Modelling Crop Yield in Agricultural Regions Using Computer Vision Technology

Abstract

The article examines new methodologies for modelling crop yield in agricultural regions of Russia based on the use of remote capabilities to get information on the field state. The proposed approach can be applied to develop indicator systems and create methodological platforms and models necessary to obtain more accurate estimates. In comparison with the traditional regression model, this method uses computer vision technology to gather additional data. Statistical hypothesis testing confirmed the significance of satellite photographs of fields for improving the accuracy of crop yield forecasting models. Traditional econometric tools were compared with various neural networks in order to discover the optimal model. The proposed tools were tested using data from 100 agricultural fields located in municipalities of 43 Russian regions, selected in proportion to the volume of crop production in this region. The conducted analysis showed the advantage of the mixed data neural network in comparison with other neural (multilayer perceptron and convolutional neural network) and regression models. In conditions of uncertainty and a large amount of data, the mixed data neural network can help obtain more accurate estimates. Additionally, while environmental factors have different effects on crop yields, they must be considered along with socio-economic characteristics. The use of new models and data types differing from table information can significantly improve the forecasting accuracy and interpretation. The analysis results can be used for examining and monitoring agricultural production in regional municipalities, determining farm resource requirements, as well as for creating sectoral and comprehensive projects and programmes for the development of the agricultural industry.