【摘要】 MODIS以其时间分辨率、光谱分辨率的优势成为全球及区域土地覆盖研究的主要数据源。但如何快速准确的提取所需土地覆盖信息一直是科学界研究的焦点问题。对于NDVI时序数列分类方面的研究很多,其中影响分类精度的一个重要因素就是NDVI的数据质量问题。本文通过试验发现经过Savizky-Golay滤波处理的NDVI时序数列能够反映植被季相变化特征,与传统的滤波效果相比有明显改善,更符合实际情况。通过分析数据的波谱曲线,滤波后的时序数列能较好的区分植被与非植被、草本(一年生)与木本(多年生)覆盖类型。但研究区内一年一熟的农作物与高盖度草地、落叶针叶林与落叶阔叶林具有相似的物候特征,仅通过NDVI序列很难区分。为解决这一问题,本研究利用MODIS地表温度(land surface temperature,LST)产品对NDVI时序数列修正,利用前5个主成分进行分类。所得分类结果用363个野外调查样区进行验证,总分类精度达到了69.15%,kappa系数为0.6499。结果表明添加LST的时序数列比单纯的NDVI夸大了覆盖类型的差异,提高了分类结果的精度。为充分发挥MODIS高时间分辨率的优势,下一步应对多源数据进行定量分析,结合植被的物候关键期识别土地覆盖类型,必将进一步提高分类精度。更多还原

【Abstract】 MODIS was built based on seven bands specifically designed for land cover monitoring,wherein an improved spectral/spatial response compared to AVHRR.this allows for greater accuracy in distinguishing different land cover categories.The paper investigated the regional land cover classification with MODIS time-series data.Northeast China is the ideal area of research on MODIS land cover classification for its pure and diversified land cover types.The normalized difference vegetation index(NDVI) increased with the growth of the vegetation,and gradually decreased after reaching the maximum at some growth stage.This characteristic is similar to hyper-spectral data,whose pixels having similar NDVI profiles are the same land cover classes;however,MODIS with high spectral and temporal resolution is more sensitive to land cover.The important factor to the accuracy classification is the quality of the NDVI.So the first step of classification is to find a method to produce the high-quality NDVI data.The maximum value composite(MVC) can eliminate the contaminated data and abnormal data in the NDVI multi-temporal image at some degree.But the profile of NDVI after compositing cannot effectively reflect the change of vegetation growth in a year.SavizkyGolay filter was used to smooth the 10-day compositing data.The result proved that the time-series data was more correlational with the vegetation growth.Through analyzing the profiles,NDVI time-series data can distinguish the perennial woody and herbaceous vegetation and non-vegetation categories depending on the seasonal differences.Grassland and cropland,needle-leaf-deciduous forest and broadleaf-deciduous forest have similar characteristics easy to be confused.We add the LST(land surface temperature) data to resolve this problem and select five principal components for classification.Five principal components are selected as the data source to join classification processing after principal analysis.Validating results with 363 filed samples,the overall classification accuracy of new time series data is 69.15% and kappa index is 0.6499.The accuracy of grassland and cropland,needle-leaf-deciduous forest and broadleaf-deciduous forest are all above 70%.From the result,we concluded that LST is correlated to altitude,latitude and other natural factors.So LST & NDVI is more sensitive to land cover than NDVI;MODIS data is good at updating the regional land cover classification.If the model of remote sensing time-series data and the key plant phenophase were established,the accuracy of land cover classification would be improved at a large degree.更多还原