Abstract:
Winter wheat is one of the most important food crops in China. Accurate and rapid forecasting of winter wheat yield at the county level has been a high demand for the decision-making on the grain in precision agriculture. The current forecast of winter wheat yield is commonly used the crop models, meteorological statistics, or remote sensing data. It is very necessary to further improve the accuracy of yield forecasting on the winter wheat, particularly for the refine prediction at the county level. In this research, a forecast model of winter wheat yield was proposed at the county level to combine the Sentinel-2, MODIS Enhanced Vegetation Index (EVI) remote sensing data, and cultivated land distribution using the Convolutional Neural Networks (CNNs) and the Back Propagation (BP) neural networks. A machine learning technology was also used to evaluate the forecast of winter wheat yield. Specifically, two modules were constructed for the planting area extraction and the yield forecast. Specifically, an improved CNN was selected to extract and identify the spectral features of winter wheat planting areas from the Sentinel-2 remote sensing data. An improved BP neural network was also used to forecast the yield of winter wheat for the subsequent residual block module. The geological features were calculated to compare the cultivated land distribution in the raw MODIS EVI remote sensing data in the yield forecast module after the identification of planting areas. The correlation between the specific information and the winter wheat yield data was verified to forecast the yield using the improved BP neural networks. Some indicators were calculated to evaluate the forecast performance and accuracy of the improved model, including the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). The experiment results showed that a better performance was achieved for the county-level winter wheat yield forecast using CNN and BP neural networks in the validation set from Shandong Province, China (2014-2016), where the R2 value was above 0.87, while the MAEs and RMSEs were lower than 269.48 and 346.56 kg/hm2, respectively. There were 52.34% of counties in Shandong Province with a relative error between the forecast and the actual yield of less than 3%, whereas, 92.98% of counties with a relative error within 9%, indicating a less than 1.2% deviation between the simulation and experiment. In the validation set from Henan Province (2015-2019), the R2 value was above 0.96, while the MAEs and RMSEs were lower than 304.84 and 418.14 kg/hm2, respectively. There were 45.90% of counties in Henan Province with a relative error between the forecast and the actual yield of less than 3%, while 91.05% of counties with a relative error within 9%, indicating a less than 1.6% deviation between the simulation and experiment. In conclusion, the higher forecast accuracies were achieved in the verification sets of Shandong and Henan Province for the county-level winter wheat yield forecast using CNN and BP neural networks. The better performance of the model can be expected to deal with the random input data, indicating the excellent generalization and high robustness for the accurate and stable forecast of winter wheat yield.