Remote Sensing Monitoring And Ecosystem Health Evaluation Methods Research Of Tropical And Subtropical Nature Reserves Based On High-resolution Images

Tropical and Subtropical nature reserves are of vital importance for maintaining biodiversity,conserving water and soil,regulating climate and other aspects.The internal surface coverage of the nature reserve is relatively special.Large tropical forests are tall and dense,with multiple levels from canopy to understory.Beyond that,the texture and color of many small or artificially planted trees are quite different.In addition,woodland and water areas occupy most of the total area,while artificial land and farmland account for very small proportions.According to the characteristics of nature reserves,a remote sensing monitoring and ecosystem health evaluation framework based on high-resolution images is constructed,in which three methods are proposed:(1)Residual Multi-Attention Network(Res MANet),which improved the representation ability of the original residual network with multi-scale convolution and attention modules and reduced the impact of sample imbalance on the results by applying a two-stage training strategy based on joint loss function.Compared with other classic semantic segmentation networks on true color and multispectral validation dataset,the mean intersection over union(MIo U)of our model were improved 2.39%-5.27% and 3.25%-7.38% and got an improvement of 3.71%-10.38% on public dataset;(2)Improved U-Net change detection model integrated short connection and pyramid pooling modules based on U-Net to fuse the deep and shallow information and extract more global context,so that it can directly detect the types of land cover changes.The improved model and the application of the joint loss function increased the overall accuracy and Kappa coefficient.The end-to-end model reduced manual intervention and alleviated the impact of sample imbalance on change detection results.Compared with U-Net on the Hainan and public datasets,the overall accuracy has increased by1.5% and 0.78%;(3)Ecological health evaluation method based on the Pressure-State-Response(PSR)framework,which uses remote sensing data as the main input and multi-source data as supplementation and determined different factors and weights for forest and wetland nature reserves.Obtain the ecosystem health evaluation value and health level of the nature reserve through operations such as min-max normalization and weighted stacking.The national and provincial nature reserves in Hainan Island were selected as the evaluation cases of three methods.The evaluation results showed that the forest or water is dominant in the land coverage of the nature reserves,while a few agricultural land and artificial buildings appeared in partial area.The proportion of unchanged categories in most nature reserves is generally more than 98%,while some of the reserves have new bare or artificial surface.The overall excellent and good rate of ecosystem health was 81.48%,the average health value of forest type reserve was 78.25,and the average health value of wetland type reserve was 68.11.