Machine Learning for improved understanding of L-A processes and feedbacks

深度学习（DL）是地球系统科学中发展非常迅速的一个领域。然而，深度学习系统通常是黑盒子，对过程理解几乎没有贡献。该项目将开发合适的深度学习和其他机器学习技术，以提高对陆地-大气（L-A）反馈过程的理解，并得出优于当前方法（如Monin-Obukhov相似理论（MOST））的定量关系。专家的跨学科合作和前所未有的数据收集将促进这一努力。我们将追求数据驱动，但结构知情的深度学习和其他机器学习方法，目标是(1)利用新一代高分辨率卫星数据确定地表温度（LSTs）和冠层湿度；(2)地表能量平衡的封闭；(3)复杂地形的地表通量关系以及对流边界层的通量梯度关系，包括超出MOST和Bulk Richardson Number （BRN）方法的夹带；(4)建立了第一个L-A反馈基础模型，用于分析相空间结构和后续任务。我们将通过四种核心DL技术来研究和推进这四个主题：(A)近似空间细颗粒L-A性质的超分辨率方法；(B)物理信息方程参数化和近似，以发现和改进L-A反馈指标；(C)在深度学习方法中包含L-A系统信息归纳偏差，以加速学习和提高泛化；(D)应用自监督和对比预训练开发L-A状态表征基础模型。所有的技术都将考虑L-A系统变量的混沌行为、不确定性和有限的可预测性。为了进一步促进对L-A系统的理解，我们将进行重要性加权并生成机器学习模型层次。因此，我们期望确定基本的和部分新颖的数据依赖关系、过程方程、过程参数化和计算单元，这些对于在大气变量之间产生准确和可推广的关系至关重要。我们将与P2合作，分析卫星频道的预测潜力。与P1， P3， P5和P6一起，我们将推导表面通量和夹带通量关系，推导能量平衡（EB）闭合项和分配蒸散发。此外，我们将领导CCWG-DL，在分析其他数据源和关系方面提供教育和咨询，包括表面粗糙度（P5）和混合高度（P1, P7）的影响。这些努力将为O1、O2、O3、O4和OS做出贡献。

Deep learning (DL) is an extremely rapidly developing field in Earth system science. However, DL systems are often black boxes, hardly contributing to process understanding. This project will develop suitably tailored DL and other machine learning techniques both to improve land-atmospheric (L-A) feedback process understanding and to derive quantitative relationships that are superior to current approaches, such as Monin-Obukhov similarity theory (MOST). The interdisciplinary collaboration of experts and unprecedented collection of data will facilitate this endeavour. We will pursue data-driven, but structurally-informed DL and other machine learning approaches, targeting (1) the determination of land-surface temperatures (LSTs) and canopy moisture with the available new generation of high-resolution satellite data; (2) the closure of the surface energy balance; (3) surface flux relationships in complex terrain as well as flux-gradient relationships in the convective boundary layer including entrainment going beyond MOST and Bulk Richardson Number (BRN) approaches; and (4) the development of a first L-A feedback foundation model to analyze the phase space structure and for downstream tasks. We will study and advance these four topics by means of four core DL techniques: (A) super-resolution approaches to approximate spatially fine-granular L-A properties; (B) physics-informed equation parameterization and approximation to discover and improve L-A feedback metrics; (C) the inclusion of L-A system-informed inductive biases in DL approaches to speed-up learning and improve generalization; (D) the application of self-supervised and contrastive pretraining for the development of an L-A state characterizing foundation model. All techniques will consider the chaotic behaviour, the uncertainty, and the limited predictability of L-A system variables. To further foster L-A system understanding, we will conduct importance weighting and generate machine learning model hierarchies. We expect to thus identify fundamental and partially novel data dependencies, process equations, process parameterizations, and computational units that are critical for the generation of accurate and generalizable relationships between atmospheric variables. In collaboration with P2, we will analyse the predictive potential of satellite channels. Together with P1, P3, P5, and P6, we will derive surface and entrainment flux relationships, derive energy balance (EB) closure terms, and partition evapotranspiration. Moreover, we will lead CCWG-DL, offering education and consultation in analysing other data sources and relationships including influences of the surface roughness (P5) and the blending height (P1, P7). These efforts will contribute to O1, O2, O3, O4, and OS.