基于时序影像及不同模型的玉米早期估产研究

For the inadequate generalization ability of the quantitative evaluation model of crop yield, the lag of forecasting time and the difficulty of establishing the optimum lead yield estimation time, this paper takes Sentinel-2 remote sensing data and the measured maize yield as the data source to research the establishment of county-scale maize yield estimation and optimum lead yield estimation time. Based on the time-series image data of maize growth-satges, through building the correlation between maize yield measured data and vegetation index, the time-series maize yield estimation model was established by MLRM (multivariate linear regression model), GPR (Gaussian process regression model) and LSTM (Long short-term memory artificial neural network model). The experimental results show that LSTM is superior to GPR and MLRM in terms of the accuracy, and reliability of the yield prediction model, the capture of the abnormal yield value, and the optimum lead yield estimation time in the time series yield estimation model established with NDVI, GNDVI and GN (NDVI and GNDVI combination) as parameters. At the same time, based on the LSTM estimation model, the NDVI time-series image data up to tasseling stage were used as parameters and the yield prediction results showed that the R² (determination coefficient) was 0. 83, RMSE(root mean square error) was 0. 26 t • ha¹, RPD(relative percent deviation) was 3. 52; The GNDVI time-series image data up to tasseling stage were used as parameters, and the yield prediction results showed that the R² was 0.79, RMSE was 0. 30 t • ha¹, RPD was 2. 87; The GN time-series image data up to tasseling stage were used as parameters, and the yield prediction results showed that the R² was 0.83, RMSE was 0.27 t • ha¹, RPD was 3. 05. Using the NDVI time-series image data as the LSTM model parameter has the optimal yield estimation, and the maize yield could be predicted 2 months in advance compared with the maize harvest stage. As a result, we developed a crop yield forecasting method in this study to predict crop yield for county-scale. It has practical significance for maize yield forecasting and provides a relevant reference for similar crop yield estimation research.