植被是陆地生态系统的重要组成部分，对物质及能量循环具有重要调节作用，是地球天然的“降温器”、“净化器”、“保护器”。全球农田植被覆盖面积约占总植被覆盖面积的11%，是保障全球粮食安全的基本单元。由于耕作制度的原因，农田植被与自然植被生长及物候特征具有显著差异。城市化被誉为气候变化的“自然实验室”，阐明城郊农田植被生长特征对城市化不同发展模式的响应规律，量化城市化导致的环境变化与城郊农田植被生长特征的时空关联，辨识城郊景观格局变化对农田植被生长特征的影响对预测未来作物产量、优化田间管理与布局、制定城郊农业发展政策等方面均具有重要理论及现实意义。

本文基于多源遥感数据、地表观测数据及再分析数据等多源数据集，研发了地理智能机器学习气温反演算法；构建了公里网格尺度地表气象要素数据集；探讨了城市化对城郊农田植被健康状况的多重影响特征；评估了城市化诱发的城郊农田区域水热状况变化与城郊农田植被健康状况变化情况的相关关系；研究了考虑不同耕作制度下不同关键物候期城郊农田植被生产力对气候变化“自然实验室”的总体和间接响应规律及其与局部气温及CO₂变化的时空关联；量化了城市化驱动的农田景观格局变化对单位农田植被生产力的影响规律；分析了不同气候背景下对城郊单位农田植被生产力影响最大的景观格局因子。主要研究结果如下：

(1) 率先提出了耦合自适应时空自相关的地理智能机器学习模型，显著地提高了公里网格尺度地表气温预估精度。本文提出了一种新型的气温重建算法，该算法通过耦合自适应时空自相关算法及融合甄选的多源异构数据集，对三种传统机器学习模型（随机森林、反向传播神经网络、支持向量机）进行了改进。改进前的模型分别记做Ori-RF、Ori-BPNN、Ori-SVM，改进后的模型分别记做Geoi-RF、Geoi-BPNN、Geoi-SVM。并通过上述6种模型反演出中国2003-2012年公里网格空间尺度的月平均地表气温数据。然后，基于多种统计方法对比算法改进前后各气温预估模型效率参数及其在不同时空尺度上的表现情况，从多维度定量评估各模型对地表气温预估的准确性及稳健性。研究结果表明：融合甄选的多源数据并耦合自适应时空自相关算法的随机森林模型(Geoi-RF)能够显著提高多云区（四川）、资料稀缺地区（青藏高原）以及城市区域的气温预测结果的准确性。此外，Geoi-RF与本研究中涉及到的其他模型相比，无论是总体还是全局，皆表现最优。其对平均气温、最高气温、最低气温预报的精准率（绝对误差小于0.5K（开尔文）的验证点占总验证点的比率）分别提高了18.51%-63.17%、15.57%-44.62%、16.53%-45.55%。此外，Geoi-RF的R²接近于1，平均绝对误差(MAE)均低于0.25K，均方根误差(RMSE)均低于0.5K，故该模型是估算大范围精细尺度地表气温较为理想的算法。

(2) 变化环境下，城市化对城郊农田植被健康状况的影响足迹及间接影响具有显著的空间异质性。本论文综合应用空间缓冲区分析、混合像元分解等多种技术手段，全面量化了城市化对不同气候背景下中国九个农业区内32个主要城市城郊农田植被的总体及间接影响。结果表明，城市化对城郊农田植被总体健康状况的影响足迹沿城乡梯度呈显著指数型递减趋势，在城市化发展较快的区域影响半径可达26km以上。城市化对中国城郊农田植被健康状况的间接影响在空间格局上总体呈现“北正南负”的空间格局。在城市化的间接影响下，位于中国北方农业区的城郊农田植被健康状况随着城市化强度的提高而改善，这种间接影响平均抵消约44％的由于不透水面扩张引起的城郊农田植被健康状况损失。而在气候条件相对温暖潮湿的南方农业区，城市化对城郊农田植被的间接影响整体多呈现负面效应。此外，在隶属北方农业区的黄淮海平原区(HHHP)及东北平原区(NCP)，城市化的间接作用对城郊农田植被健康状况的促进作用与城郊地表气温变化情况呈显著正相关。而在属于南方农业区的长江中下游平原区(MYP)，城郊土壤湿度的变化与城市化对城郊农田植被健康状况的间接影响呈显著正相关。

