中国科学院遥感与数字地球研究所-高光谱遥感应用技术研究室

【作者】             张霞                                                                                                                                            

【导师】             童庆禧；郑兰芬；张兵

【 学位年度 】    2005

【论文级别】      博士    

【关键词】          种植模式，大气纠正，作物长势，光谱指数，时间谱去噪，物候特征提取  

【Key words】   Hyperspectral Remote Sensing; Sagnac Imaging Fourier Transform Spcetrometer; Image Simulation; Spectra Recovery 

【中文摘要】 

种植模式是作物轮作的空间表达，是对作物的空间分布和作物的前后茬顺序（轮作方式）的概括。作物种植模式信息对于高效、可控农业管理都具有非常重要之意义，二者布局的合理性关系着能否充分利用季节、光能和土壤肥力，能否改善和调节作物群体的小气候条件，是提高田块的单产和总产的重要技术环节。中国地域辽阔，作物种植模式复杂多样，传统的统计方法和点上的观测数据不能及时满足政府获取种植模式变化的要求。因此，通过遥感手段获取种植模式信息具有非常重要的现实意义。本文利用光谱指数时间序列跟踪作物物候变化规律，进行种植模式提取具有一定新意和挑战性。

        本文首先围绕表征作物长势的光谱指数发展、基于图像自身的大气纠正、光谱指数时间谱去噪标准化等关键问题展开方法研究，在此基础上，选择华北平原16天合成MODIS增强型植被指数（EVI）时间序列进行去云重构、物候特征提取，构建基于最少量特征参量的简洁分类决策树，生成了面向农业应用的土地覆盖分类图，在对EVI时间谱特征参量化与优化基础上，建立了冬小麦和夏玉米提取模型，制作了华北平原典型种植模式分布图。

        研究的主要结论如下：

1. 基于图像自身的模型法大气纠正程序ACORN对高光谱图像的光谱与辐射定标要求很高，欧空局的多角度航天成像光谱仪CHRIS图像在760nm以后波段不能满足ACORN对光谱与辐射定标的要求。

2. 设计了基于MODIS图像自身反演大气参量，实现大气纠正的流程，纠正后的反射率光谱能够较好地保持地物光谱真实特征。

3. 综合运用三种构建光谱指数的方法，进行了基于航空高光谱图像光谱的氮指数发展与氮填图，以NDVI（dr745，dr699.2）为估算因子得到的叶片全氮分布图与地面实测值具有很好的一致性；

4. 基于物理模型PROSAIL模拟分析了现有叶绿素光谱指数的敏感性分析，并对综合型光谱指数TCARI/OSAVI进行了改进，验证表明，改进模型无论在叶片尺度还是在冠层尺度上，无论是对作物还是林地，都可以取得较好结果，最后将模型应用到新型的星载成像光谱仪图像，得到叶绿素填图结果；

5. 以生物量和LAI作为长势关键因子，利用长时间序列的小麦测量光谱，进行了NDVI形式光谱指数的探索。对MODIS波段模拟表明，NDSI(b19,b2）和NDSI(b19,b17）及NDSI（b19,b16）与生物量和LAI的相关性远好于MODIS自身的两种指数，而MODIS_EVI对长势的敏感性又好于MODIS_NDVI。

6. 谐波分析方法HANTS能够在去除植被指数时间序列图像上云等噪声的影响的同时，很好地保持与揭示植被指数时间谱所蕴涵的物候周期性特征，由它处理后的EVI时间谱上提取的物候关键期与观测资料一致性很好，主要土地覆盖类型的MODIS_EVI时间谱很好地体现了自然植被各自特有的消长规律，且相互间具有一定的可区分性。

7. 从MODIS_EVI时间谱的物候分析出发，提取有效表征物候差异的特征参量，结合表征地气交互作用差异的温度（LST）信息及表征地表固有的空间分异特征－坡度信息，建立分类二叉树，分类结果表明，基于较少的属性数据得到的土地覆盖分类图较之NASA USGS的MODIS土地覆盖产品在精度上有了很大改善，且与耕地面积官方统计数据的吻合度大为提高。

       从种植模式的MODIS光谱指数时间谱分析入手，提取了表征冬小麦及其后茬夏玉米与其他同季生长作物物候差异的特征，分析表明返青期和峰值时期对冬小麦提取贡献最大，而峰值时间和大小以及表征营养生长向生殖生长转化时生物量累积差异的参量（偏斜度）对玉米提取贡献最大。17个县的抽样表明，冬小麦的总体提取精度为88.38%（以当地统计局的数据为标准）。最后形成的小麦－玉米，小麦－水稻和小麦－其他种植模式的分布图，与作物熟制气候分布图具有较好的一致性。MODIS低空间分辨率和种植模式的多样化趋势是造成误差的主要原因。

