Remote sensing image surveying and mapping feature extraction system based on deep learning

Abstract

The invention relates to the technical field of computer vision and geographic information science, and discloses a remote sensing image surveying and mapping feature extraction system based on deep learning, during online operation, a query coding module in the system performs deep deconstruction and vector synthesis on a natural language instruction input by a user; the image coding module performs multi-scale coding on the target remote sensing image to generate a fine and macroscopic feature map; the preliminary matching module executes cross-modal semantic matching; the verification and fusion module is used for comparing the macroscopic environment characteristics of the candidate region with a core concept derived from query so as to effectively filter out misinformation of context inconformity; the module fuses local matching and macroscopic verification results, and drives the feature extraction module to generate high-confidence vectorized ground feature features. According to the system, accurate extraction of any complex semantics is realized, and the automation level and flexibility of remote sensing image feature extraction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and geographic information science and technology, and specifically to a remote sensing image mapping feature extraction system based on deep learning. Background Art

[0002] As the core carrier of Earth's surface information, remote sensing imagery has become a cornerstone data source in numerous fields, including modern surveying and mapping, land resources planning, urban management, environmental monitoring, and disaster response. With the rapid development of aerospace and aviation remote sensing technology, the data we can acquire has reached unprecedented levels of spatial, spectral, and temporal resolution. However, this massive increase in image data also presents new challenges: how to quickly, accurately, and intelligently extract valuable ground feature information from it has become a key bottleneck restricting the effectiveness of remote sensing data applications.

[0003] In the development of automated object extraction technology, early methods primarily relied on low- and mid-level visual information such as spectral statistical features, texture features, or geometric shapes in images. For example, Object-Based Image Analysis (OBIA) has improved object recognition to a certain extent by adopting a segmentation-then-classification strategy. However, these methods often require complex task-specific feature design and tedious parameter tuning, resulting in limited model generalization and difficulty capturing the deeper semantic meaning behind objects. Extraction accuracy and stability are significantly affected by changing lighting conditions, occlusion of objects, or significant seasonal variations in imagery.

[0004] In recent years, the rise of deep learning technology, particularly models such as convolutional neural networks (CNNs), has revolutionized the semantic segmentation of remote sensing imagery. By training end-to-end on a large number of carefully annotated samples, these models automatically learn high-level abstract features, from pixels to object categories. They demonstrate accuracy far exceeding traditional methods in extracting standard features such as roads, buildings, and water bodies.

[0005] Despite this, current mainstream deep learning-based feature extraction techniques still exhibit inherent limitations in practical applications. These models operate on a fundamentally closed-set classification problem, meaning they can only recognize a limited number of categories predefined during the training phase. For completely new feature categories not previously seen in the training data, such as the temporary extraction of "rooftop photovoltaic panels" or "cage aquaculture areas," existing models are unable to address this issue. Numerous labeled samples must be collected and the model trained from scratch, a process that is not only costly but also time-consuming, completely failing to meet the flexibility requirements of surveying and mapping production and emergency response.

[0006] Furthermore, real-world mapping needs are often complex and multifaceted, far beyond the reach of a single category label. Users may be searching for objectives that meet a set of attributes, states, or spatial relationships, such as "all abandoned factories located next to urban green spaces" or "find all temporary buildings with blue roofs on both sides of main roads." Existing semantic segmentation models simply lack the architectural capability to understand these complex, logically layered natural language instructions composed of multiple concepts. They simplify semantic understanding to pixel-level classification mapping, creating an insurmountable gap between high-level human cognitive intent and the underlying pixel processing of machines. This significantly limits the intelligence level of remote sensing information extraction technology and its application depth in a wider range of scenarios.

