Enterprise AI Analysis
Use of explainable artificial intelligence with high-resolution satellite imagery to assess water status and vine shoot growth in grapevine (Vitis vinifera L.)
Grapevine is a critical Mediterranean crop threatened by climate change, particularly drought. Traditional water status indicators (Ψpd) are labor-intensive, and a new visual index (iG-Apex) has emerged. This study leveraged remote sensing and machine learning (ML) to non-destructively monitor these indicators. We developed an explainable artificial intelligence (XAI) framework to predict Ψpd and iG-Apex in two rainfed vineyards using high-resolution multispectral imagery from Planet SuperDove. Five ML models (XGBoost, Random Forest, SVR, ElasticNet, Linear Regression) were trained and compared. XGBoost achieved the best performance for Ψpd prediction (R²=0.778), while Random Forest excelled for iG-Apex (R²=0.615). SHAP and ICE analyses revealed that visible and NIR bands were the most important predictors. These results demonstrate the potential of XAI and remote sensing for supporting grapevine management and early water stress detection in Mediterranean conditions.
Executive Impact & Key Metrics
Leverage advanced AI to transform viticulture operations with data-driven insights.
0
XGBoost R² for Ψpd Prediction
0
Random Forest R² for iG-Apex Prediction
0
Satellite Spatial Resolution
0
Total Ψpd Observations
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Agro-climatic Conditions & Ecophysiological Parameters
The study area experienced average daily temperatures peaking at 31.2 °C (DOY 196) and gradually decreasing to 16.2 °C by late September. Rainfall was irregular, with significant events in mid-August and September. Vapor Pressure Deficit (VPD) followed temperature trends, reaching a maximum of 3.25 kPa (DOY 196). Predawn leaf water potential (Ψpd) ranged from -0.82 MPa to -0.16 MPa, with a mean of -0.36 MPa, indicating a left-skewed distribution. The iG-Apex index varied from 0.05 to 0.90, with a mean of 0.56, suggesting a more symmetrical distribution. Temporal variability was observed, with Ψpd values generally decreasing (indicating more stress) from June to August, while iG-Apex showed an inverse trend.
Machine Learning Model Comparison
Five machine learning models (XGBoost, Random Forest, SVR, ElasticNet, and Linear Regression) were evaluated using SuperDove spectral bands (SDBs), vegetation indices (VIs), and a combination of both (BVI) as predictors. For Ψpd prediction, Extreme Gradient Boosting (XGBoost) using BVI achieved the best performance with an R² of 0.778, an RMSE of 0.068, and a MBE of 0.002. For iG-Apex prediction, Random Forest (RF) with BVI demonstrated superior results, yielding an R² of 0.615, an RMSE of 0.138, and a MBE of 0.001. Models performed worst when relying solely on VIs. SVR and EN showed lower performance, and LR provided the weakest outcomes.
SHAP and ICE Analyses for Interpretability
SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE) analyses were employed to provide insights into model predictions. For Ψpd prediction, coastal blue and NIR bands, along with ARVI and PSRI (VIs), were identified as the most influential predictors. High coastal blue reflectance negatively influenced Ψpd, while higher NIR reflectance positively influenced it, indicating non-stressed conditions. For iG-Apex prediction, coastal blue, NIR, red-edge, and green bands, as well as PSRI, were key. Low coastal blue reflectance positively influenced iG-Apex, while high NIR and red-edge reflectance had a positive effect. ICE graphs revealed significant heterogeneity and interaction effects among spectral features, highlighting the models' ability to capture complex, case-specific spectral signatures driven by various plant traits.
Spatial-Temporal Monitoring & Future Outlook
The best-performing XGB (for Ψpd) and RF (for iG-Apex) models were applied to SuperDove imagery from May to September 2022 to generate spatial-temporal predictions. Predicted Ψpd values started at -0.53 MPa (DOY 124) and decreased to -0.57 MPa (DOY 127), generally aligning with ground-based observations and confirming the XGB model's efficiency. Predicted iG-Apex values started at 0.319 (DOY 124) and showed an irregular increase/decrease, though the model tended to underestimate observed values early in the season and overestimate later. Hexagon maps provided a detailed spatial overview, showing lowest median Ψpd values in the northwestern zone of Field 1. The study emphasizes the potential of integrating XAI and remote sensing for non-invasive, accurate, and spatially explicit monitoring, supporting precision viticulture in a climate-stressed Mediterranean region.
0.778
XGBoost R² for Predawn Leaf Water Potential (Ψpd)
0.615
Random Forest R² for Vine Shoot Growth Index (iG-Apex)
Enterprise Process Flow
Ground Truth & Satellite Data Collection
→
Feature Engineering (SDBs & VIs)
→
Machine Learning Model Training & Comparison
→
Hyperparameter Optimization (Optuna)
→
Explainable AI (SHAP & ICE Analysis)
→
Spatial-Temporal Prediction (Hexagon Maps)
| Target | Best Model | R² | Key Predictors |
|---|---|---|---|
| Predawn Leaf Water Potential (Ψpd) | XGBoost | 0.778 |
|
| Vine Shoot Growth Index (iG-Apex) | Random Forest | 0.615 |
|
Precision Viticulture in Mediterranean Rainfed Vineyards
This study provides a practical case for integrating Explainable AI (XAI) with high-resolution satellite imagery to enhance precision viticulture. Conducted in two rainfed 'Syrah' vineyards in southern France, it addresses the critical need for non-destructive, spatio-temporal monitoring of grapevine water status (Ψpd) and shoot growth (iG-Apex) under increasing drought events in Mediterranean regions. The framework successfully demonstrates how advanced ML models, trained with Planet SuperDove multispectral data, can provide accurate predictions and actionable insights for early water stress detection, optimizing resource use, and improving crop resilience.
Quantify Your AI ROI
Estimate the potential annual savings and efficiency gains for your enterprise by implementing similar AI solutions.
Annual Cost Savings
$0
Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A typical phased approach to integrate advanced AI solutions into your enterprise, ensuring a smooth and impactful transition.
Phase 1: Discovery & Strategy Alignment (Weeks 1-4)
Comprehensive assessment of current viticulture practices, data infrastructure, and business objectives. Detailed AI solution design, including satellite data integration, model selection, and XAI framework tailoring. Finalize project scope and success metrics.
Phase 2: Data Engineering & Model Development (Weeks 5-12)
Establish secure pipelines for high-resolution satellite imagery (e.g., Planet SuperDove) and ground truth data. Develop, train, and validate custom machine learning models (XGBoost, Random Forest) for Ψpd and iG-Apex prediction. Integrate SHAP/ICE for model interpretability.
Phase 3: Pilot Deployment & Validation (Weeks 13-20)
Deploy the AI solution in a pilot vineyard environment. Conduct rigorous testing and validation against new ground observations. Collect user feedback for refinement and optimization of the prediction accuracy and explainability features.
Phase 4: Full-Scale Rollout & Ongoing Optimization (Weeks 21+)
Seamlessly integrate the AI framework into existing farm management systems. Provide comprehensive training for agronomists and decision-makers. Implement continuous monitoring, performance tuning, and model updates to ensure long-term value and adaptation to evolving climate conditions.
Ready to Transform Your Viticulture?
Connect with our AI specialists to explore how these cutting-edge insights can be tailored to your enterprise's unique needs.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat