Optimization of the soil moisture model based on hyperspectral inversion by integrating wavelet features and growth parameters of winter wheat

Soil moisture is one of the most important factors to affectaffecting the crop growth. An accurate estimation on of the soil moisture content can greatly contribute to the precision irrigation in modern agriculture. Hyperspectral data with more radiation information can provide a feasible solution to the crop soil moisture. However, the monitoring model driven by the soil moisture content can be varied in the different sensors and environments. In this study, a soil moisture monitoring model was constructed to combine the wavelet characteristics of the winter wheat canopy spectrum and leaf physiological parameters. The test site was located in the Yangling District, Shaanxi Province, China (34°17'42'N, 108°04'02'E) in the semi-humid and arid climates. Four water treatments were set up in the experiment, with three replicates and a total of 12 experimental plots. There were the consistent experimental treatments in 2021 and 2022. Non-imaging spectral data and soil moisture content under four water treatments were obtained on March 23, April 8, and April 30, 2021. The winter wheat canopy imaging and non-imaging hyperspectral data, leaf moisture content, leaf area index, chlorophyll, and field soil moisture content were then collected on February 25, March 28, April 2, April 13, April 20, April 21, May 2, May 11, and May 16, 2022. The Savitzky-Golay (SG) was used to smooth the wheat canopy spectrum. The Mexican hat wavelet family (the wavelet function in the family was Mexh) was selected as the governing function to perform an 8-scale (2¹, 2², 2³, ···, 2⁸) continuous wavelet transform on the spectral data. After that, the optimal wavelet transform scale was determined to analyze the correlation between the wavelet coefficients at different scales and soil moisture, chlorophyll, leaf area index, and leaf water content. The variable projection importance analysis was implemented to obtain the wavelet features sensitive to different physiological and biochemical indexes. Finally, the partial least squares (PLS) regression was used to monitor the soil water content of winter wheat roots. The results showed that the changes of in chlorophyll and leaf area were represented by the main influencing factors of soil moisture content on the canopy spectrum of winter wheat. The small-scale wavelet transform was enhanced the connection with the spectrum and soil moisture. In addition, it was feasible to monitor the soil moisture content of winter wheat using the physiological indexes from the wavelet features. The selected features were more applicable than those of the soil moisture content. A comparison was made on the accuracy of the models under different features. Among them, the soil moisture monitoring model with the wavelet features using chlorophyll in different year data 2021 hyperspectral non-imaging data set and 2022 imaging different scale data set performed best (R²=0.541-0.687), were better than the SMC driven model (R²=0.274-0.572). In summary, the spectral characteristics of winter wheat leaf chlorophyll and continuous wavelet transform were more applicable for the soil moisture monitoring. The finding can provide a strong reference to further improve the accuracy and stability of the soil moisture monitoring model.