[0007] Therefore, how to break through the constraints of predefined categories and enable the system to directly understand and execute open and complex feature extraction tasks described by users in natural language is a key technical problem that needs to be urgently solved in the current field of intelligent interpretation of remote sensing images. Summary of the Invention

[0008] The technical problem to be solved by the present invention is that the existing deep learning-based remote sensing image mapping feature extraction methods usually rely on a predefined closed category set, which makes it impossible to process new local object types outside the training set, and it is difficult to understand and execute semantic queries with complex restrictions, and the system lacks flexibility and scalability.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions: The first aspect of the present invention provides a remote sensing image surveying and mapping feature extraction method based on deep learning. The method realizes open understanding of natural language instructions and accurate extraction of land feature by constructing a unified visual-text semantic space and combining spatial context for hierarchical verification.

[0010] In a specific embodiment, the method first requires pre-training of the model. The model adopts a dual-tower neural network architecture, including an image encoder and a text encoder. By using a large amount of "remote sensing image-text description" data pairs, the training method is adopted by contrastive learning. Its goal is to minimize a symmetric contrast loss function , to shorten the distance between matching image-text pairs in the semantic space and push away the mismatched pairs. The loss function is defined as: ; Among them, the loss from image to text and the loss from text to image They are defined as: ; ; Where, and The first feature vectors of images and texts, is the batch size, is the temperature hyperparameter, is the cosine similarity function. Through this step, a unified visual-textual semantic space is constructed.

[0011] After completing model pre-training, this method first obtains the user's natural language query instructions when performing feature extraction.

[0012] One of the innovations of this invention is that it does not directly encode the entire query instruction, but first deconstructs it into multiple independent semantic components, such as core features, attributes, status, etc. These semantic components are then encoded into independent semantic vectors, and then dynamically combined into a final composite query vector using pre-set vector operation rules. One possible algorithm is weighted summation: ; Where, It is vector representation of semantic components, This allows the system to accurately understand and quantify the inherent logic of complex queries.

[0013] Next, the method encodes the target remote sensing image to obtain visual feature vectors for each preset area in the image. Specifically, an image encoder can be used to generate a fine-scale feature map for subsequent preliminary matching.

[0014] Subsequently, the semantic similarity between the aforementioned composite query vector and the visual feature vectors of each image region is calculated to generate a preliminary semantic similarity heat map.

[0015] Another innovation of the present invention is the introduction of a spatial semantic consistency verification mechanism. After determining candidate regions based on the preliminary heat map, the method verifies these candidate regions. Specifically, this step includes: 1. Obtain a larger-scale macroscopic visual feature vector corresponding to the candidate region. This can be achieved by generating an additional macroscopic feature map when encoding the remote sensing image.

[0016] 2. Based on the core concepts in the original natural language query instruction, a more generalized context query text is automatically generated and encoded into a context query vector.

[0017] 3. Calculate the similarity between the macro-visual feature vector and the context query vector to obtain a quantitative context consistency score.

[0018] Finally, the method fuses the semantic similarity of each region with the corresponding context consistency score to generate the final confidence map. The fusion process can be done in a weighted combination: ; Where, is the final confidence score, is the preliminary semantic similarity, is the contextual consistency score, is a fusion coefficient that balances the importance of the two. Ultimately, based on the confidence map, the features of the objects corresponding to the user's natural language query instructions can be accurately extracted.

[0019] A second aspect of the present invention provides a remote sensing image mapping feature extraction system based on deep learning, which is used to perform the above method. The system includes: The query acquisition module is used to obtain the user's natural language query instructions.

[0020] The query encoding module is used to deconstruct and encode the natural language query instruction into a compound query vector.

[0021] The image coding module is used to encode the target remote sensing image, obtain the visual feature vector of the preset area in the image, and generate feature maps of different scales.

[0022] The preliminary matching module is used to calculate the semantic similarity between the compound query vector and the visual feature vector, and generate a preliminary semantic similarity heat map.

[0023] The verification and fusion module is used to determine the candidate regions based on the preliminary heat map, perform spatial semantic consistency verification to obtain the contextual consistency score, and fuse the semantic similarity with the contextual consistency score to generate the final confidence map.

[0024] The feature extraction module is used to extract the ground feature corresponding to the natural language query instruction based on the final confidence map.

