Advanced machine learning algorithms for Canadian wetland mapping using polarimetric synthetic aperture radar (PolSAR) and optical imagery

• Main Author: Mahdianpari, Masoud
• Format: Thesis
• Language: English
• Published: Memorial University of Newfoundland 2019
• Subjects: Newfoundland; Sea ice
• Online Access: https://research.library.mun.ca/13915/; https://research.library.mun.ca/13915/1/thesis.pdf

Wetlands are complex land cover ecosystems that represent a wide range of biophysical conditions. They are one of the most productive ecosystems and provide several important environmental functionalities. As such, wetland mapping and monitoring using cost- and time-efficient approaches are of great interest for sustainable management and resource assessment. In this regard, satellite remote sensing data are greatly beneficial, as they capture a synoptic and multi-temporal view of landscapes. The ability to extract useful information from satellite imagery greatly affects the accuracy and reliability of the final products. This is of particular concern for mapping complex land cover ecosystems, such as wetlands, where complex, heterogeneous, and fragmented landscape results in similar backscatter/spectral signatures of land cover classes in satellite images. Accordingly, the overarching purpose of this thesis is to contribute to existing methodologies of wetland classification by proposing and developing several new techniques based on advanced remote sensing tools and optical and Synthetic Aperture Radar (SAR) imagery. Specifically, the importance of employing an efficient speckle reduction method for polarimetric SAR (PolSAR) image processing is discussed and a new speckle reduction technique is proposed. Two novel techniques are also introduced for improving the accuracy of wetland classification. In particular, a new hierarchical classification algorithm using multi-frequency SAR data is proposed that discriminates wetland classes in three steps depending on their complexity and similarity. The experimental results reveal that the proposed method is advantageous for mapping complex land cover ecosystems compared to single stream classification approaches, which have been extensively used in the literature. Furthermore, a new feature weighting approach is proposed based on the statistical and physical characteristics of PolSAR data to improve the discrimination capability of input features prior to incorporating them into the classification scheme. This study also demonstrates the transferability of existing classification algorithms, which have been developed based on RADARSAT-2 imagery, to compact polarimetry SAR data that will be collected by the upcoming RADARSAT Constellation Mission (RCM). The capability of several well-known deep Convolutional Neural Network (CNN) architectures currently employed in computer vision is first introduced in this thesis for classification of wetland complexes using multispectral remote sensing data. Finally, this research results in the first provincial-scale wetland inventory maps of Newfoundland and Labrador using the Google Earth Engine (GEE) cloud computing resources and open access Earth Observation (EO) collected by the Copernicus Sentinel missions. Overall, the methodologies proposed in this thesis address fundamental limitations/challenges of wetland mapping using remote sensing data, which have been ignored in the literature. These challenges include the backscattering/spectrally similar signature of wetland classes, insufficient classification accuracy of wetland classes, and limitations of wetland mapping on large scales. In addition to the capabilities of the proposed methods for mapping wetland complexes, the use of these developed techniques for classifying other complex land cover types beyond wetlands, such as sea ice and crop ecosystems, offers a potential avenue for further research.