The amount of training samples used in hyperspectral image classification can vary depending on the specific research or application, and there is no fixed rule that restricts it to only 5% of the available samples.
The number of training samples required for hyperspectral image classification depends on various factors, including the complexity of the classification task, the dimensionality of the hyperspectral data, the number of classes to be classified, and the available computational resources. In some cases, researchers may have limited access to labeled training samples due to the cost or difficulty of obtaining ground truth data for hyperspectral images. In such situations, they may resort to using a smaller subset of the available samples for training.
However, it is important to note that using a smaller number of training samples can potentially limit the performance and generalization capabilities of the classification model. In general, having a larger number of diverse and representative training samples tends to lead to better classification results. Researchers strive to find a balance between the number of training samples and the computational resources available to achieve the desired classification accuracy.
It's worth mentioning that the field of hyperspectral image classification is constantly evolving, and different studies may adopt different approaches based on their specific objectives and constraints. Therefore, it is not accurate to claim that a maximum of work in this field uses only 5% of training samples.
