Chen Yanli ^* Huang Lu Chen Cheng Xie Ying

• Guangxi Institute of Meteorological Sciences, Nanning, Guangxi Zhuang Region, China

Multi-view real-life images taken by eco-meteorological observation stations can provide high-throughput visible light (RGB) image data for vegetation monitoring, but at present, there are few research reports on the vegetation monitoring effect of multi-view images and its difference from satellite remote sensing monitoring. In this study, with the underlying surface mixed with karst bare rock and vegetation as the research object, the far-view images and near-view images of 4 eco-meteorological stations were used to compare the segmentation effect of machine learning segmentation algorithm on images from far and near views, analyze the vegetation growth characteristics of visible images from far and near views, and investigate the differences between multi-view images and satellite remote sensing monitoring. The results showed that: (1) machine learning algorithm was suitable for green vegetation segmentation of multi-view images. The segmentation accuracy of images from a near view was higher than that from a far view, with an accuracy rate of over 85%. Images captured under weak light conditions could obtain higher vegetation segmentation accuracy, and the proportion of bare rocks had no obvious influence on image segmentation accuracy. (2) The interannual variation trends of vegetation presented by different RGB vegetation indexes varied greatly, and the interannual variation difference of vegetation from a far view was greater than that from a near view. NDYI and RGBVI showed good consistency in vegetation changes from far and near views, and could also better show the interannual differences of vegetation. From the perspective of intra-year variation, various RGB vegetation indexes showed seasonal changes in different degrees. The vegetation in karst areas grew well from April to October, and the RGB vegetation indexes reached the peaks from May to June at most stations. The seasonal distribution of vegetation indexes were more obvious from a far view. (3) There was significant difference in the correlation between ground-based multi-view RGB vegetation indexes and NDVI of different satellites. In general, the correlation with FY3D NDVI was weaker than that with MODIS NDVI. Most RGB vegetation indexes from a far view had a good correlation with MODIS NDVI.