Using different training strategies for urban land-use classification based on convolutional neural networks

Urban land-use scene classification from high-resolution remote-sensing imagery at high quality and accuracy is of paramount interest for urban planning, government policy-making and urban change detection. In recent years, urban land-use classification has become an ongoing task in areas addressable primarily by remote sensing, and numerous deep learning algorithms have achieved high performance on this task. However, both dataset and methodology problems still exist in the current approaches. Previous studies have relied on limited data sources, resulting in saturated classification results, and they have difficulty achieving comprehensive classification results. The previous methods based on convolutional neural networks (CNNs) focused primarily on model architecture rather than on the hyperparameters. Therefore, to achieve more accurate classification results, in this study, we constructed a new large dataset for urban land-use scene classification. More than thirty thousand remote sensing scene images were collected to create a dataset with balanced class samples that includes both higher intra-class variations and smaller inter-class dissimilarities than do the previously available public datasets. Then, we analysed two possible strategies for exploiting the capabilities of three existing popular CNNs on our datasets: full training and fine tuning. For each strategy, three types of learning rate decay were applied: fixed, exponential and polynomial. The experimental results indicate that fine tuning tends to be the best-performing strategy, and using ResNet-V1-50 and polynomial learning rate decay achieves the best results for the urban land-use scene classification task.