arXiv 2605.199312026-05-19cs.CVcs.AIcs.LG

StruMPL：在不相交的部分监督和MNAR标签下的多任务密集回归

StruMPL: Multi-task Dense Regression under Disjoint Partial Supervision and MNAR Labels

School of Geosciences, University of Edinburgh, UK（爱丁堡大学地球科学学院，英国）；National Centre for Earth Observation (NCEO), UK（英国地球观测国家中心）；Department of Spatial Sciences, Faculty of Environmental Sciences Czech University of Life Sciences Prague, Praha, Czech Republic（环境科学学院空间科学系，捷克布拉格生命科学大学）

Reza M. Asiyabi, Juan Alberto Molina-Valero, The SEOSAW Partnership, Steven Hancock, Casey M. Ryan

备注：: 10 pages with 3 figures and 4 tables, References and Appendix 12 pages with 1 figure and 4 tables

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文针对在不相交的部分监督和MNAR标签下的多任务密集回归问题，提出StruMPL方法，通过共享编码器和可学习的物理模块，结合Augmented IPW损失函数，提高了对森林地上生物量的估计精度。

摘要：

从地球观测估计森林地上生物量（AGB）结合了两个结构上不兼容的标签源：空间borne激光雷达在数百万个位置提供冠层结构但没有生物量估计，而地面样地在数千个偏倚位置提供生物量但没有结构指标。没有单个训练样本携带所有目标变量的标签，样地标签不是随机缺失（MNAR），且生物量通过已知但生物体特异性的所有学定律与结构变量相关联。我们将其正式化为在异质不相交部分监督下的多任务密集回归问题，具有MNAR标签和任务间物理约束，并提出StruMPL方法来联合解决。一个共享编码器为每个变量回归、填补和倾向性头提供空间MNAR校正，以及一个可学习的物理模块，该模块在每个像素上评估任务间约束对模型自身预测的影响。监督损失使用Augmented IPW（AIPW）伪结果，其中在倾向性和填补基线上的停止梯度；我们证明了分析和实证上，两者对于联合优化恢复IPW加权的平稳点并保持损失有界是必要的。在两个生态上不同的生物体上，StruMPL在AGB RMSE和偏倚方面优于消融变体和最接近的已发表方法，分层分析显示AIPW减少了高AGB偏倚约54%。

英文摘要：

Estimating forest aboveground biomass (AGB) from Earth observation combines two structurally incompatible label sources: spaceborne lidar provides canopy structure at millions of locations but no biomass estimate, and ground-based plots provide biomass at thousands of biased locations but no metrics of structure. No single training sample carries labels for all target variables, plot labels are missing not at random (MNAR), and biomass is linked to the structural variables by known but biome-specific allometric laws. We formalise this as multi-task dense regression under heterogeneous disjoint partial supervision with MNAR labels and inter-task physical constraints, and propose StruMPL to address it jointly. A shared encoder feeds per-variable regression, imputation, and propensity heads for spatial MNAR correction, and a learnable physics module that evaluates the inter-task constraint on the model's own predictions at every pixel. The supervised loss uses an Augmented IPW (AIPW) pseudo-outcome with stop-gradients on the propensity and on the imputation baseline; we show analytically and empirically that both are necessary for joint optimisation to recover IPW-weighted stationary points while keeping the loss bounded. On two ecologically distinct biomes, StruMPL outperforms ablation variants and the closest published method on AGB RMSE and bias, with a stratified analysis showing AIPW reduces high-AGB bias by ~54%.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。