Joint Classification of Hyperspectral and Multispectral Images for Mapping Coastal Wetlands

Abstract

It is significant for restoration and protection of natural resources and ecological services in coastal wetlands to map different land cover types with satellite remote sensing data. Considering difficulties of wetland species classification, hyperspectral images (HSIs) with high spectral resolution and multispectral images (MSI) with high spatial resolution are considered to achieve complementary advantages of multisource data. An effective approach, named as multistream convolutional neural network, is proposed to achieve fine classification of coastal wetlands. First, regression processing is adopted to make chaotically scattered coastal wetland data more compact and different. Second, through appropriate feature extraction and feature fusion strategies, high-level information of multisource data in regression domain is fused to distinguish different land cover. Experiments on GF-5 HSIs and Sentinel-2 MSIs are carried out in order to validate the classification performance of the proposed approach in two coastal wetlands of research value in China, i.e., Yellow River Estuary and Yancheng coastal wetland. Experimental results demonstrate the effectiveness of the proposed method compared with the state-of-the-art methods in the field, especially when the number of sample size is extremely small.

Keywords

• Coastal wetlands
• convolutional neural network (CNN)
• data fusion
• hyperspectral imagery (HSI)
• least squares regression (LSR)
• multispectral imagery (MSI)