Towards Scalable Field-level Crop Yield Estimation through Integration of Crop Model and Deep Learning

Abstract

Accurate crop yield estimation, especially at the field level, has been crucial information for farmers, agricultural companies, and the government. Empirical data-driven model and process-based crop simulation model are two popular methods. The limited site-specific observations make it challenging to build either empirical or crop simulation models over wide geographical regions with such spatial granularity. In this research, a method named hybrid-calibrated crop model with Artificial neural network (HCM-ANN) was prompted, which combined crop model and deep learning algorithm to predict crop yield, with the incorporation of remote sensing data. Agricultural Production Systems sIMulator next generation (APSIMX) was selected as the crop model, and LAI data was derived from Harmonized Landsat Sentinel-2. The study was conducted in Illinois state, US, located in the US corn belt. Results showed that the accuracy of yield estimation using the HCM-ANN model is higher than that of benchmark models, which are calibrated APSIM next generation and deep learning model trained at the county level. Compared to previous yield estimation models, the proposed model has several advantages: 1) It incorporates a process-based model and deep learning, which has a physical foundation; 2} there is no need for ground-measured yield when building the HCM-ANN model, which makes the model more feasible to develop; 3) the model is scalable and generalized, which can apply to county level with fine-tuning. The hybrid-calibrated crop model with ANN approach can also estimate different crop yields with minor modifications.