Abstract: Real-time and accurate acquisition of chlorophyll content information is of great significance for timely understanding crop damage degree, guiding agricultural production, and estimating yield. To explore the optimal estimation model for the soil and plant analyzer development (SPAD) in different leaf positions of waterlogged winter wheat, a plot experiment with a waterlogging stress gradient in winter wheat field under controlled drainage and irrigation was established. The correlation between 15 commonly used hyperspectral indices and SPAD was analyzed. The SPAD estimation results of waterlogged winter wheat leaves were compared using hyperspectral indices combined with the multiple linear regression, support vector machine, BP neural network, decision tree, and random forest models. The results showed that compared with normal winter wheat, there was no significant difference for SPAD and the value of hyperspectral reflectance under short-term waterlogging (less than 3 d). The SPAD was significantly decreased after more than 9 d waterlogging, and the value was close to 0 in the later growth period. The 15 hyperspectral indices were all correlated with the SPAD (P<0.05), with the correlations between SPAD and the four indices (Ctr2, Dy, NDVI and SIPI) being the highest with the absolute correlation coefficient of 0.880, 0.868, 0.868 and 0.833, respectively. Compared with the SPAD estimation of the L1, L2, and L3 leaves, the results of average SPAD were the best, with the R² of 0.719. The SPAD estimation models constructed with random forest model was the best for different leaf positions, with R², RMSE and RE of 0.824, 4.359 and 2.96%, respectively. Therefore, the average SPAD value could be used to estimate the SPAD of waterlogged winter wheat based on hyperspectral technology, and the random forest estimation model was best.

Key words: hyperspectral characteristic index, SPAD, winter wheat, waterlogging, random forest model