Train, Validation and Test Sample P

Train, Validation and Test Sample Patches Generation for Model Design

The Last step in stage-1 is to generate model’s input data of pixel-patches that were used to prepare train, validation and test sets and also assign the appropriate label to each sample. These sample-patches are utilised by the F-CNNs model during training process to obtain the underlying pattern in the training data [44]. This experiment considers spectral information from neighbouring pixels surrounded by the target-pixels into the classification process and to incorporate contextual information, so we chose small patches of pixels from the input images to generate the train and test sample sets. The class label of the pixel located at the centre of a patch was used as the class type for that particular patch. Pixels surrounding the centre pixel were used to determine the neighbouring area for information extraction. We generated eight different sets with varying neighbourhood sizes and preferred to apply odd numbers for the neighbourhood size to obtain a specific centre pixel location. Each training sample patch is generated based on a specific neighbourhood (N) size. Figure 4 displays the structure of a training sample extracted from Landsat data with N-11 size. Samples with N-1 size actually refers to single pixels used as samples. Model training with pixel-wise samples helped us to evaluate context-based classification performance with conventional pixel-wise classification process.