基于AVHRR影像的北半球积雪识别算法

针对2000年前北半球较高时空分辨率和高精度的历史积雪范围数据缺失问题，利用NOAA-AVHRR地表反射率数据，以Landsat-5 TM生成的积雪范围影像作为参考真值，优化基于多指标的多级决策树积雪识别算法的阈值，并结合云雪混淆区分技术，生成了北半球AVHRR 1981—1999年L1级逐日积雪范围数据集。此外，针对AVHRR在高纬度地区数据完全缺失和低纬度地区数据部分缺失问题，利用微波雪深数据集进行填充，生成了北半球L2级逐日积雪范围数据集。最后，利用北半球1981—1999年间2 546个气象台站记录的雪深数据和939景Landsat-5 TM参考积雪范围影像作为验证数据，对AVHRR积雪范围数据集进行了精度验证。结果表明：L1级和L2级数据集的总体精度分别为81.8%和82.2%，用户精度分别为83.7%和83.8%，生产者精度分别为81.7%和84.2%，说明算法精度较高，错分误差和漏分误差均比较均衡。进一步利用Landsat-5 TM参考积雪范围影像对L2级数据集进行面上精度评估，发现L2级数据集的总体精度为90.3%，用户精度为90.2%，生产者精度为99.1%，L2级数据集精度较高。生成的北半球历史数据集可为全球积雪变化研究提供有效数据补充。

Snow discrimination algorithm in the Northern Hemisphere based on AVHRR image

Due to the lack of snow cover extent dataset with slightly high spatiotemporal resolution and high accuracy in the Northern Hemisphere from 1981 to 1999， we used NOAA-AVHRR version 4 surface reflectance data as the basic input data and Landsat-5 TM snow maps as reference maps， to obtain the optimal threshold of snow discriminant algorithm based on multi-indicator multi-level decision tree. Then， combined with the cloud discrimination algorithm， we produced the Northern Hemisphere AVHRR L1 （Level 1） daily snow cover extent dataset. In addition， aimed at the complete lack of AVHRR dataset in high latitudes and partial lack of data in low latitudes， we filled the gaps of AVHRR L1 dataset by the Northern Hemisphere 0.25° snow depth dataset， and generated AVHRR L2 （Level 2） daily snow cover extent dataset. Finally， taking 2 546 ground snow depth stations measurements from 1981 to 1999 and 939 Landsat-5 TM snow maps as validation data， accuracy of the Northern Hemisphere AVHRR daily snow cover dataset was accessed. The results showed that overall accuracy （OA） of AVHRR L1 and L2 dataset is 81.8% and 82.2%， user’s accuracy （UA） is 83.7% and 83.8%， and producer’s accuracy （PA） is 81.7% and 84.2%， respectively. These accuracies were relatively high， and commission error and omission error were relatively balanced. Furthermore， we used Landsat-5 TM snow maps for L2 dataset to perform accuracy assessment. The results showed that OA of L2 dataset is 90.3%， UA is 90.2% and PA is 99.1%. Therefore， our product has virtually provided more reliable snow knowledge over the Northern Hemisphere， and thereby can better serve for study on global snow cover change.