Smart deep learning by incorporating remote sensing domain knowledge in vegetation characterization

该研究计划的目标是开发创新方法，以利用地球观测和人工智能(特别是深度学习)之间的协同作用。它的动机是对调查和评估植被冠层的先进科学和创新的需求。这是由从空间、卫星或飞机(遥感)观测地球表面的技术发展推动的，也是由深度学习的发展所推动的。在过去的几年里，遥感领域的深度学习取得了巨大的增长。尽管已经取得了一些进展，但遥感领域的深度学习研究仍处于初级阶段。它主要侧重于应用和微调现有网络，深度学习方法目前只是数据驱动的，没有任何明确表达的该领域现有知识。我们没有必要学习我们已经知道的东西。为了充分发挥深度学习在给遥感数据分析带来革命性变化方面的潜力，需要对深度学习和遥感之间的深度融合进行研究。到目前为止，深度学习已被用于对土地覆盖的广泛类别进行分类，但很少用于确定森林树冠的特征和确定树种，这在地球观测中是重要但具有挑战性的。这项研究的目标是开发创新的方法，将遥感领域的知识和深度学习结合起来，以应对三个相关领域的挑战：1)单个树冠的划定，2)单个树种的分类，以及3)植被冠层生物物理参数的提取。此外，我们将利用传统的机器学习方法和先验知识来开发算法，以解决与训练数据有限、在设计深度学习网络和学习过程中纳入先验知识以及有效利用多源遥感数据有关的问题。作为这一研究计划的成果，我们将通过弥合深度学习和遥感之间的差距，确保加拿大保持在人工智能革命的前沿。研究成果将加深我们对深度学习在推进科学和工业应用方面的理解。通过提供高效和有效的方法来跟踪森林的物种、功能状态和生产力，通过拟议的研究，我们将帮助加拿大更好地管理其最大的自然资源。从野生动物栖息地测绘到生物燃料生产，再到森林防火，实际应用是重要和广泛的。最后，在未来五年内，包括4名博士、4名硕士和4名本科生在内的总部基地将接受培训，以获得在学术界或行业脱颖而出的高级技能，利用他们的专业知识增强加拿大在人工智能和地理空间技术方面的智囊团信任。

The goal of this research program is to develop innovative methods to exploit the synergy between Earth observation and artificial intelligence (specifically deep learning). It is motivated by the demand for advancing science and innovation in the inventory and assessment of vegetation canopies. This is driven by technological developments in observing the Earth's surface from space, satellites or aircraft (remote sensing), and enabled by the development of deep learning. The past several years have witnessed a massive growth of deep learning in remote sensing. Even though some progress has been made, deep learning research in remote sensing is still in its infancy. It has mainly focused on applying and fine-tuning existing networks, and deep learning approaches are currently only data-driven and without any explicitly expressed existing knowledge in the domain. It is not necessary to learn what we have already known. To fully realize the potential generated by deep learning in revolutionizing remotely sensed data analysis, research is needed for in-depth integration between deep learning and remote sensing. To date, deep learning has been used to classify broad categories of land cover, but rarely for characterizing forest canopies and identifying tree species, which is important but challenging in Earth observation. The objective of this research program is to develop innovative approaches to integrate remote sensing domain knowledge and deep learning to address challenges in three related areas: 1) individual tree crown delineation, 2) individual tree species classification, and 3) the retrieval of biophysical parameters of vegetation canopies. Moreover, we will develop algorithms by exploiting the use of traditional machine learning methods and prior knowledge to solve the issues related to limited training data, the incorporation of the prior knowledge in the design the deep learning network and in the learning process, and the effective utilization of multi-source remotely sensed data. As the outcome of this research program we will ensure that Canada remains at the forefront of the artificial intelligence revolution by bridging the gap between deep learning and remote sensing. The research results will deepen our understanding of deep learning in advancing scientific and industrial applications. By providing efficient and effective ways to keep track of the species, functional status and productivity of forests, from proposed research we will help Canada be a better steward of its greatest natural resource. The practical applications are significant and wide ranging, from wildlife habitat mapping, to biofuel production, to forest fire prevention. Finally, over the next five years, HQP including 4 PhD, 4 MSc and 4 undergraduate students, will be trained to gain advanced skills to excel in either academia or industry, using their expertise to enhance Canada's brain trust in artificial intelligence and geospatial technologies.