Agricultural Drought Model Based on Machine Learning Cubist Algorithm and Its Evaluation

Abstract

Soil moisture is the most direct evaluation index for agricultural drought. It is not only directly affected by meteorological conditions such as precipitation and temperature but is also indirectly influenced by environmental factors such as climate zone, surface vegetation type, soil type, elevation, and irrigation conditions. These influencing factors have a complex, nonlinear relationship with soil moisture. It is difficult to accurately describe this non-linear relationship using a single indicator constructed from meteorological data, remote sensing data, and other data. It is also difficult to fully consider environmental factors using a single drought index on a large scale. Machine learning (ML) models provide new technology for nonlinear problems such as soil moisture retrieval. Based on the multi-source drought indexes calculated by meteorological, remote sensing, and land surface model data, and environmental factors, and using the Cubist algorithm based on a classification decision tree (CART), a comprehensive agricultural drought monitoring model at 10 cm, 20 cm, and 50 cm depth in Gansu Province is established. The influence of environmental factors and meteorological factors on the accuracy of the comprehensive model is discussed, and the accuracy of the comprehensive model is evaluated. The results show that the comprehensive model has a significant improvement in accuracy compared to the single variable model, which is a decrease of about 26% and 28% in RMSE and MAPE, respectively, compared to the best MCI model. Environmental factors such as season, DEM, and climate zone, especially the DEM, play a crucial role in improving the accuracy of the integrated model. These three environmental factors can comprehensively reduce the average RMSE of the comprehensive model by about 25%. Compared to environmental factors, meteorological factors have a slightly weaker effect on improving the accuracy of comprehensive models, which is a decrease of about 6.5% in RMSE. The fitting accuracy of the comprehensive model in humid and semi-humid areas, as well as semi-arid and semi-humid areas, is significantly higher than that in arid and semi-arid areas. These research results have important guiding significance for improving the accuracy of agricultural drought monitoring in Gansu Province.