Our validation phase has been critical in confirming the accuracy and reliability of our innovative approaches for monitoring food security and safety in Africa. Through extensive field campaigns, remote sensing, and statistical analysis, we have assessed our crop disease models, mycotoxin prediction tools, and the AQUACROP crop growth model. In the following sections we highlight key- findings.
Crop Disease Model Validation
Field Data Integration:
Over 150 parcels in Kenya and Uganda were monitored, gathering data on crop conditions (maize, wheat, and rice) and disease symptoms. Detailed rust scoring surveys were conducted during critical growth stages.
Remote Sensing & Satellite Data:
Despite challenges such as heavy cloud cover during the critical rust period in Kenya, the project successfully integrated alternative data sources (e.g., Landsat 8 imagery) to extract relevant vegetation indices for model calibration.
Model Performance:
- In Kenya, separate models for wheat and maize were developed, yielding moderate predictive power (with R² values ranging from 0.246 to 0.591).
- By combining the datasets from both crops and regions, an enhanced predictive model was achieved with an R² of approximately 0.758 and an adjusted R² of 0.342—demonstrating robust capacity to explain disease variance while minimizing overfitting.
Mycotoxin Prediction Model Validation
Aflatoxin B1 (AFB1) Assessment:
Laboratory analyses on maize samples confirmed low concentrations of AFB1. The predictive model accurately indicated low risk for 90.4% of the parcels, ensuring that production remains within the safe limits.
Deoxynivalenol (DON) Assessment:
The DON model was validated using both predicted outputs and laboratory data:
- Most samples (78.8%) were correctly classified as low risk.
- Overall, the model achieved a global accuracy of around 76.9%, supporting its use as an early-warning tool for mycotoxin contamination.
AQUACROP Crop Growth Model Validation
Data-Driven Calibration:
The AQUACROP model was calibrated using field-collected crop data and remote sensing information from Sentinel-2 imagery.
Key parameters such as NDVI, Leaf Area Index (LAI), and Canopy Cover (CC) were derived from time-series imagery, ensuring accurate simulation of crop growth.
Yield Prediction:
The calibrated model successfully reproduced the canopy development and yield formation patterns observed in the field, reinforcing the model’s value as a decision-support tool for agricultural management.
Validation Results – Algorithm Performance and Product Specifications
Our advanced Earth Observation processing algorithms have undergone rigorous testing to ensure that our fusion and modeling approaches deliver precise, reliable results for crop monitoring and food security applications.
Key validation highlights include, Thermal Data Fusion & Diurnal Temperature Cycle Modeling.
Spatiotemporal Fusion Accuracy:
Our deep learning-based Spatiotemporal Thermal Data Fusion Network successfully integrates high-resolution thermal data with coarser resolution datasets. Validation metrics—including low RMSE and MAE, as well as high PSNR values—demonstrate that our network effectively downscales coarse Land Surface Temperature (LST) data to a finer spatial resolution that closely matches observed values.
Accurate Diurnal Temperature Modeling:
By applying an advanced Diurnal Temperature Cycle model, we accurately fit temperature curves to MODIS LST observations. This model accounts for the non-linear changes in surface temperature throughout the day, allowing us to predict LST at any given time with high precision and to normalize temporal differences between image acquisitions.
Hyperspectral Data Fusion & Radiative Transfer Model Inversion
Enhanced Image Fusion:
Our innovative fusion framework combines hyperspectral and multispectral imagery using deep learning techniques. The resulting fused products show high structural similarity and maintain essential biophysical details, ensuring that critical information for crop monitoring is preserved.
Remote Sensing & Satellite Data:
Despite challenges such as heavy cloud cover during the critical rust period in Kenya, the project successfully integrated alternative data sources (e.g., Landsat 8 imagery) to extract relevant vegetation indices for model calibration.
Reliable Parameter Retrieval:
Inversion techniques based on radiative transfer models have been validated against field measurements. The retrieval of key biophysical parameters, such as the Leaf Area Index (LAI), exhibits a strong correlation with in situ data, confirming the reliability of our methods for accurate agricultural assessments.
Overal Impact
Seamless Data Integration:
Our processing pipelines effectively harmonize data from multiple satellite sources, overcoming challenges related to varying spatial and temporal resolutions. Advanced quality control measures—such as robust cloud masking and precise radiometric scaling—ensure that only high-quality data are used in our analyses.
Operational Readiness:
The validated algorithms provide dependable, actionable insights that support data-driven decision-making in agriculture. By accurately capturing crop health dynamics and environmental conditions, our solutions are well positioned to enhance food security and optimize resource management across the region.