CEREALIA: A Digital Twin Framework for Weather-Driven Agricultural Decision-Making Under Imperfect Conditions

Background

Modern agriculture relies on precise weather data to optimize irrigation, frost protection, and pest control. However, real-world weather networks often suffer from sensor faults, calibration drift, or communication errors. These inconsistencies can lead to poor predictions and crop losses.

Our Framework: CEREALIA

CEREALIA mirrors field weather stations in real time, classifies inconsistencies using neural models, and supports resilient decision-making when perfect data is unavailable.

Key Contributions

• Designed a modular digital twin system that integrates with live and historical weather feeds.
• Evaluated nine state-of-the-art neural network models for anomaly detection in weather data.
• Demonstrated real-world deployment in a Washington State orchard using NVIDIA Jetson Orin.
• Conducted two case studies: weather data imputation and fruit surface temperature prediction with CEREALIA.

Inconsistency Generation in CEREALIA

To realistically study imperfect weather data, CEREALIA includes a noise generator module that emulates sensor faults. This allows us to inject anomalies into otherwise clean data and observe how inconsistency affects decision-making models.

We consider the following four types of inconsistencies:

• Random: unpredictable spikes in measurements, \( X_{\text{random}}(t) = X(t) \times (1 + \eta_d) \), where \( \eta_d \) is drawn randomly with probability \( d \).
• Malfunction: abrupt oscillations caused by sensor errors, \( X_{\text{malfunction}}(t) = X(t) + \epsilon \), where \( \epsilon \sim \mathcal{N}(0,\sigma^2) \times \text{intensity} \).
• Drift: gradual deviation from the true value, \( X_{\text{drift}}(t) = X(t) + \delta + \epsilon \), with offset \( \delta = X(t_0)\times\text{intensity} \).
• Bias: consistent scaling across readings, \( X_{\text{bias}}(t) = \mu \cdot \alpha \), where \( \mu \) is the mean over a window and \( \alpha \) is a bias factor.

These generators mimic common real-world sensor faults such as random spikes, faulty oscillations, long-term drifts, and constant offsets. Incorporating them allows CEREALIA to test robustness of anomaly detection and decision-support models under imperfect weather data.

Machine Learning Models for Inconsistency Detection

To classify inconsistent weather measurements, CEREALIA leverages a diverse set of nine state-of-the-art neural network models. These models capture both short-term patterns and long-term temporal dependencies in weather data.

• CNN Models: Temporal Convolutional Network (TCN) and ResNet – effective for extracting hierarchical features from time-series data.
• RNN Models: LSTM, Bi-LSTM, and GRU – designed to learn sequential dependencies across weather measurements.
• Transformer Models: Time-Series Transformer (TST) and Informer – capture long-range relationships using self-attention mechanisms.
• Hybrid Models: TST-AE and LSTM-AE – autoencoder architectures that combine sequence learning with feature reconstruction.

Together, these models allow CEREALIA to detect and classify anomalies such as random noise, sensor malfunctions, drifts, and biases with high accuracy while operating efficiently on embedded hardware.

Results Summary

We evaluated CEREALIA with nine machine learning models across multiple weather datasets. The key findings are:

• High Detection Accuracy: Most models reliably detected inconsistencies, with bias and drift faults being the easiest to classify.
• Model Variability: CNN and RNN models (ResNet, LSTM) showed the most consistent performance across all noise types, while hybrid autoencoders sometimes struggled with rare fault patterns.
• Lightweight Operation: Inference is fast (under 1 second) and memory usage is minimal (less than 0.4 MB per instance), making CEREALIA deployable on embedded devices like NVIDIA Jetson Orin.



• Scalability: Even when running 30 parallel instances, CEREALIA maintained low overhead and near real-time performance.

These results show that CEREALIA can robustly detect anomalies in weather data and support agricultural decision-making even under imperfect measurement conditions.

Case Studies

1. Weather Data Imputation

Weather stations often produce missing or faulty measurements due to sensor malfunctions or network outages. Such gaps can interrupt decision-making tasks like irrigation scheduling or fruit stress prediction.

CEREALIA addresses this issue using a generative recurrent model (C-RNN-GAN) trained on historical traces and noisy samples. The model can accurately predict missing values across key attributes (temperature, humidity, pressure, wind), ensuring uninterrupted data streams.

2. Fruit Surface Temperature Prediction

Heat stress is a major concern for fruit growers, as excessive surface temperature can cause sunburn and quality loss. Accurate, consistent weather inputs are required for triggering protective measures such as cooling, fogging, or netting. However, faulty sensor data can reduce prediction reliability.

CEREALIA incorporates nine neural models (CNN: TCN, ResNet  |  RNN: LSTM, Bi-LSTM, GRU  |  Transformer: TST, Informer  |  Hybrid AE: TST-AE, LSTM-AE) to predict fruit surface temperature from weather attributes such as canopy air temperature, wind speed, dew point, and solar radiation. When imperfect measurements were introduced, prediction errors increased significantly. With CEREALIA imputing inconsistencies, performance improved and approached the no-fault baseline.

The results demonstrate that when faulty weather feeds are used directly, surface temperature prediction errors nearly double. With CEREALIA imputing inconsistencies, accuracy improves substantially, approaching the performance of perfect sensor data.

CEREALIA implementation is publicly available: https://github.com/CPS2RL/ag-dt

Impact

CEREALIA bridges computing and agriculture, enabling more resilient, data-driven decision processes. By leveraging digital twins, we show how imperfect measurements can still lead to reliable farm management outcomes.