Estimation and forecasting of soybean yield using artificial neural networks

Abstract

In science, estimation is the calculation of a current value, while forecasting (or prediction) is the calculation of a future value. Both estimation and forecasting are based on covariates. However, whereas estimation enables greater agility in current decision making, forecasting can reveal different strategies for the future. The use of Artificial Neural Networks (ANNs) has brought improvements in accuracy to the estimation and forecasting of agricultural yield for various crops around the world. These models are part of a set of machine‐learning models, becoming an important ally not only to producers, companies, cooperatives, and to government institutions for decisions making and strategic decisions at all levels of the agricultural system. The main constraints of agricultural production are climatic conditions and soil water availability during crop cycles. We propose the use of ANNs for soybean [Glycine max (L.) Merr.] yield estimation and forecasting 2 mo before harvesting in the region of MATOPIBA, the largest and the last agricultural frontier of Brazil. This tropical agricultural area has about 73,173,485 hectares, corresponding to approximately 1.3 times the area of France. The input features for ANN were the monthly climatic conditions of air temperature, precipitation, and global radiation, as well as components of the water balance such as crop evapotranspiration, soil water storage, actual evapotranspiration, water deficiency, and surpluses during the cultivation cycle. The evaluation of ANN for yield estimation had R2 = .88 and RMSE = 167.85 kg ha–1, while the ANN for forecasting obtained R2 = .86 and RMSE = 185.85 kg ha–1.

Maize Yield Prediction using Artificial Neural Networks based on a Trial Network Dataset

“…Using another strategy, maize yield prediction was performed with machine learning with different models, and the best of them, Stacked LASSO, presented a Mean Bias Error (MBE) of 53 kg/ha and a Relative RMSE (RRMSE) of 9.5% [32]. For the soybean crop, annual productivity was predicted in the region of the Brazilian states of Maranhão, Tocantins, Piauí, and Bahia as a function of monthly climate variables (air temperature, precipitation, and global radiation) and water balance components (cultivation evapotranspiration, storage, actual cultivation evapotranspiration, water deficiency, and surplus) during plant development through a deep artificial neural network [33] with a dataset size of 920 examples, the obtained RMSE was 167.85 kg.ha −1 . The unprecedented differential in the strategy of using climatic variables subdivided into maize plant development stages that are differentially sensitive to water stress was assertive.…”

In‐season crop phenology using remote sensing and model‐guided machine learning

“…A hybrid approach to integrating biophysical models and NN methods, known as physics-guided machine learning, has been used in fields such as hydrology and materials science (Willard et al (2020). While studies have shown that NN methods are effective predictors of agronomic targets, such as gross primary production (Peng et al, 2019) and yield (Barbosa dos Santos et al, 2021), studies that guide NNs with crop models are rare. One recent study has shown that machine-learning-based yield prediction can be improved with guidance from existing crop models (Shahhosseini et al, 2021).…”

“…(2020). While studies have shown that NN methods are effective predictors of agronomic targets, such as gross primary production (Peng et al., 2019) and yield (Barbosa dos Santos et al., 2021), studies that guide NNs with crop models are rare. One recent study has shown that machine‐learning‐based yield prediction can be improved with guidance from existing crop models (Shahhosseini et al., 2021).…”

A Novel Approach to Pod Count Estimation Using a Depth Camera in Support of Soybean Breeding Applications

“…These models can automatically learn from data using a multi-layer architecture, learn from the hierarchical outputs of the previous layers, and deal with non-linearities between crop traits and yield estimation [ 31 , 43 , 44 ]. ML algorithms such as random forest and neural networks have been widely used to predict crop yield using data collected remotely [ 5 , 11 , 45 , 46 , 47 , 48 , 49 , 50 ]. It should be noted that the model accuracy is affected by the dates of the predictions [ 44 ], so it is important to consider this factor while making soybean yield predictions using satellite imagery.…”