[0025] Through the above technical solutions, the present invention establishes a complete technical chain from open semantic understanding to spatial context verification, effectively overcoming the limitations of existing technologies and significantly improving the flexibility, accuracy and intelligence level of remote sensing image surveying and mapping feature extraction.

[0026] The present invention provides a remote sensing image mapping feature extraction system based on deep learning. It has the following beneficial effects: 1. This invention breaks the reliance of traditional methods on predefined categories by constructing a unified semantic space between remote sensing image visual features and natural language text descriptions. When new feature types need to be extracted, users only need to provide the corresponding natural language query instructions to execute the task, without the need to retrain the model. This greatly improves the flexibility and scalability of the surveying and mapping feature extraction system, enabling rapid response to diverse and dynamically changing surveying and mapping needs.

[0027] 2. This invention introduces a mechanism for deconstructing and vectorizing complex natural language queries. This mechanism can decompose complex queries with multiple constraints into multiple independent semantic components and synthesize them into a precise composite query vector. This approach enables the system to deeply understand and execute refined extraction tasks, effectively extracting features with multiple combinations of specific attributes, significantly improving the accuracy and controllability of feature extraction.

[0028] 3. This invention employs a hierarchical spatial semantic consistency verification mechanism. After initially identifying candidate regions, it further verifies the consistency of the region's microscopic semantics with the semantics of its macroscopic environment. This approach, which integrates local recognition with global contextual verification, effectively filters out isolated misjudgments caused by scene complexity or visual ambiguity, significantly enhancing the reliability and accuracy of the final extraction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a schematic diagram of the functional modules of a remote sensing image surveying and mapping feature extraction system based on deep learning according to an embodiment of the present invention; Figure 2 Schematic diagram of the process of an offline model pre-training method according to one embodiment of the present invention; Figure 3 A schematic diagram of the structure of a dual-tower neural network model according to an embodiment of the present invention; Figure 4 A schematic diagram of a complex natural language query processing flow according to an embodiment of the present invention; Figure 5 The figure is a schematic diagram of the hierarchical spatial semantic consistency verification process according to an embodiment of the present invention.

[0030] Among them, 101, query acquisition module; 102, query encoding module; 103, image encoding module; 104, preliminary matching module; 105, verification and fusion module; 106, feature extraction module; 202a, image encoder; 202b, text encoder. DETAILED DESCRIPTION

[0031] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] Please see the attached Figure 1 -Attached Figure 5 The present invention provides a remote sensing image surveying and mapping feature extraction method based on deep learning. The core idea is to construct a unified semantic space that can understand visual and textual concepts in the offline stage, and introduce a query extraction mechanism that combines semantic deconstruction and spatial context verification in the online stage, thereby fundamentally solving the limitations of existing technologies when facing open and complex surveying and mapping needs.

[0033] The execution process of this method can be divided into two interrelated phases. The first phase is offline model pre-training, whose goal is to train a deep neural network model with cross-modal understanding capabilities. The second phase is online feature extraction, which uses the pre-trained model to receive and parse the user's natural language query instructions and perform high-precision ground feature extraction on the target remote sensing image.

[0034] Refer to the attached Figure 1 , Figure 1 The following is a schematic diagram of the functional modules of a deep learning-based remote sensing imagery feature extraction system according to one embodiment of the present invention. This system can be deployed in a hardware environment with high-performance computing capabilities, such as a server or workstation equipped with a graphics processing unit (GPU) and sufficient memory and storage space. The system's software environment can be based on a mainstream operating system (such as Linux) and implemented using a deep learning framework (such as PyTorch or TensorFlow) and corresponding programming languages ​​(such as Python) and libraries.

[0035] The system in the embodiment of the present invention may include: a query acquisition module 101, a query encoding module 102, an image encoding module 103, a preliminary matching module 104, a verification and fusion module 105, and a feature extraction module 106.

