【作者】 王欣；

【导师】 杜培军；

【作者基本信息】 南京大学 ， 地图学与地理信息系统， 2021， 博士

【摘要】 变化检测是通过对同一地物或自然现象在不同时间的观测确定其状态变化的过程。近年来,随着人类活动的增加,陆地表面发生了显著变化。遥感对地观测技术能够在短时间内获取地表大面积的土地覆盖信息,为大范围信息的周期性采集和动态监测提供了有效途径,广泛应用于探索由自然因素(火灾、地震和蝗灾等)和人类活动(森林砍伐、耕作与种植和城市化等)引起的地表变化。因此,有效地发现和描述多时相遥感影像中的变化是遥感科学技术的研究热点和前沿问题。随着存档数据的积累以及更为先进传感器的出现,遥感影像能够提供细节更加丰富的变化信息。然而,基于这些影像的变化检测技术方法与应用研究仍然存在许多难题。首先,当前单一层次的影像特征提取与分析方法难以全面地描述地物空间信息的变化,变化检测中没有充分考虑地表变化的空间规律;其次,多时相影像中时间维信息的发掘和利用尚有不足;第三,对于时间序列变化检测任务,基于现有时序模型的检测结果可靠性较差,在复杂地表变化场景中的稳定性有待提升;第四,目前时间序列变化检测中地表要素时空特征和变化规律的可解释性不强,对复杂环境下多要素变化之间的相互作用和影响缺乏深入的探讨。为了克服上述问题,论文根据不同类型遥感数据集的特点,在地表变化空间规律和时间规律的引导下,发掘了影像不同层次和尺度的空间特征和时间特征,提出了顾及空间规律的双时相变化检测方法、顾及时间规律的三时相变化检测方法和顾及多要素作用规律的时间序列变化检测与分析方法。论文主要研究工作与结论如下:(1)针对地表构成和变化在空间分布上具有相似性、异质性和多尺度等特点,提出了联合多层次空间信息的变化检测方法。真实地物的空间规律在遥感影像中可以通过不同层次的空间特征表达。组合包括像元级、邻域级、对象级和场景级空间特征,从多个角度全面地描述遥感影像间的变化信息。针对多层次特征的相关性大、复杂变化类型的可分性低、训练样本的需求量大等问题,采用特征选择的方式选取最有效的特征组合降低多层次特征的维数,减少信息冗余,利用迭代主动学习的策略逐步学习有价值的新样本,不断优化特征提取模型和分类器模型。结果表明,联合多层次空间特征的变化检测方法能够充分考虑地表变化的空间规律,显著提升检测结果的精度和效率,同时有效地降低检测成本。(2)探索了地表变化在多时相影像间的逻辑关系,提出了顾及时间维逻辑规律的变化检测方法。根据地表变化在三期影像间的逻辑关系,利用双时相变化检测结果之间相互验证,构建三时相地表变化模式的逻辑体系。通过对模式类型的推理判别错误检测,利用双时相检测结果的可靠性对比修正错误结果,实现了利用时间维逻辑关系弥补光谱和空间信息的不足,提升了对地表变化过程的理解。结果表明,顾及时间维逻辑关系的变化检测方法能够充分利用影像间的时间规律,修正双时相变化检测中的异常结果,有效地提升变化检测结果的精度和可靠性。(3)提出了基于集成学习策略的时序变化分析方法,实现了时间序列影像中土地覆盖变化与植被物候响应的多要素作用分析。单一分类器对不稳定时序模型的判别能力有限,结果可靠性较低。将不同基分类器集成能够充分发挥每个分类器的优势,提高对时间序列土地覆盖变化的判别性能。同时,针对目前长时间范围植被物候分析中对土地覆盖变化考虑欠缺的问题,结合连续土地覆盖变化结果,对不同区域的植被物候进行精确地时空分析。结果表明,基于多分类器集成的时序变化检测方法显著提升了连续土地覆盖变化检测的准确性。同时,在长时间土地覆盖变化的分析中顾及了地表要素间的相互作用,分析了城市化作用下植被物候时空分异和变化规律的响应结果,揭示了土地覆盖的动态变化对植被物候的直接和间接影响。更多还原

