Application of neural network technologies for crop yield forecasting

. The article discusses the problems of crop yield forecasting. A review of relevant bibliographic sources has been carried out. Crop yield factors are analyzed. Climatic and meteorological features, production and agrotechnical parameters, soil moisture and fertilizers are considered as the main factors. It is noted that a full-scale field experimental study of crop yields requires very large time and financial costs. A conclusion was made about the issues related to modeling of product forecasting. A neural network for predicting crop yields (in case of winter wheat) was built by using the Python PyTorch machine learning library.


Introduction
The issues associated with forecasting crop yields are one of the urgent problems of our time.At the same time, a current and ever increasing volume of information requires the use of new computer technologies.In particular, a good example of this technology lies within the methodology of neural network modeling.The methodology helps us identify new dependencies and patterns by using deep analysis and machine learning methods.These days, neural networks have wide areas of application and give prospects for solving various tasks, including the ones in the field of agricultural and industrial production.For instance, the following tasks can be attributed to the prospects of using neural networks in agriculture: -In pattern recognition and classification: classification and sorting of harvested crops, identification of weeds, recognition of product defects.
-In diagnostics: assessment of soil quality and the condition of plant crops, diagnostics of technical equipment malfunctions, quality control of manufactured products.
-In data clustering: landscape monitoring for verification of agricultural land.
-In forecasting: forecast of the harvest of various crops.
-In optimal management: control over the implementation of sowing operations etc.
It is obvious that a full-scale field study of agricultural yields requires substantive financial and time costs.In this regard, the issues of forecasting crop yields are becoming more relevant.The yield forecast is necessary for calculating the land need for the areas of crops, building a model of the balance of market needs and production, etc.
Although a significant number of works has already been devoted to the issues of computer modeling and forecasting crop yields, this problem cannot be considered fully developed and solved.The forecasting model is a complex multiparametric system that includes a number of technical, algorithmic and software subsystems.
The purpose of this work is to study the theoretical aspects and practices of using neural networks in agricultural and industrial production, as well as to develop and implement a neural network model for predicting crop yields by using an example case of winter wheat yields.

Neural network modeling of crop yields
Let's consider using neural networks to assess crop yields.
The work [1] provides an overview of papers related to the application of machine learning methods in agriculture.At the same time, it is noted that "crop management was observed to be at the centre of attention".In addition, "maize and wheat as well as cattle and sheep were the most investigated crops and animals".The authors also discuss the role of sensors attached to satellites and unmanned ground and air vehicles as a way of obtaining reliable input data for information analysis.
The article [2] provides an overview of research papers that use convolutional neural networks (CNN) used to solve various problems in agriculture and food production.In addition, an example of using CNN for the problem of identifying missing vegetation in a sugarcane plantation is discussed.
The issues of yield prediction are discussed in [3] in the example of kiwi yield prediction based on the concentration of nutrients in the leaf.To this end, a study was conducted using an artificial neural network.Data analysis was carried out using a multilayer perceptron.
The work [4] describes the use of neural network technologies in predicting the soils fertility level of leguminous crops and their yield.As a result, it is said that the greatest increase in yield in the studied areas is provided by the regulation of acidity and nitrogen content, as well as the value of phosphorus content.
Aspects of choosing algorithms and software environments for neural network forecasting of crop yields using retrospective data are considered in the article [5].The object of the study in the work was the time series of the accumulated long-term statistics of the yield of a group of grain crops.
Based on the analysis of the literature, it can be concluded that at the moment the methodology of neural network modeling is mainly research in nature, without wellestablished proven methods.There is no unified concept for choosing training sample parameters for building and training a model.There is a lack or inaccessibility of empirical data.