[0036] Query acquisition module 101 is responsible for providing a human-computer interaction interface to obtain natural language query instructions input by the user to describe the characteristics of the target feature. This instruction can be a simple noun, such as "water body", or a complex phrase or sentence with multiple restrictions, such as "Find all rural dirt roads that are partially covered by vegetation."

[0037] Query encoding module 102, connected to query acquisition module 101, performs deep semantic parsing and vectorization on acquired natural language query instructions. This module's core task is to convert unstructured natural language into high-dimensional, composite query vectors that can be understood and calculated by machines. This process goes beyond simply encoding the entire sentence and involves more sophisticated semantic deconstruction and synthesis operations.

[0038] Image encoding module 103 is responsible for extracting in-depth visual features from user-specified target remote sensing images. This module utilizes a powerful deep neural network to convert image pixel information into feature maps rich in semantic information. To support subsequent layered verification, this module is configured to simultaneously generate feature maps at different spatial scales, for example, a fine-scale feature map for fine-scale object matching and a macro-scale feature map for environmental context analysis.

[0039] The initial matching module 104 has its inputs connected to the query encoding module 102 and the image encoding module 103. This module performs the core semantic matching operation of the present invention, namely, calculating the semantic similarity between the composite query vector and the visual feature vectors at each location in the fine-scale feature map of the image, thereby generating a preliminary semantic similarity heat map. This heat map intuitively identifies areas in the image that are semantically consistent with the user's query intent.

[0040] The Verification and Fusion Module 105 is key to improving the accuracy of the present invention. This module receives the preliminary results generated by the Preliminary Matching Module 104 and initiates a spatial semantic consistency check. It analyzes the candidate regions identified in the preliminary match and determines whether their semantics are consistent with the semantics of their macro-environmental context. Ultimately, it fuses the preliminary matching results with the verification results to generate a more reliable final confidence map.

[0041] The feature extraction module 106 is connected to the verification and fusion module 105. Based on the final confidence map, this module uses thresholding or other image segmentation techniques to separate high-confidence areas from the background, forming the final surveyed and mapped feature. The output can be vector data (such as Shapefile) or raster data (such as GeoTIFF) that conforms to surveying and mapping industry standards, facilitating subsequent analysis and application.

[0042] The above modules work together to form a complete technical chain from natural language understanding to precise spatial positioning. The following will provide a more detailed explanation of the specific implementation methods involved in each stage and module.

[0043] Refer to the attached Figure 2 , Figure 2The figure is a flow chart of an offline model pre-training method according to an embodiment of the present invention. The feature extraction method provided by the present invention is mainly based on a fully pre-trained deep neural network model with cross-modal semantic understanding capabilities. The goal of this offline pre-training stage is to build a core model that can accurately map the visual patterns of remote sensing images and the abstract concepts of natural language into the same high-dimensional semantic space. The implementation method of this stage may include the following steps: S201. Preparation and processing of training data.

[0044] A notable feature of the present invention is that its training data does not require expensive pixel-level annotation. Instead, this step requires the preparation of a large amount of data in the form of "image-text" pairs. Each data pair consists of a remote sensing image or image slice and a natural language text describing the main content of the image.

[0045] For example, one image corresponds to the text "high-density commercial buildings in the city center", and another image corresponds to "large areas of continuous mountain farmland and terraces".

[0046] Before feeding the data into the model training, it needs to be preprocessed. For image data, uniform size normalization and pixel value standardization can be performed. For text data, tokenization is required to convert it into a numerical sequence that the model can accept.

[0047] S202. Construction of a dual-tower neural network model.

[0048] Refer to the attached Figure 3 , Figure 3 FIG2 is a schematic diagram of a dual-tower neural network model structure according to an embodiment of the present invention. This embodiment adopts a dual-tower structure consisting of an image encoder 202a and a text encoder 202b in parallel.

[0049] The image encoder 202a adopts a Vision Transformer (ViT) architecture, which is used to transform the pre-processed input image Converted into a dimensional visual feature vector The process includes segmenting the image into a series of patches, linearly projecting the patches, adding position encoding, and finally outputting a feature vector that can summarize the visual content of the entire image through a multi-layer transformer encoder network.