(3) 变化环境下，城市化对不同关键物候期农田植被累积生物量的间接影响大体呈显著的促进作用。但在不同气候背景下，这种促进作用会有明显时空异质性。城市化导致的不透水面扩张（直接影响）造成单位农田种植面积下降，致使三大农业区的单位农田植被生长期累积生物量(TINDVI)、花前生长期累积生物量(TINDVIBeforeMax)、花后生长期累积生物量(TINDVIAfterMax)对城市化强度(ISA)的整体响应方式表现为负反馈。但城市化的间接影响分别对长江中下游平原区(MYP)、黄淮海平原区(HHHP)及东北平原区(NCP)78%(80%/76%)、92%(95%/64%)、98%(82%/96%)的城市化农田子区间的TINDVI (TINDVIBeforeMax/TINDVIAfterMax)起到了促进作用。城市化对不同生长期农田植被累积生物量产生的正影响及其波动性均随城市化强度增大而显著增强。在三大农业区，城市化对不同生长期农田植被累积生物量的间接影响方差突变区间均出现在不透水面比率约50%。

(4) 以三大农业区为研究案例，研究发现城市化对城郊农田植被的间接影响均可以抵消掉部分由城市化的直接影响造成的TINDVI (TINDVIBeforeMax/TINDVIAfterMax)损失。在长江中下游平原区(MYP)、黄淮海平原区(HHHP)、东北平原区(NCP)，城市化的间接影响分别约可抵消掉17%(33%/19%)、39% (49%/12%)、40% (30%/46%)由不透水面扩张造成的城郊TINDVI (TINDVIBeforeMax/ TINDVIAfterMax)的损失。此外，在东北平原区(NCP)，城市化的间接作用对TINDVIAfterMax的促进效应相较于TINDVIBeforeMax更为显著；在黄淮海平原区(HHHP)，城市化的间接作用对TINDVIBeforeMax的促进效应相较于TINDVIAfterMax更为显著；在长江中下游平原区(MYP)，城市化的间接作用对TINDVIBeforeMax和TINDVIAfterMax的影响差异相对较小。

(5) 不同气候背景下，城郊的“热岛效应”、“碳岛效应”与城市化对城郊农田植被生长期累积生物量的间接影响的相关关系具有显著的空间异质性。偏最小二乘法回归(PLS)、地理探测器、偏相关分析及最小二乘法(OLS)的统计结果均显示“热岛效应”对生长期农田植被累积生物量的间接促进作用随着纬度的增加更为显著。在纬度相对较高的东北平原区(NCP)，城郊的“热岛效应”是促使城郊农田植被生长期累积生物量(TINDVI)增加的主要诱因，其与城市化对城郊农田植被生长期累积生物量的间接影响(TINDVI IE)的偏相关关系高达0.72(P<0.01)。在黄淮海平原区(HHHP)，“热岛效应”与“碳岛效应”均对城郊农田植被生长期累积生物量呈现显著的促进作用。而在热量条件较好长江中下游平原区(MYP)，城郊气温升高与CO₂的富集可能对TINDVI IE造成一定的负面影响。

(6) 研究结果表明农田的景观配置及组分对于单位农田植被生产力具有重要影响，且两者的相对重要性呈纬度性分布。城郊农田斑块面积的增大、农田破碎度的降低及农田稳定性的增加对不同生长期阶段单位农田植被生物量的累积及农田植被生长状况均有着明显的正向影响，且这种正向效应随着纬度的增加而增强。在城市化率固定的情况下，增加斑块面积最有利于城郊单位农田生长期累积生物量的提升。此外，随着纬度的增加，农田景观配置对单位农田生长期累积生物量的重要性逐渐增加。特别是在中国三大农业区中相对纬度最高、气候较为干冷的东北平原区(NCP)，各通过了偏最小二乘法回归(PLS)重要性检验的景观配置因子（平均斑块面积(AREA-MN)、斑块面积(CA)、最大斑块指数(LPI)、核心区指数(TCA)）对单位农田TINDVI的解释程度高于农田组分因子(ISA)20.31%-30.27%，回归系数高于农田组分因子99.40%-189.84%。

Vegetation is one of the critical parts of the terrestrial ecosystem and a key link in the cycle of mass and energy, playing a role of the earth's natural "cooler", "purifier" and "protector". The global cropland vegetation coverage area accounts for about 11% of the total vegetation coverage area, and it is the basic unit for ensuring global food security. The growth and phenological characteristics of cropland vegetation and natural vegetation are significantly different due to the planting system. Besides, urbanization is regarded as the “natural laboratory” of global change. So exploring the response of suburban cropland vegetation growth characteristics to different urbanization pattern, investigating the spatiotemporal association between the changes in environmental factors caused by urbanization and variations of vegetation growth characteristics in the suburbs, quantifying the impact of changes in landscape patterns in the suburbs on cropland vegetation growth characteristics are critical for predicting future crop yields, optimizing field management and layout, and formulating urban agricultural development policies.

