Combining K-fold cross validation with bayesian hyperparameter optimization for accuracy enhancement of land cover and land use classification

Land cover and land use (LCLU) information is crucial in different earth observation applications, such as environmental management, infrastructure planning, and urban development. Recently, deep learning has emerged as an effective technique for image processing and spatial analysis, with excellent accuracies in image classification. However, training deep learning models for LCLU classification using remote sensing data continues to be challenging. One of the main challenges when using deep learning techniques for remote sensing classification is determining the optimized hyperparameters. Bayesian hyperparameter optimization is considered to be one of the most effective methods for determining optimized hyperparameters. Therefore, this method was applied to find the optimal learning rate, gradient clipping threshold and dropout rate for ResNet18 model. Using these optimized hyperparameters to classify EuroSat dataset, the model's overall accuracy was 94.19%. To improve the exploration of the search space for better learning rate, gradient clipping threshold, and dropout rate hyperparameters, Bayesian hyperparameter optimization was combined with K-fold cross validation. The ResNet18 model achieved 96.33% overall accuracy for classifying EuroSat dataset when using the hyperparameters obtained by this process. This improvement in overall accuracy demonstrates the effectiveness of combining Bayesian hyperparameter optimization with K-fold cross validation as an enhanced technique for finding optimized hyperparameters in remote sensing LCLU classification.