[0050] The text encoder 202b uses a transformer-based pre-trained language model (such as BERT) architecture, which is used to transform the text sequence after word segmentation. Converted into a vector with the same dimension as the visual feature vector dimensional text feature vector .

[0051] These two encoders can be initialized with publicly available pre-trained weights at the beginning of training to accelerate model convergence.

[0052] S203. Semantic space alignment training based on contrastive learning.

[0053] This step is the core of cross-modal understanding. Its goal is to optimize the parameters of the two encoders so that In the dimensional semantic space, the feature vectors of semantically related images and texts are close to each other in spatial position, while those that are semantically unrelated are far away from each other.

[0054] Specifically, in a In the training batch of image-text data pairs, for any visual feature vector generated by the image encoder , whose corresponding text feature vector is generated by the text encoder Constitutes a positive sample pair. At the same time, all other text feature vectors in the batch (in ) are considered as constitute a negative sample pair.

[0055] The training process is performed by minimizing a symmetric contrast loss function The loss function is the loss from image to text and text-to-image loss It consists of two parts: ; in, The calculation method is: ; Symmetrically, The calculation method is: ; In the above formula: and Represent the first images and text -dimensional feature vector.

[0056] Represents the size of the current training batch.

[0057] It is used to measure two vectors and The cosine similarity function of similarity is defined as .

[0058] is an adjustable temperature hyperparameter that controls the sharpness of the similarity score distribution.

[0059] In each training iteration, the system calculates the total loss , and using an optimizer such as AdamW, the backpropagation algorithm calculates the gradient of the loss function with respect to all trainable parameters in the image encoder 202a and text encoder 202b, and updates these parameters accordingly. Through repeated iterations of this process, the model is driven to learn a universal semantic correspondence across visual and textual modalities, ultimately forming a stable and aligned unified semantic space, laying the foundation for the subsequent online feature extraction stage.

[0060] After completing the offline model pre-training phase, the feature extraction method provided by the present invention enters the online feature extraction phase. The core of this phase is to utilize the constructed unified semantic space to respond to and execute the user's real-time query instructions, thereby achieving dynamic and accurate extraction of any ground feature.

[0061] Refer to the attached Figure 4 , Figure 4 This is a schematic diagram of the complex natural language query processing process according to one embodiment of the present invention. The first step in the online feature extraction phase is to deeply parse and vectorize the user's natural language query instructions. This step is performed by the system's query encoding module 102 and may include the following sub-steps: S301: Obtain a natural language query instruction.

[0062] The system receives natural language query instructions input by the user through the interactive interface to describe the characteristics of the target object The complexity of this command is not limited. It can be a simple noun or a complex phrase containing multiple qualifying conditions. Taking a typical complex surveying and mapping requirement as an example, the query command that the user may enter is For: "Extracting abandoned industrial buildings with blue photovoltaic panels covering all roofs".

[0063] S302: semantically deconstruct the query instruction.

[0064] In order to achieve a deep understanding of complex query intent, an innovation of the present invention is that the system does not convert the entire query instruction into Instead of being fuzzy as a whole, it is first subjected to a detailed semantic deconstruction using natural language processing technology. This process aims to identify the basic semantic components of different roles that constitute the entire query intent. For the above example instruction, the deconstruction result can be: Core Concepts : industrial plants; state attributive : abandoned; Component Attributive :roof; appendages : Photovoltaic panels; Color Attributes :blue; This deconstruction step can be achieved with the help of technologies such as domain-adapted dependency parsing or named entity recognition, thereby accurately separating the parts of the instruction that carry different semantic information.

[0065] S303: Encode independent semantic components.

[0066] After completing the semantic deconstruction, the system will call the text encoder trained in the offline stage , for each independent semantic component Encode them separately and map them into a unified semantic space to obtain a set of independent, high-dimensional semantic vectors ,in . At this time, each vector They all represent a specific semantic concept (such as the visual concept of "industrial plant" or the color concept of "blue").

