A new spatial machine-learning model has been developed to improve precision irrigation in vineyards by predicting crop water stress, specifically the Crop Water Stress Index (CWSI). The model, tested in a wine-grape vineyard in the Judean Hills of Israel, predicts spatial variability in water stress at the level of individual vines. By incorporating terrain parameters, soil electrical conductivity (ECa), and Normalized Difference Vegetation Index (NDVI) data, the model offers valuable insights into the vineyard’s water requirements.
A spatial machine-learning model is a model that uses machine learning to analyze spatial data, which is information that represents the shape and location of geometric objects.
The integration of geospatial components, such as vine location, significantly enhances the accuracy of the CWSI predictions, enabling vineyard managers to optimize irrigation practices and reduce water usage. This research was funded by the Ministry of Agriculture and Rural Development of Israel, the European Union’s Horizon 2020 research programme, and the Ministry of Science of Spain.
The research was led by Aviva Peeters, Yafit Cohen, Idan Bahat, Noa Ohana-Levi, Yishai Netzer, Eitan Goldstein, Victor Alchanatis, Tomás Roquette Tenreiro, and Alon Ben-Gal.
Key Factors Driving CWSI Predictions
The model uses terrain parameters, soil electrical conductivity (ECa), and Normalized Difference Vegetation Index (NDVI) data to predict the water stress levels in a vineyard. NDVI, a key indicator of plant vigor, was found to be particularly important in predicting CWSI values. By integrating geospatial components into the model, including the spatial locations of vines, the prediction accuracy was significantly improved.
The model is designed to work with easily accessible data, making it practical for farmers to use in everyday irrigation decisions. By predicting CWSI with minimal input variables, the model offers a cost-effective solution to reduce the need for frequent and expensive thermal imaging campaigns, which are typically used to assess crop water stress.
Optimizing Water Use in Vineyards
The study addresses the growing need to optimize water usage in agriculture, especially in regions prone to drought or where water resources are limited. By creating management zones (MZs) based on CWSI maps, vineyard managers can adjust irrigation more precisely to the varying water needs of different parts of the vineyard. The model aims to minimize costly thermal-imaging campaigns by complementing them with more affordable multispectral imaging, which can produce NDVI maps with high spatial and temporal resolution.
CWSI, which has been shown to be effective in measuring plant water stress, is the focus of this study. In-season maps of CWSI, generated using both static and dynamic variables, can help create MZs that guide irrigation decisions. This model allows vineyard managers to make more informed decisions about water use while maintaining crop health and optimizing yields.
Machine-Learning for Complex, Non-linear Relationships
The model employs machine-learning techniques, specifically random forests (RF), to handle the complex, non-linear relationships between variables. RF models are well-suited for this task due to their ability to process large datasets and account for interactions between multiple variables. The inclusion of spatial autocorrelation in the model allows for more accurate predictions by considering the spatial relationships between vines.
To improve model accuracy, three geospatial approaches were integrated:
- Spatial Location: The (x, y) coordinates of each vine were used as predictors to account for spatial autocorrelation.
- Spatial Autocorrelation: The Getis-Ord Gi* statistic was applied to measure relationships between neighboring data points, and the resulting values were converted into z-scores.
- Weighted Spatial Autocorrelation: Z-scores from the previous step were used as weights to adjust data points, further improving model predictions.
Performance Evaluation and Findings
The model was tested in a vineyard in the Judean Hills of Israel, where Cabernet Sauvignon vines were managed using variable rate irrigation (VRI) across 20 management cells. Key input variables included static data such as slope, aspect, and soil conductivity, as well as dynamic data like NDVI and CWSI.
The results showed that the model performed best when spatial autocorrelation was accounted for through z-score transformations. NDVI emerged as the most important variable for predicting CWSI, followed by terrain and soil factors. For example, in 2017, the best performance was achieved on August 9, while in 2018, it was on August 29. The model was able to predict CWSI values with a high degree of accuracy, particularly at the management cell level.
Also read: Tohoku University Develops Leaf-Mounted Sensor to Monitor Plant Health
However, predictions were less accurate in certain cells, possibly due to extreme variability in CWSI or other factors that were not fully accounted for in the model. The study noted that further research is needed to refine the model, particularly by exploring diurnal and seasonal variations in CWSI, which could help improve irrigation thresholds.
Future Development and Research Directions
While the model has shown promising results in its current form, further work is necessary to explore its application in different geographical areas and with different crops. Testing the model in various environmental conditions, such as different grapevine varieties or soil types, would help determine its versatility and effectiveness across different agricultural settings. Additionally, the study calls for evaluating the economic impacts of adopting this model for irrigation management, including the potential benefits in terms of water conservation and yield optimization.
The research highlights the importance of considering both static (e.g., terrain, soil) and dynamic (e.g., plant water status) variables in precision irrigation models. By minimizing the number of input variables required for accurate predictions, the model offers a practical solution for farmers, helping them make better-informed decisions while keeping costs low. As such, it could serve as the foundation for future decision support tools aimed at optimizing irrigation management in vineyards and potentially in other crops as well.
This study demonstrates the potential of using a spatial machine-learning model based on random forests to predict CWSI in vineyards. By integrating easily accessible data such as NDVI, terrain parameters, and soil conductivity, and by accounting for spatial variability, the model provides an effective tool for precision irrigation. While further research is needed to adapt the model to different contexts, its ability to predict water stress with minimal data offers a promising solution for reducing the costs of irrigation management while optimizing water use.