Yield factors for winter wheat
It is known that the yield of agricultural crop production is influenced by a very large number of various factors.In [6] it is proposed to combine the factors into the following groups: -Production and agrotechnical. -Ground.
-Agrometeorological, including natural and climatic characteristics.At the same time, a regression analysis model is built, showing the relationship between the yield of winter wheat and the average daily May temperature.
The article [7] considers the issues of analyzing the yield of winter wheat varieties Zolotokosa, Ogradskaya, Favoritka.At the same time, the following main factors affecting the formation of yields were identified: sowing time, weather conditions, the level of mineral nutrition of plants, the influence of the predecessor crop and seed quality.
Thus, weather conditions encompass parameters like minimum temperatures, average daily temperatures, total precipitation, and average daily snow coverage.Among the main sources of mineral nutrition are organic fertilizers (manure, compost, peat, plant residues), mineral fertilizers (ammonium nitrate, urea, etc.), as well as soil nutrients.
An overview of the yield factors of winter wheat in Non-chernozem regions of Russia is given in [8].The purpose of this work is to study the effect of mineral fertilizers on the realization of the potential productivity of winter wheat varieties bred by the Federal Research Center Nemchinovka.At the same time, factors like agrometeorological conditions, fertilizers (phosphorus, potash, nitrogen), and microfertilizers are considered.
The work [9] is devoted to the issues of the influence of meteorological factors on the yield of winter wheat.By discussing the variety of meteorological factors, the authors conclude that the amount of precipitation in winter and soil moisture in autumn are of decisive importance.
The article [10] considers the role of abiotic factors and agricultural practices in the formation of winter wheat yield.The main yield factors are the hydrothermal conditions of the pre-sowing period and the reserve of productive moisture in the soil.
Identification of the influential environmental factors on the yield of winter wheat varieties is considered in the article [11].It is noted that the main factors affecting the indicators of the final harvest are climatic features, temperature, soil moisture and applied fertilizers.

Results
In order to study the possible use area and problems of applying neural network models to the problems of predicting crop yields, a model example of a neural network was built.
Based on the materials of statistical services and tabular data [12], a data array was prepared that is used to train the neural network.• Yield of winter wheat, centner/hectare.The Python Pandas library modules as well as PyTorch deep learning library were used to implement the neural network.
The structure of the neural network contains three layers, the number of neurons in the hidden layer is 10, the activation function is Sigmoid.Torch.optim.Adam was used as an optimizer to perform gradient descent steps.
The original Dataset was split into training data and a validation dataset to test the model.It is known that in the course of training, the neural network adjusts its synaptic weights, as a result of which functional dependencies between input and output data are revealed and their generalization is performed.This means that a successfully trained neural network is able to calculate a predictive value for data that is not in the training set.

Discussion
The neural network was trained on 2000 epochs.The results of the dynamics of improving the loss function are shown in figure 1.As shown on the graph, the convergence of the gradient descent method takes place, and as the number of epochs increases, the results of data analysis improve.Based on figures from graph 2, it can be concluded that the quality of the constructed neural network model is satisfactory.
By using the Data Analysis MS Excel package, a correlation analysis of the mutual influence of the factors of the training data set was further carried out.It was found that the following columns have the highest correlation dependence with the column "yield of winter wheat, centner/hectare": "max soil moisture in period No. 3" with correlation coefficient 0.776, "max soil moisture in period No. 4" with 0.775 and "amount atmospheric precipitation, %" with 0.772, which indicates the greatest influence of these parameters on the simulation result.

Conclusion
We can single out a number of the following problems associated with computer modeling of crop forecasts based on neural network models: -The amount of available diversified systematic data of meteorological and climate observations for building, training and testing the model is not enough.
-Data on the chemical composition of soils that affects the size of the crop is incomplete.
-The issue of the optimal structure of neural network models (dimensions of the training data sample, the number of neurons and layers in the neural network, etc.) has not been sufficiently studied.
At the same time, it can be said that the problem of yield forecasting belongs to the class of partially formalized problems that allow ambiguity in the modeling parameters.Nevertheless, it should be noted that the problem of crop forecasting is in great demand, which makes this simulation relevant in terms of its tasks and goals.
• Min soil moisture in period No. 4 (during the earing period: the ear exits from the main leaf),• Max soil moisture in period No. 4, • Hydrothermal coefficient,• Sum of active temperatures, % to norm,• Amount of atmospheric precipitation, % to the norm,