Firstly, based on comprehensive database of multi-source remote sensing data, surface observation data, and reanalysis data, the geo-intelligent machine learning algorithm is built to estimate the surface air temperature dataset on the scale of a kilometer grid. Then, this dissertation discusses the multiple impact characteristics of urbanization on the health status of suburban cropland vegetation and assesses the correlation between the variations of hydrothermal conditions and the changes of cropland vegetation health status in the suburban areas. Moreover, this dissertation studies the overall and indirect response of suburban cropland vegetation productivity to climate change “natural laboratories” under different key phenological periods and its spatiotemporal association with local surface air temperature and carbon dioxide (CO₂) variation considering cropping system. Furthermore, this dissertation quantifies the impact of urbanization-driven variations of cropland landscape pattern on cumulative biomass of cropland vegetation per unit cropland of growing period, and finds out the crucial landscape pattern factors under different climate background. The main results of this dissertation are as follows:

(1) The geo-intelligent model coupled with the adaptive spatio-temporal autocorrelation algorithm significantly improves the accuracy of surface air temperature prediction on a kilometer scale. In this study, we firstly propose a new surface air temperature reconstruction algorithm based on optimized multi-source heterogeneous datasets and three traditional machine learning models (Random Forest, Back Propagation Neural Network, Support Vector Machine) improved by coupling adaptive spatio-temporal autocorrelation algorithm. The unimproved models are recorded as Ori-RF, Ori-BPNN, Ori-SVM, and the improved models are recorded as Geoi-RF, Geoi-BPNN, Geoi-SVM, respectively. In addition, the monthly surface air temperature datasets at 1km scale of China from 2003 to 2012 are reconstructed based on the above models. Then, various statistical methods are utilized to compare the efficiency parameters of and its performance on the temporal and spatial scales of each model, and comprehensively and quantitatively assesses the accuracy and robustness of them from multiple dimensions.

The results show that Ori-RF fused with optimized multi-source data and coupled adaptive spatio-temporal autocorrelation algorithm can significantly improve the accuracy of surface air temperature prediction accuracy in the cloudy area (i.e. Sichuan) , the data-scarce areas (i.e. Tibet Plateau) and the urban areas. Moreover, Geoi-RF performs better compared with other models involved in this study, both overall and globally. The estimation accuracy, defined as the ratio of verification stations with absolute error less than 0.5K to total verification stations, of the monthly mean surface air temperature (maximum surface air temperature, minimum surface air temperature) of the Geoi-RF technique is 18.51-63.17% (15.57%-44.62%, 16.53%-45.55%) higher than that by the other techniques considered in this study. In addition, it is interesting to note that the Geoi-RF performs better with Goodness of fit (R²) close to 1, Mean Absolute Error (MAE) lower than 0.25K, and Root Mean Squared Error (RMSE) lower than 0.5K respectively. In general, the Geoi-RF technique is an ideal algorithm to estimate the surface air temperature in a large range of fine scales.

(2) The footprint and indirect impact of urbanization on the health status of suburban cropland vegetation have significant spatial stratified heterogeneity in a changing environment. Buffer analysis and mixed pixel decomposition are used to quantify the overall and indirect impact of urbanization on suburban cropland vegetation of 32 major cities in nine agricultural regions of China under different climate backgrounds. The results show that the overall impact footprint of urbanization on health status of suburban cropland vegetation in most of agricultural regions significantly exponential decrease along the urban-rural gradient, and the radius of impact footprint can reach more than 26km in agricultural areas with rapid urbanization. The indirect impact of urbanization on the health status of suburban cropland vegetation in China is generally “south positive and north negative” in terms of spatial pattern. Namely, the health status of cropland vegetation in the suburbs in the agricultural areas of northern China improves with the increase of urban intensity under the indirect impact of urbanization. This indirect impact of urbanization approximately offsets about 44% of the health status loss of cropland vegetation in the suburbs caused by the expansion of the urban impervious surface in these regions. However, in the southern agricultural areas with relatively warm and humid climate, the indirect impact of urbanization adversely affects the cropland vegetation in the suburbs. In the northern agricultural area, i.e., Huang-Huai-Hai Plain (HHHP) and the North China Plain (NCP), the indirect effect of urbanization on the health status of cropland vegetation in the suburbs is positively correlated with the increase of surface air temperature in the suburbs. And in the Middle-Lower Yangtze Plain (MYP) belonging to southern agricultural area, the variation of soil moisture in the suburbs is significantly positively related to the indirect impact of urbanization on the health status of suburban cropland vegetation.

