Deep learning for land use and land cover in aerial images

Abstract

This thesis focuses on two main types of research: classification and segmentation, addressing the challenge of aerial images using deep learning techniques.

Chapter 1 provides a brief general introduction to deep learning for land use and land cover in aerial images, followed by the research questions. Additionally, the objectives of the dissertation and its contributions are described.

In Chapter 2, a convolutional neural network (CNN) method is proposed to classify land use and land cover of five economic crops: rice, sugarcane, cassava, rubber, and longan, from the aerial images. Also, the ensemble learning method is proposed to enhance the performance of the land use classification. The work reported in this thesis used eight CNN architectures to create robust models and classify the aerial images for the classification tasks. Hence, three data augmentation techniques (rotation, height shift, and width shift) are combined when training the CNN models. Moreover, the ensemble CNN model is proposed to enhance the performance of the economic crop classification model.

Chapter 3 propose the snapshot ensemble CNN to improve the performance of the land use classification from the aerial images. The work reported in this thesis experimented with the snapshot ensemble CNN method using various learning schedules. The new drop cyclic cosine learning rate schedule, called dropCyclic, is proposed and compared with two existing learning rate schedules. The proposed learning rate schedule is evaluated on three datasets: UCM, AID, and EcoCropsAID. The results showed that the proposed dripCyclic outperformed the existing learning rate schedules on the UCM dataset. As a result, the ensemble CNN obtains better performance than using only the single CNN.

Chapter 4 proposes the instance segmentation technique to segment the water body from the aerial. The mask region-based CNN (mask R-CNN) is the instance segmentation technique to find the water resource areas for the segmentation task in this thesis. In the experiments, the mask R-CNN model could segment water bodies efficiently. Furthermore, the data augmentation techniques are included in the training process. The experimental results showed that the mask R-CNN method combined with data augmentation techniques when training obtained two times better performance than without combining data augmentation techniques.

Chapter 5 comprises two main sections: 1) Answers to the research questions 2) future work. This chapter briefly explains the proposed approaches and answers three main research questions in land use and land cover in aerial images using deep learning techniques. Two main approaches are planned to be the focused of future work, as follows. For the classification technique, I plan to replace the unweighted average method with the cost-sensitive probability method in the snapshot ensemble CNN method. For the segmentation technique, I will consider applying the new deep learning methods to enhance the performance of the water resource segmentation and other tasks.

This research makes a significant contribution in classification and segmentation for land use and land cover through deep learning-based innovations and has great potential utility in a wide range of aerial images for geographic information systems and remote sensing.

-