基于多尺度特征融合网络的新疆积雪覆盖度估算

受复杂地形和遥感数据低分辨率的影响，传统的二值化积雪遥感产品在山区和林区的积雪覆盖度计算中存在严重误算和漏算的问题，从而导致积雪覆盖度估算精度低。基于风云四号A星多通道辐射扫描计（AGRI）新疆地区的遥感影像数据，提出一种多尺度特征融合网络的积雪覆盖度估算方法。通过深度残差网络和特征金字塔模式对卷积层各个阶段的特征信息进行重构，融合深层和浅层特征的多重语义信息，同时结合AGRI数据高时间分辨率的特性，拟合光谱信息和地理因素间的非线性关系，从而提高数据源和特征信息的整体利用率。实验结果表明，相比MOD10_FSC、BP-ANN_FSC和ResNet_FSC方法，该方法在A1~A4样本区中相关系数均值和解释回归模型的方差得分均值最高可提高8和6个百分点，且其均方误差均值仅为0.1，能够获得较高精度的积雪覆盖度估算结果。

Xinjiang Fractional Snow Cover Estimation Based on Multi-Scale Feature Fusion Network

Binary fractional snow remote sensing products are affected by complex terrain and low resolution of remote sensing data. The traditional binary fractional snow remote sensing products have serious miscalculation and omission in the calculation of Fractional Snow Cover(FSC)in mountainous and forest areas,resulting in low accuracy of FSC estimation.Based on the remote sensing image data of the Advanced Geosynchronous Radiation Imager(AGRI)of FY-4 A in Xinjiang,a FSC estimation method based on multi-scale feature fusion network is proposed. The feature information of each stage of convolution layer is reconstructed through depth residual network and feature pyramid mode,and the multiple pieces of semantic information from deep and shallow feature layers are fused. Combined with the characteristics of high time resolution of AGRI data,the proposed method can better fit the nonlinear relationship between spectral information and geographical factors,improving the overall utilization of data source and feature information. The experimental results show that compared with the MOD10_FSC,BP-ANN_FSC and ResNet_FSC methods,in the four sample areas A1~A4,the mean value of correlation coefficient and the mean value of the variance score of the interpretation regression model are increased by up to 8 and 6 percentage points,and the mean square error is only 0.1,providing more accurate FSC estimation results.