Comparison of multi-source satellite images for classifying marsh vegetation using DeepLabV3 Plus deep learning algorithm

• Published in: Ecological indicators Vol. 125; p. 107562
• Main Authors: Liu, Man, Fu, Bolin, Xie, Shuyu, He, Hongchang, Lan, Feiwu, Li, Yuyang, Lou, Peiqing, Fan, Donglin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2021; Elsevier
• Subjects: DeepLabV3 Plus; Intelligent classification; Marsh vegetation; Satellite images; Spatial resolution; Spectral range; Marsh vegetation; DeepLabV3 Plus; Intelligent classification; Satellite images; Spectral range; Spatial resolution
• ISSN: 1470-160X; 1872-7034
• DOI: 10.1016/j.ecolind.2021.107562

•DeepLabV3 Plus algorithm has a strong ability to classify marsh vegetation.•There are significant differences in classification results between different spatial resolutions.•The fusion of Gaofen-2 and Sentinel-2A images improves the overall accuracy.•Synergistic use of spectral band and index achieved 0.928 overall accuracy.•Multi-source data achieve 16%–64% improvement in classification accuracy of marsh vegetation. The accurate classification of wetland vegetation is essential for rapid assessment and management. The Honghe National Nature Reserve (HNNR), located in Northeast China, was studied. The multi-scale remote sensing data of a new generation of Chinese high-spatial-resolution earth observation satellites Gaofen-1 (GF-1), Gaofen-2 (GF-2), Ziyuan-3 (ZY-3), and international earth observation satellites Sentinel-2A and Landsat 8 OLI were selected as sources. Based on the DeepLabV3 Plus deep learning model, 12 intelligent marsh vegetation classification models were constructed. We quantitatively analyzed the applicability and identification ability of DeepLabV3 Plus for classifying complex marsh vegetation. We discuss the differences in accuracy of marsh vegetation classification with different remote sensing data sets. The spatial resolution of remote sensing data sets ranges from 30 m to 0.8 m, and spectral bands range from blue bands (450 nm) to shortwave infrared bands (2280 nm). The specific conclusions of this study are as follows: (1) The DeepLabV3 Plus model better identified marsh vegetation, but there were significant differences in the classification accuracy of 12 DeepLabV3 Plus intelligent identification models. (2) Under the same conditions of the spectral bands (four Blue ~ NIR bands), the accuracy of deep-water marsh vegetation classification gradually increased as spatial resolution improved. For shallow-water marsh vegetation, when the accuracy of vegetation classification increased to a certain level, the classification accuracy decreased with the improvement of spatial resolution, which indicated that high-resolution images reduced pixel mixing to a certain extent, but for some vegetation types, the internal spectral difference increased, which made classification more difficult. (3) The increase of spectral bands improved the classification of marsh vegetation, while the classification accuracy of models with spectral indices was better than that of models only including spectral bands. (4) The accuracy of marsh vegetation classification was greatly improved by combining spectral bands and spectral indices. (5) The classification of the five sensor satellite images had statistical differences between models with different spatial resolutions and models with different spectral ranges.