| dc.description.abstract | Remote sensing (RS), a critical technology for large-scale-monitoring Earth-observing systems (EOS), plays an important role in Earth science and other related fields where physical, biological, and chemical properties of the Earth can be characterized in a non-contact manner. One of the most widely used techniques to analyze land-cover information utilizing RS images is pixel-level classification, where each individual pixel will be classified with a semantic label by employing machine learning (ML) algorithms. In recent decades, with the tremendous developments in hardware and software, computational capabilities have improved dramatically, which has facilitated the development of deep-learning (DL) algorithms derived from traditional ML. Among various deep learning models, recurrent neural networks (RNNs), which are able to process sequential inputs by utilizing a series of internal state, attracts more attention for the purpose of handling multi-temporal RS image analysis due to their intrinsic recurrent structure. However, their application on single-image classification still needs more investigations, especially from the perspective of sequential feature extraction. To address this limitation, we propose a similarity measurements-based sequential feature extraction method for singe RS image classification using long short-term memory (LSTM), a special class of RNN. For a given pixel the proposed framework utilizes the spectral information of those pixels collected from the whole image instead of the individual spectral vector of its own. And its classification performance on two standard datasets demonstrates its effectiveness compared with other benchmark algorithms.
However, the computational time cost of the aforementioned approach is a critical issue as all pixels in that single RS image need to be considered during similarity measurement for every pixel. That brings more difficulties, especially for processing large-scale RS image. Therefore, building upon the previous work, we improve that model by adding segmentation map as an additional criterion for shrinking searching range from the whole image to selected segments. Within such a segmentation map, homogeneous pixels will be aggregated into adjacent segments. Thus, the similarity measurement will be split into two phases, including segment-level similarity calculation and pixel-level similarity calculation. Experimental results obtained from three benchmark hyperspectral RS images and large-scale satellite images illustrate that the proposed approaches achieve promising classification performance with lower computational time cost.
Besides classification, another application of deep learning is object detection where interest of object is allowed to identify and localized in an image or video. In this study, our focus is to detect fallen trees from large-scale aerial photos. Therefore, we develop a framework to automatically create image and annotation dataset which can be utilized as input data for deep learning based object detection model. Quantitative and qualitative classification results illustrate that, for isolated fallen trees, our model achieves promising results considering recall index. However, many false positive detection results are generated as well. Then we further investigate the influence of those fallen trees located on the boundary of extracted sub-image and find that using geographical coordinates to calculate accuracy metrics can achieve better results. Furthermore, we also explore the issue of overlapping fallen trees and observe that overlapping fallen tree introduce more detection errors which can be alleviated to some extend by considering fallen tree clusters. However, detecting overlapping fallen trees is still a challenging task, especially for those large scale datasets collected from the real world. | |
