Abstract:To estimate crop yield in field management, hyperspectral data and yield data during the reproductive growth period of soybeans through two years of field experiments were collected. Seven spectral indices were calculated based on first-order spectral reflectance at various growth stages. These indices included the ratio index (RI), difference index (DI), normalized difference vegetation index (NDVI), soil-adjusted vegetation index (SAVI), triangular vegetation index (TVI), modified normalized difference index (mNDI), and modified simple ratio (mSR). A correlation analysis between the spectral indices and soybean yield data were conducted by using the correlation matrix method. The best wavelength combinations to be used as the optimal spectral indices related to soybean yield were extracted. Finally, the five spectral indices with the highest correlation coefficients with soybean yield at different growth stages were selected as input variables for the model. Support vector machine (SVM), random forest (RF), and back propagation neural network (BPNN) were utilized to construct soybean yield estimation models and conducted validation. The results indicated that the spectral indices calculated at different growth stages (full flowering stage (R2), full pod stage (R4), and seed filling stage (R6)) all exhibited a correlation coefficient greater than 0.6 with yield, showing a strong correlation. Among these, the spectral index FDmSR at the full pod stage had the highest correlation with soybean yield, reaching 0.717. The optimal model for soybean yield estimation was built using first-order spectral indices from the full pod stage in combination with RF as input variables, achieving a validation set R2 of 0.85, and RMSE and MRE values of 272.80kg/hm2 and 5.12%, respectively. The research outcome can provide a theoretical basis and practical reference for crop yield estimation based on hyperspectral remote sensing technology.
