Design of adaptive spatiotemporal information and control systems

当今遥感系统数据收集、处理和联网能力的急剧提高，使遥感坚定地走上了成为一项普及技术的道路，以各种方式影响着社会。现在迫切需要的是强大的目标检测和场景解释方法，能够自动处理和解释这些丰富的信息。该项目以学习控制和模式识别/图像处理的协同作用为基础，提出了主要在环境应用中遇到的复杂、动态时空系统的建模、监测和控制的创新技术。总体目标是开发一种综合设计方法，解决与自适应处理大量多维、不同种类传感器数据有关的问题。更具体的目标包括设计和优化生物启发的注意视觉算法的理论和方法基础，以解决异质稀疏遥感图像中时空显著的问题，以及开发新的迭代学习方法来解决病态目标检测和分类问题。方法论办法是在缺乏基本事实参考信息的情况下确定的，这些信息可能很难获得，成本也很高。它包括开发混合技术，将基于核的方法和相关的机器学习范例(确定性学习和迭代学习)相结合，用于时变、不确定的系统监测和控制应用。这一跨学科项目将帮助培养各级研究人员，其培训计划将因与行业合作伙伴的合作带来的机会而显著增强。除了对新的科学知识的贡献外，这项研究的结果将直接造福加拿大社会，有助于满足对尖端信息技术的需求，以有效地监测和管理加拿大广阔领土上的自然资源。

Steep increases in the data gathering, processing and networking capabilities of today's remote sensing systems have put remote sensing firmly on the path to become a pervasive technology, affecting society in a variety of ways. What is now critically needed are powerful object detection and scene interpretation methods amenable to automated processing and interpretation of this wealth of information. This project builds on the synergy of learning control and pattern recognition/image processing to propose innovative techniques for modeling, monitoring, and control of complex, dynamic spatiotemporal systems as encountered mainly in environmental applications. The overall objective is to develop an integrated design methodology that addresses issues relating to the adaptive processing of large quantities of multi-dimensional, heterogeneous sensor data. More specific objectives include research on theoretical and methodological foundations for the design and optimization of biologically-inspired attentive vision algorithms for spatiotemporal saliency in heterogeneous, sparse remote sensing imagery, and the development of novel iterative learning methods to solve ill-posed object detection and classification problems. The methodological approach has been defined in the context of the scarcity of ground truth reference information, which can be difficult and costly to acquire. It involves the development of hybrid technologies integrating kernel-based methods and pertinent machine learning paradigms (deterministic learning and iterative learning) for time-variable, uncertain system monitoring and control applications. This interdisciplinary project will help develop research personnel at all levels, with its training program significantly enhanced by the opportunities arising from the collaboration with industrial partners. Apart from the contribution to new scientific knowledge, the results of this research will directly benefit Canadian society, helping to meet the need for sophisticated information technologies to efficiently monitor and manage natural resources in Canada's vast territory.