(3) In a changing environment, the indirect effects of urbanization generally have a significant promotion effect on cumulative biomass of suburban cropland vegetation in different key phenological periods, but it will be significantly distinctive under different climate backgrounds. The decline in cultivated area triggered by impervious surface expansion (direct impact) brings about the result that the accumulated biomass during the growing period (TINDVI), the accumulative biomass during the pre-anthesis growth period (TINDVIBeforeMax), and the accumulative biomass after flowering (TINDVIAfterMax) of cropland vegetation per unit of cropland decreases with the increase of urbanization intensity (ISA). However, the indirect impact of urbanization enhances the TINDVI (TINDVIBeforeMax/TINDVIAfterMax) at 78% (80%/76%), 92% (95%/64%), 98% (82%/96%) of the places along the intensity gradient in MYP, HHHP, NCP, respectively, and the relative enhancement and volatility dramatically increases with urban intensity. Besides, the indirect effects of urbanization on the cumulative biomass of cropland vegetation in different growing periods are all around the 50% impermeable percentage in the three major agricultural areas of China.

(4) In the three major agricultural areas of China, the indirect impact of urbanization on suburban cropland vegetation can offset part of the TINDVI (TINDVIBeforeMax/ TINDVIAfterMax ) loss caused by the direct impact of urbanization. In the MYP, HHHP and NCP, the indirect impact of urbanization can offset approximately 17% (33%/19%), 39% (49%/12%), 40% (30%/46%) of the TINDVI (TINDVIBeforeMax/ TINDVIAfterMax) loss caused by impervious surface expansion, respectively. In addition, the indirect effect (IE) of urbanization on TINDVIAfterMax is more significant than TINDVIBeforeMax in the NCP. Besides, in the HHHP, the indirect effect of urbanization on TINDVIBeforeMax is more critical than TINDVIAfterMax. And the indirect effect of urbanization on TINDVIBeforeMax is approximately equivalent to that of TINDVIAfterMax in the MYP.

(5) The correlation between the effect of “heat island” and “carbon island” in the suburbs and the indirect impact of urbanization on the cumulative biomass of suburban cropland vegetation during the growing period has significant spatial heterogeneity under different climate backgrounds. The results from partial least squares regression (PLS), geographic detectors, partial correlation analysis and least squares regression (OLS) all show that the indirect promotion of "heat island effect" on the cumulative biomass of cropland vegetation in the growing period (TINDVI IE) increases with increasing latitude. In the NCP with relatively high latitudes, the “heat island effect” in the suburbs is the main cause of the increase in TINDVI IE, and the partial correlation between them reaches to 0.72 (P<0.01). In the HHHP, both the heat island effect and the carbon island effect significantly promote the increment of TINDVI IE. However, in the warmer MYP, the increase of mean air temperature and the enrichment of carbon dioxide (CO₂) in the suburbs may have a slightly negative effect on TINDVI IE.

(6) The landscape configuration and composition of cropland has great influence on unit cropland vegetation productivity, and their relative importance varies with latitude. The increase of cropland patch area in the suburbs, the reduction of cropland fragmentation and the increment of cropland stability all pose a significant positive effect on the accumulation of cropland biomass and cropland vegetation growth status at different stages of the growth period and this effect increases with increasing latitude. Given a certain urbanization rate, increment of the cropland patch area is most effective way to increase the biomass of unit cropland in the growing period. In addition, as the latitude increases, the importance of cropland landscape configuration to the TINDVI gradually increases, especially in the NCP with the relatively highest latitude and dry climate among the three major agricultural areas of China. The explanatory power of landscape configuration factors (AREA-MN, CA, LPI, TCA) that passed the partial least squares (PLS) importance test to TINDVI of unit cropland is higher than the landscape composition factors (ISA) 20.31%-30.27%, and the regression coefficient of it is higher than ISA 99.40%-189.84%.