Abstract—The possibilities of using machine learning for estimating the effect of the complex of weather and agrotechnical factors on the yield of agricultural crops and for yield forecasting were investigated. Numerical simulations were carried out using materials of long-term field experiments in the forest-steppe zone of Kemerovo oblast. Continuous observation data for 2013–2018 for the main crops, wheat and barley, were used to train the model. The Random Forest Classifier machine-learning algorithm was used for calculations. The accuracy was defined as the ratio of the number of correct predictions for the test sample to the total number of test cases. When the model was trained using information on current agricultural practices and weather conditions of the previous year (average monthly temperatures and precipitation) as input data, the accuracy for wheat was 0.81, 0.87 for barley, and 0.84 on average for the crops. To estimate the information content in the data of weather fluctuations of the previous year and the effect of the agronomic factors of the current year on the accuracy of the yield forecast, the model was trained in two alternative ways using modified input data. In one case, we considered only the weather image of the previous year based on the monthly average temperature and precipitation. The second case accounted for the agricultural practices used, while the weather data were reduced to just one value, the average annual temperature. The accuracy for the crop (without distinction between barley or wheat) in the case of considering only the weather factors of the previous year was 0.7. In the case of accounting mainly for agricultural practices, with minimal consideration of weather factors, the accuracy was 0.73. These results suggest that each of the groups of factors considered (weather for the previous period and planned agricultural practices) made a comparable contribution to the expected accuracy of the yield forecast.

摘要—研究了​​使用机器学习来估计天气和农业技术因素的复杂性对农作物产量的影响以及进行产量预测的可能性。使用克麦罗沃州森林草原地区的长期野外实验材料进行了数值模拟。该模型使用了2013-2018年主要农作物小麦和大麦的连续观测数据。计算使用了随机森林分类器机器学习算法。准确度定义为测试样本的正确预测数与测试用例总数之比。当使用有关当前农业实践和上一年天气状况（平均每月温度和降水）的信息作为输入数据对模型进行训练时，小麦的准确度为0.81，大麦的准确度为0.84，农作物的平均准确度为0.84。为了估计上一年的天气波动数据中的信息内容以及当年的农学因素对单产预报准确性的影响，使用修改后的输入数据以两种替代方式对模型进行了训练。在一个案例中，我们仅根据月平均温度和降水量考虑了上一年的天气图像。第二种情况说明了所使用的农业实践，而天气数据减少到仅一个值，即年平均温度。仅考虑上一年的天气因素时，农作物的准确性（大麦或小麦之间没有区别）为0.7。在主要以农业作法核算的情况下，在不考虑天气因素的情况下，准确度为0.73。这些结果表明，考虑的每组因素（上一时期的天气和计划的农业实践）对单产预测的预期准确性做出了可比的贡献。