摘要:
遥感和深度学习方法结合应用于自然资源的监测、评估是一种高效的方法。该研究综合考虑珠海一号高光谱影像的特性,在HRNet网络中引入3D卷积模块提出3D-HRNet网络用于自然资源调查监测的语义分割模型,并以遥感影像计算生物丰富度指数、植被覆盖指数、水网密度指数、土地应力指数、污染负荷指数和环境约束指数构建生态指数(ecological index,EI)生态评价模型,对北京市北部部分区域进行自然资源监测和评估。结果表明: ①3D-HRNet模型提取自然资源的平均总体精度为0.83,F1分数为0.83,Kappa系数为0.73,比HRNet模型分别高0.04,0.04和0.06,比3D-CNN模型分别高0.04,0.05和0.06,说明3D-HRNet模型提取高光谱影像的自然资源结果比HRNet模型更好,即3D卷积模块能更好地利用高光谱间特性提取信息; ②利用EI生态评价模型对北京市北部部分区域2020年生态环境进行评价,其EI平均值为68.2,反映了区域内生态状况良好,与北京市生态环境状况公报结论接近,说明遥感用于EI生态评价的可行性,为区域生态状况的时空分析提出创新性方法。
Abstract:
The combination of remote sensing and deep learning is efficient in the monitoring and evaluation of natural resources. Based on the comprehensive consideration of the characteristics of Zhuhai-1 OHS hyperspectral images, this study established the 3D-HRNet architecture by introducing the 3D convolutional module into the HRNet architecture and applied it to the semantic segmentation model for natural resources survey and monitoring. Using remote sensing images, this study established an ecological index (EI) evaluation model by calculating the species richness index, vegetation index, water network density index, land stress index, pollution load index, and environmental regulation index. Then, the model was employed to monitor and evaluate the natural resources in partial areas in the northern part of Beijing. The results show that: ① when being used to extract the natural resources, the 3D-HRNet model yielded average overall accuracy, a F1 score, and a Kappa coefficient of 0.83, 0.83, and 0.73, respectively, which were 0.04, 0.04, and 0.06 higher than those of the HRNet model, and 0.04, 0.05 and 0.06 higher than those of 3D-CNN model, respectively. This suggests that the 3D-HRNet model can extract natural resources from hyperspectral images more effectively than the HRNet model. In other words, the 3D convolutional module can utilize the inter-hyperspectral features to extract information more effectively; ② The eco-environment of partial areas in north Beijing in 2020 was evaluated using the EI evaluation model, with an average EI value of 68.2. This reflects a good ecological state in the study area, highly consistent with the conclusion of the Report on the State of the Ecology and Environment in Beijing, demonstrating the feasibility of remote sensing for ecological assessment. Therefore, this study provides an innovative method for the spatio-temporal analysis of the regional ecological state.
下载: