Grain moisture content at the filling stage is one of the most important indicators during maize breeding. However, the destructive sample (such as threshing) has been widely used to measure the moisture content of 200 grains in the middle of the ear. This study aims to address the challenge of threshing in moisture detection. A general model was proposed for the quantitative analysis and in-situ detection of maize grain moisture at the grain-filling stage using near-infrared spectroscopy. The GA-IRIV-DS strategy was established for spectral data processing. The genetic algorithm (GA) and the iteratively retained information variable (IRIV) secondary wavelength screening were combined to extract the effective moisture variables from the spectral data. The direct standardization algorithm (DS) was combined to correct the spectral data of maize spike tip grains at the filling stage. The spectrum of the middle 200 grains was also used as the baseline. The moisture quantitative analysis model was established on the two test samples of the middle 200 grains and the ear tip grains. As such, the general model was established for quantitative analysis of maize grain moisture at the filling stage. After the GA + IRIVN wavelength screening, the effective variables were extracted to reduce the feature space dimension for high prediction accuracy. In the parameter setting of the GA algorithm, the individual fitness increased with the increase of crossover rate and mutation rate. The highest value was achieved when the crossover rate was 0.5 and the mutation rate was 0.2. The variable threshold of 0.5 was set to retain the important features without too many redundance, in order to enhance the prediction accuracy of the improved model. In the IRIV algorithm, the residual variance was no longer significantly reduced, after the number of principal components increased to 10. The 5-fold cross-validation of the variables was provided for a better balance between the calculation cost and the evaluation accuracy. At the same time, the algorithm performed 10 iterations. As such, more effective moisture variables were extracted for the high accuracy of the improved model. Then the conversion matrix of DS was used to reduce the difference between the prediction and the modeling sample. Moreover, the spectral difference was reduced by 59.4 % after correction. The GA + IRIVN with the wavelength screening was combined to extract the spectral features, in order to further verify the effectiveness of the GA-IRIV-DS spectral data processing. The moisture quantitative model was constructed using partial least squares (PLSR). The DS was also combined with the full spectrum and various typical wavelength screenings. The test results show that the GA-IRIVN-DS-PLSR model on the two-sample prediction sets performed better than the full spectrum than the rest. The coefficient of determination (R²) reached 0.971 5 and 0.901 2, respectively, for the prediction of the middle and spike tip grain samples. The root mean square error of prediction (RMSEP) decreased by 80.10 % and 64.60 %, respectively, compared with the full spectrum. The generalization was improved in the moisture quantitative model with the near-infrared spectral using GA-IRIVN-DS spectral data processing. A feasible reference can also be provided to reduce the sample damage for the high efficiency of detection in the process of maize breeding.

In maize breeding, limitations on sampling quantity and associated costs for measuring maize grain moisture during filling are imposed by factors like the planting area of new varieties, maize plant density, effective experimental spikes, and other conditions. However, the conventional method of detecting moisture content in maize grains is slow, damages seeds, and necessitates many sample sets, particularly for high moisture content determination. Thus, a strong demand exists for a non-destructive quantitative analysis model of maize moisture content using a small sample set during grain filling. The Bayes-Merged-Bootstrap (BMB) sample optimization method, which built upon the Bayes-Bootstrap sampling method and the concept of merging, was proposed. A critical concern in dealing with small samples is the relationship between data distribution, minimum sample value, and sample size, which has been thoroughly analyzed. Compared to the Bayes-Bootstrap sample selection method, the BMB method offers distinct advantages in the optimized selection of small samples for non-destructive detection. The quantitative analysis model for maize grain moisture content was established based on the support vector machine regression. Results demonstrate that when the optimal resampling size is 1000 times or more than the original sample size using the BMB method, the model exhibits strong predictive capabilities, with a determination coefficient (R²)>0.989 and a relative prediction determination (RPD)>2.47. The results of the 3 varieties experiment demonstrate the generality of the model. Therefore, it can be applied effectively in practical maize breeding and determining grain moisture content during maize machine harvesting.