Utilizing Collocated Crop Growth Model Simulations to Train Agronomic Satellite Retrieval Algorithms.

Area of Science:

• Agricultural Science
• Remote Sensing
• Data Science

Background:

• Satellite remote sensing offers global, high-revisit monitoring of crop health and yields.
• Integrating crop growth models with satellite data can enhance retrieval accuracy.
• Existing methods may lack the precision needed for regional-scale in-season crop monitoring.

Purpose of the Study:

• To develop and validate a novel approach for training satellite-based agronomic retrieval algorithms.
• To utilize collocated crop growth model simulations and satellite measurements for algorithm development.
• To predict in-season crop state variables and yields using bidirectional long short-term memory networks (BLSTMs).

Main Methods:

• Trained BLSTMs using Agricultural Production Systems sIMulator (APSIM) maize simulations and Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data.
• Evaluated algorithm performance using k-fold cross-validation for in-season state variables.
• Compared predictions to United States Department of Agriculture (USDA) ground-truth data for county-level yields and state-level phenology.

Main Results:

• BLSTMs accurately retrieved key maize state variables (leaf area index, biomass) with R² values from 0.4 to 0.8.
• Predicted county-level maize yields with R² values between 0.45 and 0.6.
• Accurately predicted state-level phenological dates (emergence, silking, maturity) with R² values from 0.75 to 0.85.

Conclusions:

• The proposed methodology effectively integrates crop models and satellite data for robust agronomic variable retrieval.
• This approach enables the development of satellite products for monitoring field-scale crop growth globally.
• Future work involves applying this methodology with field-scale simulations and fused satellite data for enhanced precision.