Review of Machine Learning Approaches for Modern Agrometeorology

With the development of smart meteorology and precision agriculture, modern agrometeorology tasks demand for efficient analyzing and processing of massive agricultural and meteorological data, including multi-source remote sensing images. Machine learning technology can powerfully contribute to the development of agrometeorology and the innovation of agrometeorological service mode. A targeted overview on the related work of machine learning in modern agrometeorology domains is given, including mapping and zoning, detection and observation, yield prediction, and parameter prediction, with specially focuses on deep learning approaches for agrometeorology and the latest research progress in recent years. From the aspect of mapping and zoning, machine learning technology can be combined with remote sensing images to map land cover and crop types in different scales, and can also be combined with remote sensing data, soil data and statistical data to make thematic maps of crop growth and vegetation quality and to zone crop management areas. From the aspect of detection and observation, machine learning technology is successfully used to detect weeds in field images. Deep learning technology is used in plant phenotype observation, disease and pest detection, obstacles and anomaly detection, fruit counting and so on with high accuracy, which could greatly improve the level of agrometeorology automatic observation. From the aspect of yield prediction, machine leaning technology combined with remote sensing time series data, meteorological data and soil data is successfully used to predict the yield of different crops in different scales. Machine learning technology also has great application potential in loss assessment for agrometeorological disasters. From the aspect of parameter prediction, the hydrological, soil and crop parameters concerned by agrometeorology tasks such as evapotranspiration, leaf area index, soil moisture and nitrogen can be accurately inverted and predicted by the combination of machine learning technology, meteorological data and remote sensing data. Overall, among the traditional machine learning approaches, support vector machine and artificial neural network are the most widely used and the most ideal methods. In recent years, ensemble-based methods such as random forest and gradient boosting machine have generally achieved higher accuracy than kernel methods, while deep learning approaches have achieved higher accuracy than ensemble-based methods in some tasks. In the future, it is necessary to verify the applicability and advancement of more different machine learning approaches, especially deep learning approaches in more different agrometeorological tasks, and choose the most suitable machine learning technology for each specific task in modern agrometeorological services according to the data using, which will help to meet new challenges and opportunities of the modern agrometeorology development.