This paper compares different artificial intelligence (AI) models in order to develop the best crop yield prediction model for the Midwestern United States (US). Through experiments to examine the effects of phenology using three different periods, we selected the July–August (JA) database as the best months to predict corn and soybean yields. Six different AI models for crop yield prediction are tested in this research. Then, a comprehensive and objective comparison is conducted between the AI models. Particularly for the deep neural network (DNN) model, we performed an optimization process to ensure the best configurations for the layer structure, cost function, optimizer, activation function, and drop-out ratio. In terms of mean absolute error (MAE), our DNN model with the JA database was approximately 21–33% and 17–22% more accurate for corn and soybean yields, respectively, than the other five AI models. This indicates that corn and soybean yields for a given year can be forecasted in advance, at the beginning of September, approximately a month or more ahead of harvesting time. A combination of the optimized DNN model and spatial statistical methods should be investigated in future work, to mitigate partly clustered errors in some regions.

This paper describes the development of an optimized corn yield prediction model under extreme weather conditions for the Midwestern United States (US). We tested six different artificial intelligence (AI) models using satellite images and meteorological data for the dominant growth period. To examine the effects of extreme weather events, we defined the drought and heatwave by considering the characteristics of corn growth and selected the cases for sensitivity tests from a historical database. In particular, we conducted an optimization for the hyperparameters of the deep neural network (DNN) model to ensure the best configuration for accuracy improvement. The result for drought cases showed that our DNN model was approximately 51–98% more accurate than the other five AI models in terms of root mean square error (RMSE). For the heatwave cases, our DNN model showed approximately 30–77% better accuracy in terms of RMSE. The correlation coefficient was 0.954 for drought cases and 0.887–0.914 for heatwave cases. Moreover, the accuracy of our DNN model was very stable, despite the increases in the duration of the heatwave. It indicates that the optimized DNN model can provide robust predictions for corn yield under conditions of extreme weather and can be extended to other prediction models for various crops in future work.

Many process-based models for carbon flux predictions have faced a wide range of uncertainty issues. The complex interactions between the atmosphere and the forest ecosystems can lead to uncertainties in the model result. On the other hand, artificial intelligence (AI) techniques, which are novel methods to resolve complex and nonlinear problems, have shown a possibility for forest ecological applications. This study is the first step to present an objective comparison between multiple AI models for the daily forest gross primary productivity (GPP) prediction using satellite remote sensing data. We built the AI models such as support vector machine (SVM), random forest (RF), artificial neural network (ANN), and deep neural network (DNN) using in-situ observations from an eddy covariance (EC) flux tower and satellite remote sensing data such as albedo, aerosol, temperature, and vegetation index. We focused on the Gwangneung site from the Korea Regional Flux Network (KoFlux) in South Korea, 2006–2015. As a result, the DNN model outperformed the other three models through an intensive hyperparameter optimization, with the correlation coefficient (CC) of 0.93 and the mean absolute error (MAE) of 0.68 g m⁻² d⁻¹ in a 10-fold blind test. We showed that the DNN model also performed well under conditions of cold waves, heavy rain, and an autumnal heatwave. As future work, a comprehensive comparison with the result of process-based models will be necessary using a more extensive EC database from various forest ecosystems.

Evapotranspiration (ET) is an important component of the Earth’s energy and water cycle via the interaction between the atmosphere and the land surface. The reference evapotranspiration (ET₀) is particularly important in the croplands because it is a convenient and reasonable method for calculating the actual evapotranspiration (AET) that represents the loss of water in the croplands through the soil evaporation and vegetation transpiration. To date, many efforts have been made to retrieve ET₀ on a spatially continuous grid. In particular, the Moderate Resolution Imaging Spectroradiometer (MODIS) product is provided with a reasonable spatial resolution of 500 m and a temporal resolution of 8 days. However, the applicability to the local-scale variabilities due to complex and heterogeneous land surfaces in countries like South Korea is not sufficiently validated. Meanwhile, the AI approaches showed a useful functionality for the ET₀ retrieval on the local scale but have rarely demonstrated a substantial product for a spatially continuous grid. This paper presented a retrieval of the daily reference evapotranspiration (ET₀) over a 500 m grid for croplands in South Korea using machine learning (ML) with satellite images and numerical weather prediction data. In a blind test for 2013–2019, the ML-based ET₀ model produced the accuracy statistics with a root mean square error of 1.038 mm/day and a correlation coefficient of 0.870. The results of the blind test were stable irrespective of location, year, and month. This outcome is presumably because the input data of the ML-based ET₀ model were suitably arranged spatially and temporally, and the optimization of the model was appropriate. We found that the relative humidity and land surface temperature were the most influential variables for the ML-based ET₀ model, but the variables with lower importance were also necessary to consider the nonlinearity between the variables. Using the daily ET₀ data produced over the 500 m grid, we conducted a case study to examine agrometeorological characteristics of the croplands in South Korea during the period when heatwave and drought events occurred. Through the experiments, the feasibility of the ML-based ET₀ retrieval was validated, especially for local agrometeorological applications in regions with heterogeneous land surfaces, such as South Korea.