[0067] S304. Dynamically synthesize a composite query vector.

[0068] This is the key step in the present invention to achieve precise query control. The system dynamically combines the set of independent semantic vectors obtained in the previous step into a single composite query vector that can summarize the user's complete query intent through a preset vector operation rule. In this embodiment, the synthesis is performed by weighted linear combination, and the calculation formula is as follows: ; In the above formula: Represents the final synthesized composite query vector.

[0069] is the total number of semantic components deconstructed from the query instruction.

[0070] It is The independent semantic vectors corresponding to the semantic components.

[0071] is assigned to The weight coefficient of each semantic component. This weight coefficient is not fixed, but can be automatically assigned according to the role of the semantic component. For example, the industrial plant as the core of the query can be given a higher weight, while blue as a modifying attributive It can be assigned a relatively low weight.

[0072] Through the complete process of "deconstruction-independent encoding-weighted synthesis" mentioned above, the system transforms a meaningful natural language query instruction into a mathematical object with clear directionality and structured information in the semantic space, namely, a compound query vector This vector not only contains all relevant semantic concepts, but also reflects the primary and secondary relationships of these concepts through the difference in weights, laying the foundation for subsequent high-precision semantic matching in remote sensing images.

[0073] The query encoding module 102 generates a compound query vector Afterwards, the online feature extraction process enters the stage of matching the semantic query with the actual remote sensing image. This stage is completed collaboratively by the image encoding module 103 and the preliminary matching module 104. Its purpose is to preliminarily screen candidate regions that semantically match the user's query intent from the large-scale visual data. This process specifically includes the following sub-steps: S305: Perform multi-scale encoding on the target remote sensing image.

[0074] When the user specifies the target remote sensing image to be analyzed After that, the image encoding module 103 will call the image encoder trained in the offline stage The image is processed. Unlike outputting only one global feature vector, the image encoder in this embodiment is configured to output multi-scale feature maps. Specifically, the system generates at least two feature maps with different spatial resolutions: Fine-scale feature maps :The feature map retains a high spatial resolution, and each position on it Each corresponds to one dimensional visual feature vector , the vector It represents the visual content of a small local area (e.g., a pixel or a small patch) on the image. This feature map is the basis for subsequent target positioning.

[0075] Macroscale feature map This feature map has low spatial resolution and is obtained by downsampling the fine-scale feature map or by taking the output of a shallower layer of the model network. Each location on it corresponds to a macroscopic visual feature vector, which summarizes the overall visual pattern and scene information of a larger area in the image (for example, a block or a region). This feature map provides essential environmental context for subsequent spatial context verification.

[0076] S306: Generate a preliminary semantic matching graph.

[0077] The preliminary matching module 104 performs the core cross-modal semantic matching calculation in this step. It converts the single compound query vector generated by the query encoding module 102 into and the fine-scale feature map generated by the image encoding module 103 Perform calculations.

[0078] Specifically, the system calculates and On every position The visual feature vector The cosine similarity between It quantitatively measures the degree of semantic consistency between the visual content of the local area in the image and the user's complete query intent. The calculation formula is as follows: ; By fine-scale feature maps By performing the above calculations on all positions on the image, the system can generate a preliminary semantic similarity heat map aligned with the original image space. In this heat map, the grayscale value or color value of each position corresponds to its calculated similarity score Regions with higher scores indicate that their visual content is closer to the semantic description of the user's query, and are therefore highlighted on the image. These highlighted regions form a preliminary candidate set of target objects to be extracted, providing input for subsequent, more refined verification and fusion steps.

