Developing a deep learning network suitable for automated classification of heterogeneous land covers in high spatial resolution imagery
University of New Brunswick
The incorporation of spatial and spectral information within multispectral satellite images is the key for accurate land cover mapping, specifically for discrimination of heterogeneous land covers. Traditional methods only use basic features, either spatial features (e.g. edges or gradients) or spectral features (e.g. mean value of Digital Numbers or Normalized Difference Vegetation Index (NDVI)) for land cover classification. These features are called low level features and are generated manually (through so-called feature engineering). Since feature engineering is manual, the design of proper features is time-consuming, only low-level features in the information hierarchy can usually be extracted, and the feature extraction is application-based (i.e., different applications need to extract different features). In contrast to traditional land-cover classification methods, Deep Learning (DL), adapting the artificial neural network (ANN) into a deep structure, can automatically generate the necessary high-level features for improving classification without being limited to low-level features. The higher-level features (e.g. complex shapes and textures) can be generated by combining low-level features through different level of processing. However, despite recent advances of DL for various computer vision tasks, especially for convolutional neural networks (CNNs) models, the potential of using DL for land-cover classification of multispectral remote sensing (RS) images have not yet been thoroughly explored. The main reason is that a DL network needs to be trained using a huge number of images from a large scale of datasets. Such training datasets are not usually available in RS. The only few available training datasets are either for object detection in an urban area, or for scene labeling. In addition, the available datasets are mostly used for land-cover classification based on spatial features. Therefore, the incorporation of the spectral and spatial features has not been studied comprehensively yet. This PhD research aims to mitigate challenges in using DL for RS land cover mapping/object detection by (1) decreasing the dependency of DL to the large training datasets, (2) adapting and improving the efficiency and accuracy of deep CNNs for heterogeneous classification, (3) incorporating all of the spectral bands in satellite multispectral images into the processing, and (4) designing a specific CNN network that can be used for a faster and more accurate detection of heterogeneous land covers with fewer amount of training datasets. The new developments are evaluated in two case studies, i.e. wetland detection and tree species detection, where high resolution multispectral satellite images are used. Such land-cover classifications are considered as challenging tasks in the literature. The results show that our new solution works reliably under a wide variety of conditions. Furthermore, we are releasing the two large-scale wetland and tree species detection datasets to the public in order to facilitate future research, and to compare with other methods.