Simulation of Soybean Phenology with the Use of Artificial Neural Networks

Abstract

The phenology of soybean explicitly indicates environmental changes and strongly depends on temperature and day length. We adapted an artificial neural network model to predict the time to flowering in nine early maturing soybean accessions in the Northwest region of Russia. We added scaling constants for network inputs, optimized the high and low temperature thresholds and base day length, and implemented a new model written in Python using the Keras and TensorFlow libraries. Experimental data obtained in 1999–2013 in Pushkin (in the Leningrad region) and Kuban were used for training 121 model parameters; after training the mean-root-square error became smaller, 0.026. The investigated accessions had a reduced upper temperature threshold compared to the literature data (23 instead of 30°С) and increased low temperature threshold (12 instead of 5°С). The extension of day length from 12 to 13 h confirmed the adaptation to a longer day. The average prediction error was improved by approximately 2 days compared to the previous model of temperature minima. We generated daily weather for different future greenhouse gas emission scenarios and predicted time to flowering for nine soybean accessions in a changing climate for 2019–2030 and two planting days, that is, May 1 and May 10. The predicted time to flowering decreases to 2030 for most accessions and scenarios but may remain constant or fluctuate in several cases. The difference in the mean between 39.21 days in the experimental data and 36.33 days in the modeling results for 2030 is statistically significant according to Mann–Witney–Wilcoxon criterion (5423.5, P = 0.0097 < 0.01). Consequently, the results confirmed the predictive power of the developed model.