[0079] Refer to the attached Figure 5 , Figure 5: is a schematic diagram of the hierarchical spatial semantic consistency check process according to an embodiment of the present invention. Although the preliminary semantic matching step can effectively identify areas that are visually similar to the query semantics, it only relies on the comparison of local visual features, and sometimes produces some isolated, contextually unreasonable false alarms due to the complexity of the scene or the diversity of the objects themselves. For example, a blue billboard may be visually similar to a "blue photovoltaic panel" and be highlighted. To solve this problem, the present invention introduces a key innovative step, namely, hierarchical spatial semantic consistency check, which is performed by the check and fusion module 105 and can specifically include the following sub-steps: S307: Candidate region identification and context positioning.

[0080] The system first performs a preliminary semantic similarity heat map By setting a preset similarity threshold , all similarity scores in the graph are Higher than The connected regions of the image are identified to form a set of candidate regions (or candidate patches). For each identified candidate region, the system records its spatial position and range on the image. Then, using this spatial position information, the system locates the candidate region in the macro-scale feature map. The corresponding, larger context area in the image is extracted from this location, which can represent the overall environmental background of the area. .

[0081] S308: Generate and match context queries.

[0082] This step is to verify whether the "identity" of the candidate region is consistent with its "environment". The system does not use the original, detailed composite query vector Instead of comparing with the macro environment, a more general context query is intelligently generated.

[0083] Specifically, the system will go back to step S302 to check the original query. The system then automatically derives a more general contextual query text based on the core concept. , such as "industrial area" or "urban functional area".

[0084] The system then uses this contextual query text Input to the trained text encoder In the query vector, we get a context query vector : ; S309: Calculate the spatial semantic consistency score.

[0085] After obtaining the macroscopic visual feature vector representing the environment of the candidate area and a contextual query vector representing reasonable environmental semantics The system can then calculate the cosine similarity between the two to obtain a quantitative spatial semantic consistency score The score is calculated as follows: ; It should be noted that for all pixels in the same macro context area , they share the same contextual consistency score .

[0086] The spatial semantic consistency score It quantitatively measures the consistency between the visual features of the macro environment in which the candidate region is located and the contextual semantic concepts derived from the original query instruction. The score indicates that the local recognition result of the candidate region is semantically consistent with its macro environment, thereby enhancing the possibility that the candidate region is a true target. The score indicates that there is a semantic mismatch between the candidate region and its surrounding environment, providing a direct technical basis for identifying and suppressing false detections caused by local visual similarities.

[0087] After completing the hierarchical spatial semantic consistency check of the candidate regions, the online feature extraction process enters the final decision-making and generation phase. The purpose of this phase is to effectively fuse the local semantic matching information obtained in the previous steps with the macro-environment consistency information to generate a high-confidence final result and convert it into standardized surveying and mapping feature features. This process is performed by the verification and fusion module 105 and the feature extraction module 106, and can specifically include the following sub-steps: S310: Confidence score fusion and final confidence map generation.

[0088] In this step, the verification and fusion module 105 performs a , the similarity score obtained in the preliminary semantic matching The context consistency score obtained in the spatial semantic consistency check Perform weighted fusion to calculate the final confidence score The fusion calculation can be realized by linear combination: ; In the above formula: is the image position The final confidence score of .

[0089] is a preset fusion coefficient, which ranges from 0 to 1. This coefficient is used to balance the importance of local fine matching results and macro context verification results in the final decision. For example, when the query being processed focuses more on the unique visual features of the object itself, the value of When the query is more dependent on the relationship between the object and its environment, the value.

[0090] By performing this fusion calculation for all locations in the image, the system generates a final confidence map Compared to the initial semantic similarity heat map, this final confidence map integrates multi-level semantic information, resulting in a more reliable result that can effectively suppress false positives due to contextual inconsistencies while enhancing the confidence level of the true target.

[0091] S311. Generation and output of final land feature features.

[0092] The feature extraction module 106 receives the final confidence map As input, the final feature extraction operation is performed. First, the system applies a final confidence threshold Binarize the image and convert all confidence scores Higher than The position of is marked as the target foreground, and the rest of the positions are marked as the background, thus generating a binary mask map.

