Integrating Neural Networks and Physical Modeling for Microclimate Precision

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Dec 3 2023

In an article published in the journal Nature, researchers introduced a microclimate model combining neural networks and terrain features, leveraging physical laws and data-driven algorithms. The research scrutinized the impact of using global climate data (ERA5) versus local meteorological station data on microclimate model accuracy.

Background

Microclimate, encompassing localized climatic conditions at the meter (m) scale, plays a pivotal role in precision agriculture and natural resource management. Understanding this complex phenomenon involves navigating intricate interactions between global and local processes, such as solar radiation shading, vegetation-induced temperature buffering, and wind sheltering effects. Precision agriculture aims to enhance accuracy in comprehending microclimate variations, considering factors like temperature, humidity, precipitation, and wind patterns that profoundly influence crop growth.

Existing microclimate monitoring systems often rely on Internet of Things (IoT) technologies, enabling real-time data collection and transmission. These systems and climate data facilitate a holistic analysis of microclimate patterns. Traditional approaches utilize physical laws like fluid dynamics and heat transfer, integrating linear data-driven methods for downscaling and detailed microclimate modeling. However, these methods might fail to capture the intricate relationships inherent in microclimate dynamics.

The advent of artificial intelligence (AI) introduced non-linear modeling techniques, particularly neural networks, offering a promising avenue for unraveling the complexities of microclimate interactions. Neural networks excelled in discerning non-linear relationships hidden within data, making them valuable tools for understanding the nuanced variations in temperature and humidity. Integrating localized sensor data with global climate information allowed these networks to capture intricate relationships, providing a comprehensive description of microclimate variations.

Yet, the accuracy of neural networks depended on the quality of climate data used for training. Global climate datasets like ERA5, with 25 km-scale resolution, may lack the specificity required for meter-scale precision in agricultural fields. This research addressed this gap by introducing an innovative microclimate model that combines physical laws and neural networks, evaluating the impact of global climate data versus local meteorological station data on microclimate modeling accuracy. Through this, the study aimed to contribute insights into the optimal data sources for accurate microclimate predictions, crucial for informed decision-making in precision agriculture.

Methods

The study focused on developing a microclimate model for Bergamo, Italy, using a combination of physical modeling and neural networks. The study area's terrain data, validated through a digital surface model, revealed low-inclination slopes with a nearby 700m high mountain. Climatic conditions in Lombardy exhibited variations, and meteorological data from the ERA5 database (25 km resolution) and a local Lombardy Regional Environment Protection Agency (ARPA) station were used. The microclimate model integrated these data with a digital surface model to predict temperature and humidity at a 2m resolution.

The physical model incorporated short-wave radiation, long-wave radiation, and wind speed, with the latter adjusted for local terrain. This information is fed into feed-forward neural networks trained on data from 25 sensors, collecting temperature and humidity from November 2022 to June 2023. The neural networks used Keras and TensorFlow, with architecture optimized through grid search.

The study acknowledged sensor failures during calibration and recording but emphasized compensatory effects due to the sensor distribution. Neural network performance was evaluated based on Log-Cosh loss, mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). Furthermore, a method was proposed to interpret the neural network's black-box nature by systematically replacing input variables and observing their impact on predictions.

Study Results

The study applied a physical microclimate modeling approach to generate high-resolution variations in physical variables related to temperature and relative humidity. Utilizing global climate data from the ERA5 database and local data from the ARPA station, the model depicted spatiotemporal variations in short-wave radiation, long-wave radiation, wind speed, and more across the study area.

The feed-forward neural networks trained on this data successfully predicted local temperature variations. The ERA5-trained network exhibited a high level of accuracy (R² = 0.98) with MAE of 0.61°Celcius(C). The ARPA-trained network performed slightly better, demonstrating the effectiveness of local meteorological data.

For relative humidity predictions, the ERA5-trained network achieved moderate accuracy (R² = 0.65) with a MAE of 5.67%. The ARPA-trained network showed similar performance (R² = 0.70) with a lower MAE of 5.34%, emphasizing the challenges in predicting relative humidity accurately.

Comparing the importance of input variables, both networks prioritized reference temperature and humidity, indicating their significant role in microclimate modeling. Long-wave radiation appeared less influential.
Extending predictions to the entire study area, temperature maps illustrated the superiority of the ARPA-trained network in capturing localized variations caused by shadows. However, relative humidity predictions exhibited lower reliability.

Conclusion

In conclusion, this paper presented an advanced microclimate model that combined physics and deep learning techniques to accurately simulate temperature and humidity variations at a 2m scale in the Lombardy foothills. The model integrated physics principles to process climate data and terrain properties, which were then input into a feed-forward neural network for predicting local temperature and humidity. The network was trained and tested using data from 25 sensors in the study area.

Testing with two datasets, ERA5 and ARPA, showed the model's strong accuracy, particularly in temperature prediction. Using neural networks improves prediction accuracy compared to linear regression, acknowledging the complex and non-linear nature of microclimate.

To enhance relative humidity modeling, advanced sensors measuring additional variables could be beneficial. Improved comparisons between global and local climate data would be more meaningful with meteorological stations within the study area. The findings contributed valuable insights for selecting input climate data in microclimate modeling, with implications for optimizing processes in precision agriculture.