CN-121982557-A - Port and dock identification method and system based on high-resolution optical satellite image

CN121982557ACN 121982557 ACN121982557 ACN 121982557ACN-121982557-A

Abstract

The application discloses a port and dock identification method and system based on high-resolution optical satellite images, comprising the following steps of S1, preprocessing input high-resolution optical satellite images, S2, inputting the preprocessed images into a preset deep learning network model, outputting a segmentation probability map with continuous structure and fine boundary, wherein the preset deep learning network comprises a double-branch feature extraction network, a multi-scale feature fusion network and a structure restoration network which are sequentially connected, and S3, post-processing the segmentation probability map with continuous structure and fine boundary to obtain dock vector polygon data with topological relation, so that port and dock identification is completed. Compared with the prior art, the method solves the problems that the prior model mainly depends on image spectrum and texture information, and is insufficient in utilization of space semantics and context information and easy to confuse.

Inventors

DUAN HONGSHAN
ZHAO JUN
SUN SHAOJIE
WANG JIALIN
QU ZIHONG
LIN XIAOBO
Qilin Spring Batch
HUANG GUANXIAN
SHI XIAOCHUN
Xu Gengran

Assignees

广东省国土资源测绘院
中山大学

Dates

Publication Date: 20260505
Application Date: 20251209

Claims (10)

1. A port and dock identification method based on high-resolution optical satellite images is characterized by comprising the following steps: S1, preprocessing data of an input high-resolution optical satellite image; S2, inputting the preprocessed image into a preset deep learning network model, and outputting a segmentation probability map with continuous structure and fine boundary, wherein the preset deep learning network comprises a double-branch feature extraction network, a multi-scale feature fusion network and a structure restoration network which are connected in sequence; and S3, carrying out post-processing on the segmentation probability map with continuous structure and fine boundary to obtain wharf vector polygonal data with topological relation, and completing the recognition of the wharf and the port.
2. The port and dock identification method based on the high-resolution optical satellite image according to claim 1, wherein the dual-branch feature extraction network adopts a parallel architecture of a convolutional neural network branch and a visual state space model branch and is used for cooperatively extracting local detail features and global context information of the high-resolution optical satellite image, the multi-scale feature fusion network receives the local detail features and the global context information of the high-resolution optical satellite image, aggregates the local detail features and the global context information into multi-level fusion features and inputs the multi-level fusion features into a structure restoration network, and the structure restoration network comprises a dual-path supervision and structure perception loss function of a main segmentation path and an auxiliary edge path and outputs a segmentation probability map which is continuous in structure and has a clear boundary based on the multi-level fusion features.
3. The port and dock identification method based on the high-resolution optical satellite image is characterized in that the convolution neural network branches conduct convolution operation on the input high-resolution optical satellite image, edge, texture and corner local detail features are extracted, the visual state space model branches conduct linear complexity modeling on the input high-resolution optical satellite image, semantic association among long-distance pixels in the image is established, global long-distance context relation features are extracted, the convolution neural network branches and the visual state space model branches conduct feature interaction and fusion in a plurality of preset middle layers, then conduct feature fusion through a channel splicing operation and conduct self-adaptive feature weighted addition mode after passing through a 1x1 convolution layer, and multi-scale fusion features with local detail definition and global context consistency are generated.
4. The method for identifying the port and dock based on the high-resolution optical satellite image according to claim 3, wherein in each fusion stage, a preset deep learning network model performs downsampling or projection on the high-resolution feature map branched out by the convolutional neural network, simultaneously performs spatial remodeling on the global long-distance context feature branched out by the visual state space model, performs stitching and feature fusion after aligning the spatial dimensions of the two to obtain a multi-scale fusion feature.
5. The method for identifying the port and the dock based on the high-resolution optical satellite image according to claim 4, wherein the multi-scale feature fusion network receives multi-scale fusion features including different levels and output by a dual-branch feature extraction network, performs up-sampling operation through a top-down path, performs element-by-element addition with high-resolution features from shallow layers, processes input features in parallel through multi-rate cavity convolution operation, captures context information of a dock target at different scales, outputs top-level fusion features, and introduces a channel attention mechanism in all levels to learn and calibrate importance weights of different feature channels in identifying the dock, the shoreline and the yard.
6. The method for identifying the port and dock based on the high-resolution optical satellite image according to claim 5, wherein the structural restoration network is an encoder-decoder structure, the structural restoration network receives top-level fusion features output by the multi-scale feature fusion network, the high-resolution optical satellite image is introduced into a dual-path supervision after being subjected to shallow layer features extracted by branches of the convolutional neural network in a jump connection mode, a main segmentation path outputs a segmentation probability map of a dock area through up-sampling and convolution operation, an auxiliary edge path only processes the originally input high-resolution optical satellite image in the same structure and outputs a pixel-level shoreline edge probability map, and the structural perception loss function consists of three parts of weighted summation of segmentation loss, edge consistency loss and connectivity penalty loss.
7. The method for identifying the port and dock based on the high-resolution optical satellite image is characterized by further comprising a training process aiming at the deep learning network model, and specifically comprises the steps of preparing a training set comprising a high-resolution optical satellite image sample and a pixel-level dock area label graph and a shorelin edge label graph corresponding to the high-resolution optical satellite image sample, adopting an end-to-end joint training mode, taking a loss function perceived by the structure as an optimization target, iteratively updating all parameters of the network by using a counter propagation algorithm, adopting a data enhancement technology to optimize an application environment, wherein the data enhancement technology comprises the steps of simulating storage yard goods shielding, ship berthing, imaging shadows, water flare, image geometry and color disturbance, and adopting multiple rounds of iterative training until model loss converges to a preset value.
8. The method for recognizing the port and dock based on the high-resolution optical satellite image according to any one of claims 6 and 7, wherein the preprocessing process comprises denoising, adjusting orientation and trimming the input high-resolution optical satellite image, performing regular grid sampling on the high-resolution optical satellite image with a size exceeding the network input limit by using a sliding window with a preset size of 1024 x 1024 pixels, and setting the moving step length of the sliding window to be smaller than the window size to ensure that an overlapping area exists between adjacent windows, thereby preventing the dock object from being cut at the window boundary to cause information loss.
9. The method for recognizing the wharf based on the high-resolution optical satellite image according to claim 8, wherein the post-processing process comprises the steps of firstly binarizing and vectorizing a panoramic segmentation probability map output by the structural restoration and refinement network to obtain a preliminary wharf polygon boundary line, then analyzing the wharf polygon boundary line based on a preset geometric rule, wherein the geometric rule comprises a polygon length-width ratio threshold and an adjacent polygon adjacency relation judgment to recognize a potential wharf boundary line segment, detecting a fracture endpoint existing in the boundary line segment, calculating the distance between any two spatially adjacent fracture endpoints and the directional compatibility of the line segments, connecting the two endpoints by using a smooth curve if the distance between the two endpoints is smaller than the preset spatial tolerance threshold and the directional compatibility meets the preset angular tolerance requirement, repeating the process until the two endpoints cannot be connected with a new endpoint, thereby reconstructing the topological continuity of the wharf structure, eliminating the false fracture generated by model prediction or window splicing, and finally outputting the topological vector data of the geometric rule wharf.
10. A port and dock identification system based on high-resolution optical satellite images, which is used for the port and dock identification method based on high-resolution optical satellite images according to any one of claims 1 to 9, and is characterized by comprising an image preprocessing module, a feature extraction module and a post-processing module; The image preprocessing module performs data preprocessing on the input high-resolution optical satellite image and inputs the data into the feature extraction module, the feature extraction module is provided with a preset deep learning network model, the feature extraction module receives the preprocessed high-resolution optical satellite image and outputs a segmentation probability map with continuous structure and fine boundary, and the post-processing module performs post-processing on the segmentation probability map with continuous structure and fine boundary to obtain wharf vector polygon data with topological relation and complete port and wharf identification.

