关键词: SRTM DEM, 神经网络, 多光谱影像, 灰色关联分析, 方差膨胀因子

Abstract:

【Objective】SRTM DEM is a publicly available DEM accessible at no cost. However, it is well known that SRTM DEM has a large vertical deviation. In order to improve the accuracy of SRTM DEM in cultivated land, the effects of topography on the temporal and spatial distribution of soil physical and chemical properties and crop growth were analyzed to mine the factors of interaction with topography, so as to obtain a digital elevation model for precision agriculture. 【Method】This paper took Helen Dongxing Agricultural Machinery Cooperative in Heilongjiang Province as the study area, the actual ground elevation data were collected, and SPOT-6, Sentinel-2A remote sensing images and SRTM DEM were obtained. The variables, such as normalized differential moisture index (NDMI), normalized difference vegetation index (NDVI), Tasseled Cap Brightness (TCB) and potential solar radiation (PSR), were extracted, and the effects of topography on them were analyzed. Extreme Learning Machine (ELM) and back Propagation Neural Network (BPNN) were used to improve horizontal spatial resolution and vertical accuracy of SRTM DEM. The accuracy was verified with the actual ground elevation point and compared with the DEM generated by the UAV and the ZY-3. 【Result】The correlation degree between SRTM, NDVI, NDMI, TCB and the improved elevation were more than 90%, which were important factors for improving SRTM DEM. In the whole study area, the RMSE_P of BPNN method was 0.98, the RMSEP of R²_P was 0.98, and the RMSE_P of Elm was 1.00, R²_P was 0.90. In the flat area, the RMSE_P of BPNN method was 0.84 and the RMSE_P of ElM was 1.00. In the fluctuation area, the RMSE_P of BPNN method was 0.99, and the RMSE_P of ELM was 0.94. The vertical accuracy of DEM obtained by this method was higher than that of DEM generated by ZY-3. It provided a new idea for improving the spatial resolution of SRTM. 【Conclusion】 The introduction of auxiliary information of SRTM, NDVI, NDMI and TCB was helpful to improve the spatial resolution and vertical accuracy of SRTM DEM, so as to obtain high accuracy DEM. The accuracy of digital elevation model obtained by BPNN method improved by SRTM DEM was higher than that obtained by ELM method as a whole. In addition, the further research showed that the BPNN method was more suitable for the acquisition of high-precision DEM in the flat area, and the ELM method was more suitable for the acquisition of high-precision DEM in the undulating area. The research results improved the accuracy of the existing DEM and provided data support for accurate farmland management zoning and digital soil mapping.

Key words: SRTM DEM, neural network, multispectral image, grey relational analysis, variance expansion factor