Refining KNN Classification of Hyperspectral Images through PCA and Hyperparameter Optimization Techniques

Recognition hyperspectral images can be detected for the analysis of objects to provide scientific reference value, how to use high spatial resolution hyperspectral data to identify the image and classification research accurately is the current more mature research results, but the parameters need to be further optimized is one of the focus of the current vigorous research. The article takes the classical three public datasets, Indian Pines, Botswana, and Kennedy Space Center, as examples, and uses PCA to perform initial dimensionality reduction on the datasets. In tuning the algorithm’s hyperparameters, the hyperparameters are computed separately for different proportions of the training set using k-fold cross-validation. The recognition ability of the hyperspectral image dataset is explored using the K-nearest neighbor. The results show that: a) Compared with the setting of random parameters, the accuracy of calculating the optimal hyperparameters using k-fold cross-validation is more accurate, and its Overall Accuracy (OA), Average Accuracy (AA), and Kappa Coefficient values can be improved between 2% and 5%; b) The amount of data in the training set can directly affect the classification accuracy, with the highest scores under the highest provisioned 90% training set. Overall, the proportion of the training set, hyperparameter optimization, and classification algorithm all have different degrees of influence on hyperspectral image recognition accuracy. In the process of hyperspectral image recognition, prioritizing removing uncontrollable noise, optimizing the hyperparameters of machine learning algorithms, and utilizing a large amount of data will improve the recognition accuracy of traditional machine learning on hyperspectral images.