Land cover classification through Convolutional Neur-al Network model assembly: A case study of a local rural area in Thailand

Abstract

Recent Convolutional Neural Network (CNN) has shown great potential in image classification, segmentation and object detection. Land cover takes advantage of CNN development for a large type of applications as water management and urban growing. However, to perform a land cover with numerous features – classes in classical CNN terminology, CNN models require a significant number of layers and neurons, resulting in high computational costs. To address this problem, a methodology is proposed in this paper to build a land cover using the aggregation of several CNN models. The overall process is based on 7 steps. The first two steps are the dataset creation and arrangement in smaller dataset fit for the specific features to detect. Then, a CNN architecture is built and validated on each sub-dataset corresponding to each class. Post-processing is conducted on each prediction before assembling the results. In the last step, a data cleaning is performed, giving the final land cover. The land cover of a rural area in Thailand is performed as a demonstration of the method, using satellite images with a resolution of 0.15 m/pixel. A 5-class (buildings, crops, forests, roads, and wastelands) dataset is created, consisting of a total of 1 million tiles of 64 × 64 pixels. The prediction results using the developed CNN model show an accuracy greater than 90% for each class, except for the road class where the accuracy only reaches 72%. Post-processing is performed on each of the 5 predictions. Only the 4 best results are retained and assembled to obtain the land cover, which generally corresponds to buildings, crops, forests, and wastelands. This method enables to identify by substitution with improved accuracy the last class whose prediction is the least accurate, and which generally corresponds to roads due to their small width relative to the tile size. The proposed methodology to perform a land cover by aggregating the prediction of different CNN models is found to predict correctly the land cover of two areas, especially roads can be classified, demonstrating the usefulness of the approach.