青藏高原不同土地覆盖类型下积雪面积判别算法优化

Optimization of snow area discrimination algorithm under different land cover types in Qinghai-Tibet Plateau

MODIS V006版本数据仅提供了归一化积雪指数(NDSI),而用户往往关心的是直观的积雪分类,包括积雪范围或积雪覆盖率。美国国家冰雪数据中心推荐全球积雪范围最佳的NDSI阈值为0.4,但是青藏高原地形复杂多样,积雪斑块化特征明显,单一阈值并不能精确地判识不同下垫面上的积雪。青藏高原被称为地球的第三极,是中国三大稳定积雪区之一,蕴藏了大量的淡水资源。随着全球气候变暖,青藏高原地区积雪融化时间提前,冰川融水增加,影响河流水量,造成洪涝灾害,进而影响人类正常生产生活,因此通过确定不同下垫面阈值,改善传统阈值的积雪高估低估现象,提高积雪识别精度,进而更准确地探究青藏高原积雪状况,显得尤为迫切。本文以青藏高原为研究对象,首先生成MODIS逐日无云NDSI序列并进行验证;其次对应站点雪深数据与NDSI序列,证实在下垫面为林地和非林地的区域,去云NDSI序列与站点雪深均有良好的对应关系,确定不同下垫面最优阈值范围;最后在最优阈值范围内通过混淆矩阵确定最优阈值。计算得出,林地NDSI=0.03时,总体精度最高为94.02%,在该NDSI之下,高估误差OE和低估误差UE分别为1.21%和4.60%;非林地NDSI=0.26时,总体精度OA最高为94.27%,在该NDSI之下,高估误差OE和低估误差UE分别为0.51%和5.03%。因此选取优化后林地阈值为NDSI=0.03,非林地阈值为NDSI=0.26。为避免地面常规观测资料尺度上的局限性,本文采用高精度的Landsat 8 OLI卫星数据识别结果,作为“真值”对优化后阈值的判别结果进行“像元—像元”级别的验证。在定量验证中,优化后NDSI阈值对MOD10A1 V006积雪判别结果的总体精度OA为84.21%,高估误差OE为5.33%,低估误差UE为10.46%;传统阈值对MOD10A1 V006积雪判别结果的总体精度OA为82.86%,高估误差OE为1.48%,低估误差UE为15.66%。可以看出在定量验证中,优化后阈值的积雪判别精度更高。同时在定性验证中,积雪大面积集中的区域,新的阈值与传统阈值提取效果均相对较好;积雪相对分散破碎的区域,优化后阈值能提取出大量积雪,传统阈值则不能。这表明考虑不同土地覆盖类型下的NDSI阈值优化可以有效地提高青藏高原积雪判别精度,为NDSI在积雪识别中的应用提供有力的支撑,有助于更准确地了解该地区积雪分布状况。

The data provided by MODIS V006 version is the Normalized snow cover Index(NDSI),but the ma⁃jority of users are often concerned with the intuitive snow cover classification results,including snow cover ex⁃tent or snow cover rate.The National Snow and Ice Data Center(NSIDC)recommended 0.4 is the best NDSI threshold for the global snow cover.However,the Qinghai-Tibet Plateau has complex and diverse terrain and obvious snow patch characteristics,so a single NDSI threshold of 0.4 cannot accurately distinguish the snow cover on different underlying surfaces.The Qinghai-Tibet Plateau,known as the third pole of the earth,is one of the three stable snow areas in China and contains a large amount of fresh water resources.With global warm⁃ing,the Tibetan plateau ahead of time,the snow is melting glaciers,increase,affect the rivers of water,caus⁃ing floods,and thus affect the normal production of human life,so determining different underlying surface threshold,improve the traditional threshold value of snow overestimated underestimate phenomenon,improve the identification accuracy of snow,and then more accurate study of the Tibetan plateau snow conditions,Is par⁃ticularly urgent.In this study,This paper takes the Qinghai-Tibet Plateau as the research object.Firstly,MO⁃DIS daily cloud-free NDSI sequence is generated and its reliability is verified.Secondly,the underlying surface is forested and non-forested areas,and the NDSI sequence of cloud removal has a good corresponding relation⁃ship with the snow depth at the site.NDSI can accurately reflect the snow melting phenomenon of the pixel where the station is located.Determine the optimal threshold range of different underlying surface;Finally,the optimal threshold was determined by confusion matrix within the optimal threshold range.When NDSI=0.03,the highest overall accuracy was 94.02%.Under this NDSI,the overestimation error OE and underestimation er⁃ror UE were 1.21%and 4.6%,respectively.When NDSI=0.26 for non-forestland,overall accuracy(OA)is 94.27%.Under this NDSI,he overestimates error(OE)and underestimate error(UE)are 0.51%and 5.03%,respectively.Therefore,the optimized threshold of forestland is NDSI=0.03,and that of non-forestland is NDSI=0.26.Because snow cover is a large scale phenomenon,the conventional observation data are mostly point scale observations.In order to avoid the limitations in the scale of conventional ground observation data,this pa⁃per uses the high-precision Landsat 8 OLI satellite data identification results as the“truth value”and the snow discrimination results of the optimized threshold and the snow discrimination results of the traditional threshold to verify the“pixel to pixel”level.In quantitative verification,the overall accuracy(OA)of the optimized NDSI threshold to MOD10A1 V006 snow discrimination results is 84.21%,the overestimation error OE is 5.33%,and the underestimation error(UE)is 10.46%.The overall accuracy(OA)of the traditional threshold for the snow discrimination results of MOD10A1 V006 is 82.86%,the overestimation error(OE)is 1.48%,and the underestimation error(UE)is 15.66%.It can be seen that in the quantitative verification,the snow dis⁃crimination accuracy of the optimized threshold is higher.At the same time,it can be seen from the qualitative verification that the new threshold and the traditional threshold are relatively good at snow recognition in the area with large area of concentrated snow.In the region with relatively scattered and broken snow,the optimized threshold can identify a large number of snow,while the traditional threshold cannot identify the same number of threshold.These results indicate that NDSI threshold optimization considering different land cover types can ef⁃fectively improve the accuracy of snow discrimination on the Qinghai-Tibet Plateau,which provides a strong support for the application of NDSI in snow recognition.It is helpful to understand the snow distribution in this area more accurately.