遥感图像时间序列可以捕捉植物生长和物候变化的季节特征，在土地覆盖分类、环境变化、作物监测等领域得到了广泛应用，已成为描述全球气候变化以及其对陆地生态系统影响的重要指标。但是遥感数据受到大气污染和传感器自身硬件问题的影响，在地表反射率产品中会观察到残余云或未检测到的云影，这总是导致后续下游产品的时间和空间不一致，在应用需要的时间和地点往往找不到直接观测到的高质量遥感图像数据，因此常常需要进行地表反射率数据的重建。而以前的NDVI、反射率重建大多基于低分辨率的遥感影像，本研究提出在Google Earth Engine（GEE平台）利用基于离散余弦变换的约束最小二乘回归方法（penalized least-squarere gression based on discrete cosine transform，DCT-PLS）算法来重建高分辨率的10m的Sentinel-2 NDVI和反射率数据，可以为相关地学研究提供支持。本研究思路先是在MATLAB用离散余弦变换的DCT-PLS算法对典型植被样本点NDVI和反射率时间序列进行重建，并与SG滤波进行比较，评价精度，改良算法。然后将DCT-PLS时间序列滤波算法移植到GEE平台进行大尺度区域的NDVI、反射率重建，并评价不同参数设置下的图像效果。然后选取农作物种类丰富的典型实验区，应用算法在GEE上生成高时空分辨率图像集，并下载数据。最后，选取反映不同作物生育期差异的时间序列特征因子，对研究区的作物进行时间序列的精细分类，利用地面调查数据验证分类精度。

具体得到以下结论：

（1）发现DCT-PLS算法能很好地识别出受污染的NDVI点，而SG滤波不能识别，DCT-PLS算法重建的NDVI时间序列更加平滑，并且在阶数N=24，控制平滑的程度的正标量s选取13-20时重建的效果最好，留一法均方根误差RMSE_cv均小于0.1，R²均达到90%以上。而在NDVI时间序列重建过程中剔除了受污染的异常数据的基础上，进一步使用DCT-PLS算法重建多波段反射率数据，得到的反射率时间序列平滑，符合植被光谱响应特征和植被生长规律。

（2）在GEE平台重建NDVI、反射率图像的过程中，本研究按重建日期及其临近日期影像云量多少的不同情况对两个实验区进行分析，得到以下结论：在临近日期高质量遥感影像较多的情况下， ①s=16时能很好地重建植被地物的NDVI、反射率时间序列，R²、RMSE_fit均达到达到理想情况，并且植被越多，拟合效果越好，而当s增大会导致过度平滑，R²会降低、RMSE_fit会增大；②s=16时对水体部分可能不能很好地进行重建，需要提高s的值，比如s=30、50、100等能够较好地进行拟合重建；③在临近日期高质量遥感影像较少的情况下， s=16时不能很好地重建，需要提高s的值，比如s=50、100、150等可以得到较稳定的重建效果。

（3）在张家口市实验区的农作物分类中，使用了基于简单去云填补的、DCT-PLS方法重建的NDVI时间序列特征点的决策树分类和基于农作物生长旺季时期的平均影像的随机森林分类三种方法。重建后的决策树分类效果明显高于未重建的决策树分类，而虽然重建后决策树分类与随机森林分类结果有些许差异，差异主要在玉米、其他作物类别，而在大部分区域覆盖比较接近，并且决策树分类和随机森林分类都有较高的总体精度与Kappa系数，并且决策树分类的总体精度与Kappa系数优于随机森林分类。这证明时间序列数据在作物种类精细分类中具有比较好的应用价值。

Remote sensing image time series can capture the seasonal characteristics of plant growth and phenological changes, and have been widely used in land cover classification, environmental change, crop monitoring and other fields, and have become an important indicator to describe global climate change and its impact on terrestrial ecosystems. However, the remote sensing data subject to atmospheric pollution and the sensor’s hardware imperfection, residual clouds or undetected cloud shadows are observed in surface reflectance products, which always leads to time and space inconsistencies in subsequent downstream products, and directly observed high-quality remote sensing image data are often not found at the time and location needed for the application, so reconstruction of surface reflectance data is often required. The previous NDVI and reflectance reconstruction researches are mostly based on low-resolution remote sensing images. This study proposes to use the penalized least square regression based on discrete cosine transform (DCT-PLS) algorithm on Google Earth engine (GEE) to reconstruct 10 m resolution Sentinel-2 NDVI and reflectance data, which can provide support for relevant geological research. The idea of this research is to reconstruct the NDVI and reflectance time series of typical vegetation species samples with DCT-PLS in MATLAB, compare them with Savitzky-Golay (SG) filter, evaluate the accuracy and improve the algorithm. Then, the DCT-PLS time series filtering algorithm was transplanted to the GEE platform for large-scale regional NDVI and reflectance reconstruction, and the resulting image effects under different parameter settings were evaluated. Finally, a typical experimental area with rich crop species was selected, and the algorithm was applied to generate a high spatial-temporal resolution image on GEE, and the data was downloaded. The time series characteristic factors reflecting the differences of different crop phenological were selected to carry out fine classification of crops in the study area, and the classification accuracy was verified by ground survey data.

The specific conclusions are as follows:

(1) It is found that the DCT-PLS algorithm can identify the contaminated samples in NDVI series well, while the Savitzky-Golay filter cannot. And when the order N=24 and the real positive scalar s which controls the degree of smoothing is selected from 13 to 20, the reconstruction effect is the best. The root mean square error (RMSE_cv) of the leave-one-out method is less than 0.1, and the R² reaches more than 90%. On the basis of removing the contaminated abnormal data in the NDVI time series process, the DCT-PLS algorithm is further used to reconstruct the multi-band reflectance data. The obtained reflectance image is clear, and the vegetation spectral curve confirmed to the growth law.

(2) In the process of reconstructing NDVI and reflectance images on GEE platform, This study analyzes the two experimental areas according to the cloud amount of the image on the reconstruction date and its adjacent date, and obtains the following conclusions: ① When s=16, the NDVI and reflectance time series of vegetation objects can be well reconstructed. Both R² and Fitted root mean square error (RMSE_fit) reach the ideal ssituation, and the more vegetation, the better the fitting effect. When s increases, it will lead to excessive smoothing. R² will decrease and RMSE_fit will increase; ②The water body part may not be reconstructed well when s=16, and the value of s needs to be adjusted, such as s = 30,50, 100, etc; ③ When there are few high-quality remote sensing images near the date, s = 16 can not be reconstructed well. It is necessary to adjust the value of s, such as s = 50, 100, 150, etc. to obtain a more stable reconstruction effect.

(3） In the crop classification of Zhangjiakou experimental area, three methods are used: decision tree classification based on the feature points of rough cloud removal and filling NDVI time series, decision tree classification based on the feature points of NDVI time series reconstructed by DCT-PLS method and random forest classification based on average images in peak season of crop growth. The classification performance of decision tree classification based on the feature points of reconstructed NDVI time series is singnificantly higher than that of the feature points of NDVI time series based on rough cloud removal filling. Although the results of the decision tree method base on the featurepoints of reconstructed NDVI time series and ramdom forest method, differences are mainly in corn and other crop categories, both decision tree classification and random forest classification have hreached satisfactory performance. The decision tree classification has better overall accuracy and Kappa coefficient than random forest classification. This proves that time series data has good application value in fine classification of crop species.