SSCBKA-optimized machine learning for hyperspectral classification of maize seeds for agricultural quality control

The high latitude of hyperspectral data poses significant difficulties for data processing. Machine learning has demonstrated unique advantages in handling high-latitude data. However, traditional machine learning is limited in model performance due to unreasonable setting of internal parameter values. Therefore, this paper proposes the use of an improved Black-winged Kite algorithm to optimize the parameters of a machine learning algorithm for processing hyperspectral data. First, maize seeds were collected from the test site, and hyperspectral data of the maize seeds were obtained using hyperspectral instruments. Secondly, to enhance the global search ability of the Black-winged kite algorithm, this paper proposes a new improved mechanism called nonlinear time-varying perturbation sine and cosine strategy, which incorporates this strategy into the migration behavior of the black-winged kite algorithm. At the same time, to enhance the population diversity of the black-winged kite algorithm, the vertical and horizontal crossover strategy is introduced into the position update of the algorithm. Again, to verify the effectiveness of the proposed algorithm, the proposed algorithm was compared with seven meta-heuristic optimization algorithms and ten newer meta-heuristic optimization algorithm variants in dimensions of 10, 30, 50, and 100 based on the CEC2017 test function. Finally, the improved algorithm was applied to hyperparameter optimize the extreme random tree and random forest respectively and compared with the other nine models based on accuracy, precision, recall, and F1-Scorse. Experimental results show that the algorithm proposed in this paper converges the fastest and optimizes the best compared with other algorithms. The accuracy of the improved extreme random tree has been 88.268%, the precision has been 94.805%, the recall has been 81.100%, and the F1-Scorse has been 87.425%. The improved random forest achieved 89.385% accuracy, 88.172% precision, 91.100% recall, and 89.617% F1-Scorse.