Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 149607

149607 - A Hybrid Crop Modeling Approach for Predicting Biomass Production of Sorghum 

Sorghum (Sorghum bicolor (L.) Moench), as a versatile C4 drought-resistant and nutritionally valuable crop, is integral to food security and biofuel production in many parts of the world. Biomass sorghum yields can be influenced by a multitude of factors, including environmental conditions, management practices, and genotypic variability. Reliable yield predictions are essential for optimizing agronomic interventions, resource allocation, and supply chain logistics. Consequently, researchers have explored various modeling approaches, ranging from process-based crop simulations to data-driven crop models, to address this challenge. 

Process-based crop models have been widely used to predict sorghum yield by integrating various physiological processes, environmental factors, and management practices. Process-based models describe processes at a fine-scale timebase; they incorporate many parameters that require a large number of field experiments to calibrate. Unlike the process-based models which are developed based on the human-knowledge of the crop science, the data-driven models aim to build a mathematical relationship between the input data and the output. The data-driven modeling approach has two major limitations. First the the blackbox structure between the input and output data makes their results are sometimes less insightful due to the lack of knowledge on the crop science. Another difficulty is that model performance is sensitive to data quantity and quality, which makes it difficult to apply this to other crops with insufficient data. 

In this paper, we propose a new hybrid crop model for the biomass sorghum growth process simulation. The proposed model uses a detailed descriptive sorghum growth model to describe how sorghum phenotypes are determined by genotype, environment, management and their interactions during the growth period; data-driven techniques are used to calibrate genotypic parameters from field experimental data. Unlike conventional crop models, the proposed model treats genotypes as variables rather than input sets, separating the impact of genotype, environment and management on sorghum growth timing and phenotype from the parameter definition. We can thus overcome the complex coefficient calibration process in the process-based model and avoid obtaining uncertain parameters derived from field experiments where the results are inevitable under a combined influence of genotype, environment and management. Another strength of our hybrid crop model is that it offers a flexible framework comprising multiple modules that mirror the crop growth physiology. The specific combination of these modules is contingent upon the accessibility of data. In contrast to other crop models which mandate a predetermined set of datasets, this adaptive approach eliminates the need to impute or make assumptions about missing or unavailable data before modeling can begin. 

Our sorghum growth model is designed based on the available data mentioned before. It has a customized module structure according to the granularity of modeling approach that is afforded by available data. In this paper, we are focusing on tracking the phenotype during the sorghum growth process, especially the total biomass weight which is determined by the dry leaf wright plus the dry stem weight. Based on these, we build a sorghum growth model based on the module structure contains stress, tillering, growth, water, photosynthesis, transpiration, respiration, and phenology. Our hybrid crop model approach can separate input data, output data, genotype specific properties, intermediate variables, and output variables. Other crop models like APSIM and DSSAT use parameters that are jointly determined by genotype and environment interactions. The hybrid crop approach calibrate the parameters using the data rather than using predetermined coefficients. This difference provides the advantages includes combination with most state of the art data-driven calibration algorithms and adapting with breeding algorithms.

The training results that average relative root mean square error (RRMSE) is quite low (with an error of 8.87%). The test RRMSE 8.48% is similar to the train result. The model also indicates the temporal progression of plant height and biomass accumulation in various organs (stem, leaves, roots) during the growth cycle. And we also performed another series of tests to check whether the input planting and harvest dates are the best options for the weather in 2019.

Presenting Author: Yanbin Chang Oklahoma State University

Presenting Author Biography: Yanbin Chang is a fourth-year Ph.D. student in the School of Industrial Engineering & Management at Oklahoma State University. His research focuses on developing advanced crop yield prediction models using a combination of optimization techniques, machine learning algorithms, and deep learning architectures.

Authors:

Yanbin Chang Oklahoma State University
Ni Zheng Oklahoma State University
Maria G Salas-Fernandez Iowa State University
Lizhi Wang Oklahoma State University