地球同步气象卫星多源遥感图像三维球面拼接方法研究

Since the earth is non-perfect sphere and GSMS images are high spatial resolution, GSMS remote sensing image stitching on spherical surface is different with the traditional method of image stitching on plane in theoretical model, computation complexity and resolution. .To solve these problems and realize engineering applications, a novel method of 3D spherical stitching of geo-stationary meteorological satellite multi-source remote sensing image is proposed in this project. This new method is put forward based on several ideas:.1..A nonlinear model is proposed to describe the multi-view stereo geometry transformation between 2D image and 3D sphere. However, it is an underdetermined problem to solve model parameters when the sensor’s inner and outer parameters are unknown. To estimate sensor parameters, we combine deep convolution neural network (CNN) with topology statistical characteristics of edges..2..Invariant and unique topology similarity measure parameters and edge topology structure coding method will be studied for global matching which are used to reduce feature mismatching and leakage matching. .3..Focusing on 3D spherical registration of image, a CNN-based method for GSHHG landmark generation and matching will be studied..4..To improve spatial resolution, landmark recovery method is put forward based on codec of landmark topology structure. Gradient distribution consistency of Poisson equation and edge distribution based on forest structure will be studied to improve the global spatial resolution. .Our main contributions are: a theoretical framework of GSMS remote sensing image 3D spherical stitching for non-standard spherical object; a solution for underdetermined 3D sphere registration model based on deep learning and prior knowledge of GSHHG landmark model; achieving high resolution spherical mosaic based on topology and gradient distribution.

非标准球面高像素高分辨率GSMS多源遥感图像三维球面拼接，在理论模型、计算量、精度等方面向传统二维拼接提出挑战。申请课题提出GSMS多源遥感图像三维球面拼接新方法：提出二维图像与三维球面间多视点立体几何非线性模型，针对传感器内外参数未知时该模型参数求解的欠定问题，提出深度卷积神经网络估计模型参数，针对非标准球面提出多模融合的全局配准方法，从而建构非标准球面三维球面拼接理论框架；提出兼具不变性及唯一性的边缘特征拓扑结构编码及相似性度量方法，抑制特征误匹配及漏匹配；研究基于深度卷积神经网络的GSHHG地标图生成方法，提出利用地标先验知识的匹配方法，实现三维球面配准；提出基于地标拓扑稳定性的地标拓扑结构编解码地标恢复方法，及基于梯度分布一致性的泊松方程及结构森林分辨率提高方法，实现均匀高分辨率三维球面拼接。项目提出的GSMS图像三维球面拼接理论框架及关键技术对下一代GSMS产品研发具有积极意义。

目前的图像拼接方法多要求场景近似位于同一平面，而地球同步气象卫星(GSMS)遥感图像的拍摄目标为三维球体，直接使用二维拼接模型将导致严重误差。为此本项目深入探讨了GSMS多源遥感图像三维球面拼接方法，将二维数据投影到三维球面上并确定每个像素点的经纬度信息。.针对非标准球面高像素高分辨率GSMS多源遥感图像三维球面拼接在理论模型、计算量、精度等方面的挑战，研究了GSMS多源遥感图像三维球面拼接新方法，主要包括：研究了二维图像与三维球面间多视点立体几何非线性模型，建构非标准球面三维球面拼接理论框架；提出了基于深度神经网络优化的局部描述子学习方法，针对局部描述子学习的不足，提出了难样本三元组采样策略和改进的Triplet损失函数，通过在公开数据集上实验证明，最终生成的局部描述子DHNet具有增强的类内聚合性和类间分离性；提出了基于多尺度金字塔模型、几何拓扑以及拓扑拟合的三种特征匹配配算法，基于斜率约束的误匹配校正法，基于邻域空间一致性，分别从特征初匹配、特征误匹配删除和特征漏匹配补全这三个层面实现特征匹配，提升了配准精确度为96.51%的同时，将召回率提升至90.89%，速度相比同类方法有所提升（GTM算法和RSOC算法的时间复杂度都为O(N的三次方），WGTM算法的复杂度为O（N的平方乘以log（N）），KNN-TAR算法、ISSC算法、RFVTM算法与本项目算法的时间复杂度为O（N的平方））；研究了基于深度卷积神经网络的GSHHG地标图生成方法，通过建立多项式拟合映射模型完成图像间的全局配准，得到每个点的经纬度信息，针对传统最佳缝合线生成算法容易产生结构误差的问题，提出基于结构最优的三维缝合线方法，实现均匀高分辨率三维球面拼接。.项目研究的关键技术在中国气象卫星数据库中风云二号气象卫星的图像上进行实验，实验验证了这些方法的有效性，对下一代GSMS产品研发具有参考价值。