Study on transfer learning ability for classifying marsh vegetation with multi-sensor images using DeepLabV3+ and HRNet deep learning algorithms

• Published in: International journal of applied earth observation and geoinformation Vol. 103; p. 102531
• Main Authors: Liu, Man, Fu, Bolin, Fan, Donglin, Zuo, Pingping, Xie, Shuyu, He, Hongchang, Liu, Lilong, Huang, Liangke, Gao, Ertao, Zhao, Min
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2021; Elsevier
• Subjects: DeepLabV3+ algorithm; HRNet algorithm; Marsh vegetation classification; Multi-scale segmentation; Satellite images; Transfer learning; Multi-scale segmentation; HRNet algorithm; DeepLabV3+ algorithm; Satellite images; Transfer learning; Marsh vegetation classification
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2021.102531

•DeepLabV3+ and HRNet models have better classification ability in combination images with low-spatial resolution.•Classification ability in marsh vegetation with narrow spectrum is better than that larger spectral range.•DeepLabV3+ and HRNet models have good transfer learning abilities for classifying marsh vegetation.•Transfer learning ability is better in images with different spatial resolutions in similar spectral ranges.•Integration of object-based segmentation and pixel-based classifications improves accuracy by 4.45–5%. The verification of the transfer learning ability of the convolutional neural network in the classification of natural vegetation is relatively lacking. In this paper, 16 combination scenarios of multispectral images in marsh vegetation were constructed. The influence of the combination with different spatial resolution gradients and spectral dimensions on the classification accuracy of marsh vegetation was systematically studied. Multi-sensor images were used to evaluate the transfer learning ability of DeepLabV3+ and HRNet algorithms in marsh vegetation, and analyse the transfer learning effect of two algorithms in different spatial resolution and spectral ranges. The majority voting method was used to fuse the classifications of high, medium and low spatial resolution images. Based on the largest area method, the fusion results were integrated with multi-scale segmentation to explore the classification ability of the integration of pixel-based classification and object-based classification. The average accuracies of different spatial resolutions to vegetation in multispectral images were statistically analysed in order to quantitatively study the classification ability of spatial resolution to marsh vegetation. The results indicated that: (1) image combination improved the classification accuracy of marsh vegetation in low-resolution images, and decreased the classification accuracy of vegetation in high- and medium-resolution images based on DeepLabV3+ and HRNet algorithms; (2) when GF-1 and Sentinel-2A images were used for combination, the spectral range increased by 1565–1655 nm and 2100–2280 nm from 450–900 nm, the classification accuracy of GF-1 image improved by 0.93–1.77%, and the classification accuracy of Sentinel-2A image decreased by 2.34–4.15%; (3) DeepLabV3+ and HRNet algorithms both have good transfer learning capabilities in the classification of marsh vegetation, but the transfer learning ability was better in images with different spatial resolutions in similar spectral ranges than in images with different spectral ranges; (4) the integration of object-based segmentation and pixel-based classifications (DeepLabV3+ and HRNet algorithms) improved the accuracy, and the growth rate of overall accuracy reached 4.45–5%; (5) the classification of water in images with different spatial resolution gradients had the largest difference, and the accuracy of deep-water marsh vegetation was lower than that of shrub and shallow-water marsh vegetation.