Training and test sets were prepare

Training and test sets were prepared for each different neighbourhood size separately. Consequently, eight sets of training and test data were prepared for training the convolutional model. For each set, we consider a 90–10% ratio for dividing the input dastaset into train and test sets. The test set was not involved in the process of model training. This set was used to assess the classification performance of the trained model. The training set further were subdivided into train and validation set with a 80–20% ratio. Each training set covers a total of 4.50 million random selected samples (1.50 million training samples for each class type).

2.2.3. Validation during Model Design

A validation set is also required in this process which helps to assess the generalisation ability of the trained model on validation dastaset. If the validation performance is deteriorating while the training performance is improving over the epochs, the model assumes to be not well trained and has a very low generalisation ability. In deep learning terms it is called overfitting [44]. The training stage ended after 2000 iterations and finally the trained network was applied on test samples to generate classification outputs.