Abstract:

Rapid and non-destructive monitoring of the water status of maize （Zea mays） is important for water status diagnosis and irrigation management. Spectral indices serve as crucial tools for non-destructive real-time estimation of crop leaf water content （LWC）. However， traditional spectral indices are sensitive to external environmental factors， resulting in reduced prediction accuracy when they are used to estimate LWC. Machine learning （ML） algorithms demonstrate distinct advantages in predicting crop water status， particularly when applied in precision agriculture and crop water status monitoring. Therefore the aims of this study were to enhance the accuracy of LWC estimation by integrating spectral indices with ML approaches， with an overall goal to facilitate efficient water resource utilization during maize cultivation. Field experiments with varying water gradients were conducted in typical maize cultivation regions of Inner Mongolia during 2023-2024. The hyperspectral reflectance of maize leaves were measured across three critical growth stages， and then correlation analyses were conducted between maize LWC and 13 water-sensitive spectral indices. To develop LWC estimation models， spectral features selected via the ReliefF technique were used as input variables for three ML algorithms-partial least squares regression （PLSR）， random forest （RF）， and Gaussian process regression （GPR）. The results demonstrate that among the 13 hydrological indices， the modified DATT index exhibited optimal predictive performance （coefficient of determination R2=0.52）. However， its accuracy was affected by the growth stage and leaf canopy position， limiting its effectiveness for LWC monitoring. Integrating full-spectrum data （350-2500 nm） with ReliefF-selected spectral indices into ML algorithms enhanced the accuracy of LWC estimates by 7%-45%. Models utilizing spectral indices as input features demonstrated superior overall performance， with the RF and GPR models explaining 88%-89% of LWC variability. Independent validations confirmed the robustness of the models， with coefficient of determination R2 values of 0.89 （RF） and 0.88 （GPR） and root mean square error values of 1.95% and 2.04%. Our results show that the synergistic combination of spectral indices with RF/GPR algorithms had cascading effects， significantly improving the accuracy of LWC estimation. This methodology provides a reliable approach for monitoring maize water status and establishes a scientific foundation for the development of precise integrated water-fertilizer management systems.

Key words: maize leaves, water content, spectral index, machine learning