【Abstract】 Change detection is a process of determining changes of the same object or natural phenomenon through observations at different time.In recent years,with the increase of human activities,the land surface has undergone significant changes.Remote sensing earth observation technology can obtain land cover information of a large region in a short time,providing an effective way for periodic collection and dynamic monitoring of large area information.It has been widely used in exploring land surface changes caused by natural factors(e.g.,fires,earthquakes,locust plagues.)and human activities(e.g.,deforestation,farming and planting,urbanization.).Therefore,effectively discovering and describing the changes in multi-temporal remote sensing images is a research hotspot and frontier issue in remote sensing science and technology.With the accumulation of archived data and the emergence of more advanced sensors,remote sensing images can provide more detailed change information.However,there are still many problems in the methodology and application of change detection based on these images.First,the current single-level feature extraction and analysis methods cannot fully characterize the change of spatial information in images.The spatial patterns of land surface change is not fully considered in change detection process.Secondly,the exploration and utilization of temporal information among multitemporal images is obviously insufficient.Third,for time series change detection tasks,the reliability of the results based on existing time series model is poor,whose stability in complex land surface change scenarios needs to be improved.Fourth,the interpretability in spatio-temporal characteristics and change patterns of land surface elements in the current time series change detection needs to be enhanced.There is a lack of in-depth discussion on the interaction and influence of multi-element changes in a complex environment.In order to overcome the aforementioned problems,according to the characteristics of different types of remote sensing datasets,the spatial and temporal image features at different levels and scales were excavated under the guidance of the spatial and temporal patterns of land surface changes.Three change detection approaches were proposed in the thesis,including bi-temporal change detection approach with consideration of spatial pattern,tri-temporal change detection approach with consideration of temporal pattern,and time series change detection and analysis approach with consideration of the pattern of multi-element interaction.The main research work and conclusions of the thesis are as follows:(1)According to the characteristics of similarity,heterogeneity and multiscale in the spatial distribution of land surface composition and change,a change detection approach based on multi-level spatial information integration was proposed.The spatial patterns of real ground objects can be expressed in different levels of spatial features in remote sensing images.The change information between remote sensing images can be comprehensively described from different angles by combining pixel-level,neighborhood-level,object-level and scene-level spatial features.However,the problems such as high correlation of multi-level features,low separability of complex change types,and large demand for training samples arose.Therefore,feature selection was carried out to find the most effective feature combination to reduce the feature dimension and information redundancy.The iterative active learning strategy was also adopted to search for new valuable training samples step by step,which continuously optimized the models of feature extraction and change detector.The results showed that the change detection method combined with multilevel spatial features fully considered the spatial patterns of land surface changes.It significantly improved the accuracy and efficiency of change detection results and effectively reduced change detection cost.(2)The logical relationship between land surface changes in multi-temporal images was explored,and a change detection approach with consideration of the logical rules in temporal space was proposed.According to the logical relationship among land surface changes in tri-temporal images,bi-temporal change detection results can be mutually verified,and a logical system of tri-temporal land surface change pattern was constructed.The incorrect patterns were identified by reasoning the type of pattern,and the anomalous results were modified through the reliability comparison among bi-temporal change detection results.The logical relationship in temporal space was utilized to make up for the insufficiency of spectral and spatial information,improving the understanding of the process of land surface change.The results showed that the change detection approach that took the logical relationship of the land surface changes into account made full use of the temporal patterns among images to modify the anomalous outcomes from original bi-temporal change detection,effectively improving the accuracy and reliability of change detection results.(3)A time series change analysis approach based on ensemble learning strategy was proposed,realizing the multi-element interaction analysis of land cover change and vegetation phenology response in time series images.The ability of a single classifier in discriminating unstable time series model was limited,leading to low reliability of the results.Integrating different base classifiers gave full play to the advantages of each classifier,improving the performance of discriminating time series land cover changes.Meanwhile,in view of the lack of consideration of land cover changes in the current long-term vegetation phenology analysis,accurate spatiotemporal analysis of vegetation phenology in different regions was realized by integrating the continuous land cover change results.The results showed that the timeseries change detection approach based on the integration of multiple classifiers significantly improved the accuracy of the continuous land cover change detection.Moreover,the interaction among land surface elements was taken into account in the analysis of long-term land cover changes.The response results of the spatio-temporal differentiation and change patterns of vegetation phenology under the condition of urbanization were analyzed.The direct and indirect influences of land cover dynamic changes on vegetation phenology were revealed.更多还原