关键词: 高分六号卫星, 红边波段, 蓄积量遥感反演, 随机森林模型, 西宁市

Abstract: The red band and the near-infrared band are often used as sensitive bands for remote sensing retrieval of forest stock volume, however the red edge band between the two is often ignored. To investigate whether the red edge band is sensitive to the accuracy of forest stock volume retrieval, the wide imagery data of GF-6 remote sensing satellite (GF-6 WFV), the DEM, the second survey data of forest resources in 2014 of Xining city were used, and the multiple linear regression model (MLR), the random forest regression model (RF) were employed. The predicting variables were collected from spectral characteristic, vegetation index, topographic factor and image texture. These variables were divided into two groups, the one is no red edge band (No Red-Edge), and the other is red edge band added (Red-Edge Added). The results show that: (1) the selected principal components which were derived from the texture variables of two groups by PCA method mainly explained the texture features of red, near-infrared and red-edge 1 band of the image. (2) Regarding the spectral variables, the surface reflectance of red and near-infrared band were selected because of the high correlation with the forest stock volume in No Red-Edge group, and in Red-Edge Added group, the surface reflectance of the red-edge 1 band was selected. Regarding the vegetation index variables, the Normalized Difference Vegetation Index (NDVI) and Soil Adjusted Vegetation Index (SAVI) were selected in No Red-Edge group, and in Red-Edge Added group, the MERIS Terrestrial Chlorophyll Index (MTCI) was selected. (3) In Red-Edge Added group, the R2 of the RF model is better than that of the MLR model, 0.6719 and 0.5487, respectively, and the root mean square error (RMSE) of the RF model is smaller than that of the MLR model, 26.3m³·ha⁻¹ and 20.8m³·ha⁻¹, respectively. (4) After eliminating the effect of the model on the accuracy of forest stock volume retrieval, the R2 between the retrieval and the observed values in Red-Edge Added group increased by 11.6%, and the RMSE in Red-Edge Added group decreased by 9.1% compared to No Red-Edge group. Our results suggested that the red-edge band significantly improved the accuracy of the coniferous forest stock volume retrieval in Xining city. This study has high potential value in the remote sensing retrieval of forest stock volume.

Key words: GF-6 remote sensing satellite, Red edge band, Forest volume retrieval, Random forest model, Xining city