【Abstract】

According to FAO, cropping pattern is defined as the spatial representation of crop rotations, which embodies the spatial arrangement of crops and the yearly crop sequences. Cropping pattern plays an important role for effective and controllable agricultural management. Generally speaking, rational cropping pattern decides to a certain extent whether the planter can take beat advantage of the physical constraints of soil, climate and availability of irrigation to supplement rainfall and in part a response to market opportunities. Remote sensing data with its synoptic and frequently repeated coverage provides the undoubtedly best way to obtain this information and even the case especially with the development of spaceborne high-quality imaging spectrometers such as MODIS. However, studies have been restricted to crop phenology detection or extraction of multiple cropping index based on the multi-temporal high resolution images such as TM or coarse resolution satellite derived vegetation index time series. Few studies on remote sensing cropping system can be found in the literature. The objective of this dissertation is to make an attempt to extract cropping system information from temporal profile of optimized spectral index time-series (or temporal profile).

      This dissertation begins with the study on key issues related to atmospheric correction based on image itself, spectral index development to characterize the crop growth and the denoising and normalization of the temporal profile of the spectral index. On this basis, the yearly 16-day composite MODIS Enhanced Vegetation Index (EVI) time series covering North China Plain are collected. Noise-reducing and phonological feature extraction are performed on this data. A brief but reasonable decision tree based on less features is built so that one accurate agriculture-oriented land cover map is gained with an overall accuracy of 75.5％. Winter wheat and maize is extracted by the optimized features from the spectral index profile, and the typical cropping pattern map is gained.

      The main conclusions are drawn as follows:

1. The model-based atmospheric correction code ACORN is quite demanding for spectral and radiometric calibration performance of hyperspectral instrument. The new spaceborne imaging spectrometer CHRIS can’t meet the requirements of ACORN at least for bands beyond 750nm.

2. An atmospheric correction flow is designed for MODIS, the key atmospheric parameters could be inverted pixel by pixel. The corrected MODIS reflectance spectrum has good consistence in shape with the in-situ measured spectrum.

3. Three algorithms are utilized to develop nitrogen-sensible spectral index based on the rebuilt airborne hyperspectral image spectra and nitrogen mapping is carried out. The nitrogen map estimated by NDVI(dr745, dr699.2) as indicator is quite consistent with the value range of the ground truth data.

4. The sensibility analysis of the popular chlorophyll spectral indices is made by the canopy radiative transfer model PROSAIL simulation. An improvement is made on the integrated spectral index TCARI/OSAVI. Validation shows that the improved integrated index is sensitive to chlorophyll for crop and forest whether at foliar scale or canopy scale. Chlorophyll maps are gained successfully for two temporal CHRIS images.

5. Choosing biomass and leaf area index (LAI) as the best indicators of crop growth, the optimal NDVI-type spectral index of MODIS is explored by simulating the MODIS bands by long term ground measured wheat spectra. It turns out, NDSI(b19,b2）, NDSI（b19,b16）and NDSI(b19,b17）all have much stronger correlations with both biomass and LAI than the two MODIS vegetation indices, while MODIS_EVI is more sensible indeed to growth indicators than MODIS_NDVI.

6. The harmonic analysis algorithm for time series (HANTS) not only can reduce the noise effects in the MODIS EVI image cube, but also can reveal the periodical characteristics implied in the temporal EVI profile. The phenological stages determined by HANTS-processed EVI profiles agree with regular observations.

7. Based on temporal analysis of MODIS_EVI profile, five phenological features are extracted to measure the calendar difference, in combination with land surface temperature to measure interaction difference of surface-atmosphere, and slope to characterize the spatial variability of land cover classes, a brief but reasonable decision tree is built to make an accurate agriculture-oriented land cover map with an overall accuracy of 75.5% , obviously higher than that of NASA USGS product. Most of all, when compared with the cropland area from official statistics, our classification shows much higher consistence with an overall mean square root error of 34.0507 kilo-hektare versus 66.1205 kilo-hectare by USGS product.

      Phenological features of winter wheat and its succession maize are extracted and analyzed. Re-green stage and peak stage(heading stage) are of great importance for winter wheat extraction, while peak stage (silking stage) and the skew parameter which characterizes the accumulated biomass difference before and after silking stage are most important for maize extraction. The overall accuracy of wheat mapping is 88.38% when compared with official statistics in county unit. The cropping system map of wheat-maize, wheat-rice, and wheat-other crop is quite consistent with the rotation feasible map regarding on climate distribution. The error is caused mostly by the coarse resolution of MODIS and the great spatial variability in regional scale.