遥感影像语义分割知识迁移的元学习方法

In recent years, with the promotion of deep learning, large-scale application of big data and graphics processor unit (GPU), intelligent interpretation of remote sensing images faces many opportunities and challenges. The robustness of automatic interpretation and semantics segmentation of remote sensing images is still difficult to meet the requirements of practical application, such as lack of training data, insufficient generalization ability of models, and visual ambiguity of learning objectives. To solve the problems at pixel level, target level and scene level, we proposes a meta-learning method for knowledge transfer of semantics segmentation of remote sensing images. By constructing meta-learning units for knowledge transferring at different levels of "sample-pixel-target-scene", the model has the ability of learning and inferring at different levels under the condition of imbalanced and inadequate training samples. In this way, the problem of insufficient recognition robustness caused by sample distribution, spectral difference, directional scale change and scene context superposition in the task of semantics segmentation of remote sensing image is eliminated. The outcomes of this project can not only provide a new calculation and analysis framework for automatic remote sensing images interpretation, but also be effectively applied to remote sensing monitoring of land and resources related tasks, which has important research and application value.

近年来，随着深度学习、大数据和图形处理器（GPU）大规模应用的推动，遥感影像智能解译面临诸多机遇和挑战。遥感影像的自动解译和语义分割在鲁棒性方面仍然难以达到实际应用的需求，主要问题有：训练数据缺乏、模型泛化能力不足、学习目标具有视觉歧义性等。本项目针对遥感影像像素级、目标级、场景级语义分割存在的问题，提出遥感影像语义分割知识迁移的元学习方法，通过构建“样本-像素-目标-场景”不同层级的知识迁移元学习单元，使模型在训练样本不均衡、不充分条件下，具备不同层次的学习和推估能力，以消除遥感影像语义分割任务中，因样本分布、光谱差异、方向尺度变化，以及场景上下文叠加造成的识别鲁棒性不足等问题。本项目不仅能为遥感影像自动解译提供新的计算和分析框架，研究成果还能有效应用于国土资源遥感监测等任务中，具有重要的研究和实际应用价值。

近年来，随着深度学习、大数据和图形处理器（GPU）大规模应用的推动，遥感影像智能解译面临诸多机遇和挑战。然而，遥感影像的语义分割在鲁棒性方面仍然难以达到实际应用的需求。本项目针对遥感影像像素级、目标级、场景级语义分割存在的问题，提出语义分割知识迁移的元学习方法，通过构建“样本-像素-目标-场景”不同层级的知识迁移元学习单元，使模型具备不同层次的学习和推估能力，以消除遥感影像语义分割任务中，因样本分布、光谱差异、方向尺度变化，以及场景上下文叠加造成的识别鲁棒性不足等问题。项目构建了多级别的遥感影像元学习样本库平台LuoJiaSET，同时搭建了“像素-目标-场景”多层级的遥感专用深度学习框架及元学习模型库平台LuoJiaNET，成果荣获2020年地理信息科技进步特等奖、2021年湖北省人工智能重大创新成果奖等奖励。项目从样本到框架模型，建立了一套全栈国产化遥感影像自动解译新计算和分析框架，有效支撑了国土资源遥感监测等任务。