just knowing the general weather outside. By filling in the missing sensor pages, the AI could finally see the full picture.

The Bottom Line

The paper shows that if you don't have enough historical data from your smart sensors, you don't have to wait years to get it. You can use a "time-travel" trick (backcasting) to reconstruct what those sensors likely recorded in the past, using reliable weather station data. This gives your AI a much bigger library of examples to learn from, leading to much better predictions of how much fruit will be harvested.

Important Note: The researchers emphasize that this was a "retrospective" study. They looked back at past data to prove the method works. They did not claim this method is perfect for predicting the future in real-time right now, nor did they claim it works for every type of farm or climate, but it is a powerful tool for farms that are just starting to collect data.

Technical Summary: Enhancing Strawberry Yield Forecasting with Backcasted IoT Sensor Data and Machine Learning

Problem Statement
The paper addresses a critical bottleneck in digital agriculture: the scarcity of historical sensor data required to train effective AI-based crop yield forecasting models. While Internet of Things (IoT) technologies offer real-time monitoring of environmental parameters (e.g., temperature, humidity, soil moisture), deploying these sensors over multiple growing seasons is often cost-prohibitive and logistically difficult, particularly in dynamic farm environments. Consequently, many farms possess historical yield records but lack the corresponding granular sensor data, limiting the ability to train robust machine learning (ML) models that can support informed decision-making, such as labor optimization during harvest.

Methodology
The study was conducted in a real-world strawberry production setting in Scotland, UK, utilizing two types of polytunnels (heated "Seaton" and unheated "Multispan"). The authors deployed IoT sensors to collect environmental data over two growing seasons (2023–2024) while also gathering historical yield data spanning four seasons (2021–2024).

To bridge the data gap for the two seasons (2021–2022) lacking sensor coverage, the authors developed a backcasting approach rather than traditional forecasting.

Backcasting Mechanism: Instead of predicting future states from past data, the model uses archived future meteorological data (from the UK Met Office) to reconstruct missing past sensor observations. Specifically, for a target time $t$ , the model uses a look-ahead window of external weather covariates ( $t+1$ to $t+k$ ) to predict the internal polytunnel sensor value at $t$ .
Data Synthesis: The backcasting models were trained on the 2023–2024 deployed seasons to learn the relationship between external weather and internal tunnel conditions. These models were then applied to historical Met Office data from 2021–2022 to generate synthetic sensor data.
Model Selection: Recognizing that deep learning models (e.g., LSTMs, Transformers) require large datasets not available in single-farm settings, the authors employed ensemble tree-based methods: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost).
Forecasting Pipeline: The final yield forecasting models were trained on three datasets: (a) real sensor data only, (b) real data augmented with synthetic backcasted data, and (c) various feature combinations (yield-only, yield + weather, yield + sensor, yield + sensor + weather).

Key Contributions

Real-World IoT Implementation: The paper documents a comprehensive deployment of IoT sensors in polytunnels, capturing water usage, internal/external temperature and humidity, soil conditions, and photosynthetically active radiation (PAR), while highlighting practical challenges like power failures and sensor dropouts.
Backcasting Methodology: The authors propose a novel solution for limited historical sensor data by generating synthetic polytunnel sensor observations using external weather archives. This approach is distinct from generative adversarial networks (GANs) as it leverages continuous, reliable external data sources rather than latent noise, making it suitable for environments with severe data scarcity.
Empirical Validation: The study validates that combining synthetic backcasted data with real data significantly outperforms models trained exclusively on limited real data.

Results

Backcasting Accuracy: The backcasting models successfully reconstructed missing sensor values. The GBDT model generally performed best for specific features, while RF and XGBoost were selected for others based on the lowest Root Mean Squared Error (RMSE).
Yield Forecasting Performance:
- Models trained on real data only showed that adding environmental features often degraded performance compared to a "yield-only" baseline, likely due to insufficient data volume to learn complex feature interactions.
- Models trained on synthetic + real data demonstrated significant improvements. The GBDT model achieved the best results, reducing RMSE by 19.55% and MAE by 7.75% compared to the real-only baseline when using the "yield + sensor" feature set.
- The "yield + sensor" combination outperformed "yield + sensor + Met Office," suggesting that localized micro-climate data (synthesized from the backcasting) is more predictive than regional weather data alone.
- A window size of 6 future time steps was identified as optimal for the backcasting input, balancing temporal context and noise.

Significance and Claims
The paper claims to be among the first studies to reconstruct missing growing-season IoT observations from external weather archives specifically for downstream crop yield forecasting in operational polytunnel agriculture.

The authors emphasize that their work addresses a pragmatic challenge faced by individual farms: the inability to accumulate sufficient historical sensor data for predictive modeling. By demonstrating that synthetic data augmentation via backcasting can enhance forecasting accuracy, the study offers a viable pathway for small-scale agricultural settings to leverage AI despite data scarcity.

Modesty and Limitations
The authors maintain a modest tone regarding their claims:

Retrospective Nature: The evaluation is explicitly retrospective. While 2023 yield labels were held out for testing, the backcasting model was calibrated using sensor-weather relationships that included 2023 data. The authors acknowledge that a stricter prospective evaluation would require recalibrating without any data from the evaluation year.
Causal Mechanisms: The improved performance is interpreted as evidence of practical predictive utility under data augmentation, not as proof that the models have learned causal agronomic mechanisms.
Generalizability: The study was conducted on a single Scottish farm. The authors note that future work is required to validate the transferability of this approach across diverse climates and polytunnel designs.
Baseline Comparison: The study compares against an autoregressive machine learning baseline but does not include a naive persistence baseline (e.g., predicting next week's yield equals this week's), avoiding claims of outperforming the simplest possible predictor.