Evaluation of Hyperparameter Combinations of the U-Net Model for Land Cover Classification

Abstract

The aim of this study was to select the optimal deep learning model for land cover classification through hyperparameter adjustment. A U-Net model with encoder and decoder structures was used as the deep learning model, and RapidEye satellite images and a sub-divided land cover map provided by the Ministry of Environment were used as the training dataset and label images, respectively. According to different combinations of hyperparameters, including the size of the input image, the configuration of convolutional layers, the kernel size, and the number of pooling and up-convolutional layers, 90 deep learning models were built, and the model performance was evaluated through the training accuracy and loss, as well as the validation accuracy and loss values. The evaluation results showed that the accuracy was higher with a smaller image size and a smaller kernel size, and was more dependent on the convolutional layer configuration and number of layers than the kernel size. The loss tended to be lower as the convolutional layer composition and number of layers increased, regardless of the image size or kernel size. The deep learning model with the best performance recorded a validation loss of 0.11 with an image size of 64 × 64, a convolutional layer configuration of C→C→C→P, a kernel size of 5 × 5, and five layers. Regarding the classification accuracy of the land cover map constructed using this model, the overall accuracy and kappa coefficient for three study cities showed high agreement at approximately 82.9% and 66.3%, respectively.