长时序高质量的归一化差值植被指数是理解植被动态变化的基础，在精确预估地表植被的季节性模式、年际变化趋势等方面发挥着关键作用。这些数据对追踪和分析植被对气候变化、人类活动和自然干扰的响应，评估植被生产力变化，以及监测荒漠化、森林退化和生态系统恢复过程至关重要。然而，目前长时序NDVI数据存在的时空不一致的问题，这在一定程度上限制了数据的使用。例如，AVHRR NDVI数据虽然时间覆盖长，但空间分辨率相对粗糙，无法精确评估植被生长趋势；MODIS NDVI数据记录了2000年以后的数据，无法追溯至2000年之前的历史信息。为解决以上问题，本研究应用机器学习和深度学习技术重建了1982-2015年16天/1km NDVI数据集。具体做法如下：

（1）基于多变量随机森林模型（MRF）方法，较好解决了2000年以后MODIS NDVI数据时空不连续的问题。利用质量控制文件筛选出高质量的MODIS NDVI数据，并将其作为基准数据集，同时考虑低质量区域NDVI出现的背景信息，将地理数据、气候数据及生态数据作为辅助数据，采用MRF进行了低质量数据的重建工作。验证结果显示MRF模型的均方根误差（RMSE）值为0.0346，决定系数（R²）为0.9740，平均绝对误差（MAE）为0.0238。时序曲线与空间分布图的结果进一步验证了MRF模型在精确捕获NDVI的变化趋势及准确反映地表植被真实状态方面的能力。本研究成果为后续基于深度学习方法进行的空间降尺度研究提供了高质量的样本数据。

（2）提出了一种多尺度残差网络空间降尺度（MRCNN）方法，提升了2000年以前AVHRR NDVI数据的空间分辨率，并将AVHRR NDVI降尺度的1982-2000年数据集与MODIS NDVI数据续接，形成了一套1982-2015年时间跨越34年16天的空间分辨率为1km的NDVI产品。具体工作如下：本研究结合多尺度网络结构和残差网络结构形成一种新型网络结构。MRCNN空间降尺度方法网络能通过学习AVHRR NDVI数据（低分辨率）以及辅助数据（30m DEM以及CLCD）与高质量样本数据之间的映射关系，实现不同空间分辨率间的转换。验证结果显示RMSE从0.2177减少至0.1789，降低了17.82%，而MAE从0.0969减少至0.0672，减幅达到30.65%。这些改进结果表明了方法在提高预测精度方面的有效性。此外，R²从0.5106增加到0.6926，增长了35.64%，反映了模型对数据变异的解释能力有显著提升。在图像质量评估方面，峰值信噪比（PSNR）从23.2185提升至26.3729，增加了13.59%，表明图像的整体质量和清晰度有所改善。结构相似性指数（SSIM）也从0.8897提高至0.9199，增加了0.0302，增幅为3.394%，显示图像的结构保真度得到提升。这些综合结果表明，MRCNN方法在提升AVHRR NDVI数据的空间分辨率和整体图像质量方面表现优异。

（3）针对地形复杂场景，引入注意力机制，发展了高性能空间尺度降尺度模型（HPSR），进一步提升了复杂区域降尺度的准确性。研究表明，HPSR算法显著减少了MAE和RMSE（MAE降至0.062和RMSE降至0.074），与AVHRR NDVI相比降幅分别为57.24%和56.21%。同时，将PSNR提高了80%和SSIM提高了23.94%。此外，通过对不同时间尺度和地物类别的适应性分析发现HPSR展现出对复杂地形和物种多样性丰富区域的良好适应性。

本文的研究解决了目前公里级长时序NDVI数据集时空一致性不足的问题，形成了一套时间分辨率为16天且空间分辨率为1km的高质量NDVI数据集，提升了植被变化趋势估算的准确性和可靠性，具有重要的理论价值和实际意义。