Description

Port and dock identification method and system based on high-resolution optical satellite image Technical Field The invention relates to the technical field of image processing, in particular to a method and a system for identifying a port and a dock based on high-resolution optical satellite images. Background The explosive growth of global trade makes the operation efficiency and security monitoring of ports as seagoing hubs vital. The high-resolution optical satellite remote sensing technology provides data support for realizing large-scale and periodic port facility monitoring. The automatic identification of the ports and the wharfs based on the satellite images has application potential in the fields of port planning, shipping scheduling, illegal construction monitoring, national defense safety and the like. The prior art still has the defects when carrying out automatic and intelligent identification on a port and a dock, wherein the early identification method is edge detection, hough transformation detection straight line and the like, mainly depends on pixel spectral characteristics, texture analysis and simple geometric shape matching, has high requirements on image quality, is extremely easy to be interfered by complex environments around the dock such as storage yard goods, ship shielding, shadow and water flare, has low identification precision and poor robustness, and is difficult to cope with the diversity of different port layouts; the identification method based on traditional machine learning improves the identification capability to a certain extent by manually designing features and combining a classifier, but the method has the characteristics of complex engineering, relies on expert experience, has limited characteristic characterization capability, is difficult to describe a complex space structure of a port and a dock, which is formed by a plurality of elements such as berths, breakwater, cranes, yards and the like, has insufficient generalization capability, has become a mainstream technology of geospatial target extraction particularly based on a semantic segmentation model and an instance segmentation model of a convolutional neural network, however, has the core problems that the structural integrity is difficult to ensure, the port and the dock are usually formed by combining a continuous and slender linear structure (such as a dock shoreline) and a planar structure (such as a yard), the universal segmentation model is easy to predict the continuous shoreline into broken fragments, or can not accurately distinguish the tightly adjacent berths, the extracted structural contour is broken, the topological relationship is wrong, is difficult to be directly used for geometric measurement and space analysis, the general segmentation model is difficult to be applied to the port and the complex structure challenge, the internal facility scale is huge, the method has the advantages that the method has extremely high requirements on the characteristic extraction and fusion capability of the model from a macroscopic kilometer breakwater to a microscopic single crane, the general model is easy to lose detail boundaries of small-scale facilities or fuzzy large-scale facilities when processing such multi-scale targets, the spectrum reflection characteristics are different due to the fact that wharf materials are changeable, the spectrum characteristics of certain roads and squares are possibly similar to those of wharf, the general model has limited distinguishing capability on such phenomena, and the false alarm rate and the omission ratio are high. In view of the above needs and the drawbacks of the prior art, the present application provides a method and a system for identifying a port and a dock based on high-resolution optical satellite images. Disclosure of Invention The invention provides a port and dock identification method and a port and dock identification system based on high-resolution optical satellite images, which solve the problems that in the prior art, structural integrity and continuity of port and dock extraction are difficult to maintain, fracture, burrs and irregular boundaries are easy to generate, the geometric accuracy of an extraction result is poor, and the accurate measurement of key parameters such as dock length, dock number and the like in practical application cannot be met. The primary purpose of the invention is to solve the technical problems, and the technical scheme of the invention is as follows: the invention provides a port and dock identification method based on high-resolution optical satellite images, which comprises the following steps of: S1, preprocessing data of the input high-resolution optical satellite image. S2, inputting the preprocessed image into a preset deep learning network model, and outputting a segmentation probability map with continuous structure and fine boundary, wherein the preset deep learning network comprises a double-branc