关键词：归一化差值植被指数，低质量区域重建方法，多变量随机森林，空间降尺度方法，多尺度残差网络，注意力机制

Long-term, high-quality Normalized Difference Vegetation Index (NDVI) data is fundamental in understanding vegetation dynamic changes, playing a key role in accurately estimating the temporal trends, seasonal patterns, interannual variations, and long-term trends of surface vegetation. These data are essential for tracking and analyzing vegetation responses to climate change, human activities, and natural disturbances, evaluating changes in vegetation productivity, as well as monitoring desertification, forest degradation, and ecosystem recovery processes. However, the spatiotemporal inconsistency in current long-term NDVI data somewhat limits their utility. For instance, while AVHRR NDVI data offer extensive temporal coverage, their spatial resolution is relatively coarse, making it challenging to precisely evaluate vegetation growth trends. On the other hand, MODIS NDVI data only record information post-2000, lacking historical data pre-2000. To address these limitations, this study employed machine learning and deep learning techniques to reconstruct a 16-day/1km NDVI dataset spanning from 1982 to 2015. The specific methodology is outlined as follows:

(1) Based on the Multivariate Random Forest (MRF) model, the study effectively addressed the temporal and spatial discontinuities in MODIS NDVI data after 2000. Specifically, high-quality MODIS NDVI data was selected using quality control files as the reference dataset. Background information of low-quality NDVI regions was considered, and geographic, climate, and ecological data were incorporated as auxiliary data to reconstruct the low-quality data using MRF. Validation results showed that the MRF model had a Root Mean Square Error (RMSE) value of 0.0346, a Coefficient of Determination (R²) of 0.9740, and a Mean Absolute Error (MAE) of 0.0238. The results from temporal curves and spatial distribution maps further validated the MRF model's ability to accurately capture NDVI trends and reflect surface vegetation conditions. This research provides high-quality sample data for subsequent spatial downscaling studies using deep learning methods.

(2) A Multi-Scale Residual Convolutional Neural Network Spatial Downscaling (MRCNN) method was proposed to enhance the spatial resolution of AVHRR NDVI data before 2000. The downscaled AVHRR NDVI dataset from 1982-2000 was integrated with MODIS NDVI data to form a 1982-2015 NDVI product spanning 33 years with a spatial resolution of 1km. By combining multi-scale network structures and residual network structures, the MRCNN spatial downscaling method learned the mapping relationship between low-resolution AVHRR NDVI data and auxiliary data (30m DEM and CLCD) with high-quality sample data to achieve resolution conversion between different spatial scales. Validation results showed that the RMSE decreased from 0.2177 to 0.1789, a reduction of 17.82%, while the MAE decreased from 0.0969 to 0.0672, a reduction of 30.65%. These improvements indicate the effectiveness of the method in enhancing prediction accuracy. Additionally, the R² increased from 0.5106 to 0.6926, a growth of 35.64%, reflecting a significant enhancement in the model's ability to explain data variations. In terms of image quality assessment, the Peak Signal-to-Noise Ratio (PSNR) increased from 23.2185 to 26.3729, a 13.59% increase, indicating an improvement in overall image quality and clarity. The Structural Similarity Index (SSIM) also improved from 0.8897 to 0.9199, an increase of 0.0302, representing a 3.394% increase, showing an enhancement in image structural fidelity. These comprehensive results demonstrate the excellent performance of the MRCNN method in enhancing the spatial resolution and overall image quality of AVHRR NDVI data.

(3) For complex terrain scenes, an attention mechanism was introduced to develop a High-Performance Spatial Resolution Downscaling Model (HPSR), further improving the accuracy of downsizing complex regions. The study showed that the HPSR algorithm significantly reduced the MAE and RMSE (with MAE reduced to 0.062 and RMSE to 0.074), achieving reductions of 57.24% and 56.21% compared to AVHRR NDVI. Moreover, the PSNR was increased by 80% and the SSIM was increased by 23.94%. Additionally, adaptive analysis of different time scales and land cover categories revealed that the HPSR demonstrated good adaptability to complex terrains and regions with rich species diversity.

The study addressed the issue of inadequate spatiotemporal consistency in current kilometer-scale long-term NDVI datasets, thereby establishing a high-quality NDVI dataset with a temporal resolution of 16 days and a spatial resolution of 1km. This advancement enhances the accuracy and reliability of vegetation change trend estimation, holding significant theoretical and practical implications.

KEY WORDS: Normalized Difference Vegetation Index, Low-Quality Area Reconstruction Method, Multivariate Random Forests, Spatial Downscaled methods, Multi-scale Residual Networks, Attention Mechanisms.