[0093] To further improve the regularity of the results, the system can perform a series of morphological post-processing operations on the binary mask image, such as using erosion and dilation algorithms to eliminate small noise points, fill tiny holes inside the target, and smooth the edges of the target.

[0094] Finally, the system converts the post-processed binary mask image into vectors, transforming connected foreground pixel regions into closed polygons. These polygons and their geospatial coordinates are organized into industry-standard data formats, such as Shapefile vector files or GeoJSON files, and output as the final extracted features. These outputs can be directly loaded into Geographic Information System (GIS) software for subsequent mapping applications such as cartography, spatial analysis, and data updates.

[0095] Through the above embodiments, the present invention realizes an open and high-precision response to natural language instructions by constructing a unified visual-text semantic space and combining the processes of complex query deconstruction, multi-scale matching and hierarchical spatial semantic verification, effectively solving the flexibility and accuracy problems of existing technologies in extracting land features.

[0096] It should be understood that the above embodiments are only exemplary descriptions of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art should know that, based on the technical solutions disclosed in the present invention, various modifications, combinations or equivalent replacements can be made to the above embodiments, and these changes do not deviate from the spirit and scope of the present invention. For example, the specific network structures of the image encoder and text encoder described in the embodiments (such as the visual transformer or the BERT model) can be replaced by other network structures with equivalent or similar functions; the specific values ​​of the weight coefficients, fusion coefficients, and various thresholds can be adaptively adjusted according to the actual application scenarios, and these adjustments also fall within the scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the appended claims.

Claims

1. A remote sensing image mapping feature extraction method based on deep learning, characterized in that: The following steps are involved: Get natural language query instructions; Encoding the natural language query instruction into a compound query vector; Encoding the remote sensing image to obtain a visual feature vector of a preset area in the remote sensing image; Calculating the semantic similarity between the compound query vector and the visual feature vector to generate a preliminary semantic similarity heat map; Determining candidate regions based on the preliminary semantic similarity heat map; Performing a spatial semantic consistency check on the candidate region to obtain a context consistency score; fusing the semantic similarity and the context consistency score to generate a final confidence map; Extracting the feature of the object corresponding to the natural language query instruction according to the final confidence map.

2. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 1, characterized in that: The step of encoding the natural language query instruction into a compound query vector comprises: Deconstructing the natural language query instruction into multiple semantic components; Encoding the multiple semantic components into independent semantic vectors respectively; The independent semantic vectors are combined into the composite query vector through a preset vector operation rule.

3. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 2, characterized in that: The vector operation rule is to perform weighted summation on the independent semantic vectors.

4. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 1, characterized in that: The step of performing spatial semantic consistency checking on the candidate region comprises: Obtaining a macroscopic visual feature vector corresponding to the candidate area; Generate a context query text according to the natural language query instruction, and encode it into a context query vector; The similarity between the macroscopic visual feature vector and the context query vector is calculated to obtain the context consistency score.

5. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 4 is characterized in that: The step of generating a context query text according to the natural language query instruction includes: The core concepts in the natural language query instruction are extracted, and the context query text is generated according to the core concepts.

6. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 1, characterized in that: The step of fusing the semantic similarity and the context consistency score comprises: The semantic similarity and the context consistency score are weightedly combined to obtain a final confidence score, and the final confidence map is generated based on the final confidence score.

7. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 1, characterized in that: The method further includes a model pre-training step, wherein the model pre-training step includes: A dual-tower neural network model is used, which includes an image encoder and a text encoder; Using data pairs consisting of remote sensing images and corresponding text descriptions, the twin-tower neural network model is trained through comparative learning to construct a unified visual-textual semantic space.

8. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 7, characterized in that: The contrastive learning adopts a symmetric contrastive loss function to maximize the feature vector similarity of matched data pairs and minimize the feature vector similarity of unmatched data pairs.

9. The remote sensing image surveying and mapping feature extraction method based on deep learning according to claim 4, characterized in that: The step of encoding the remote sensing image comprises: An image encoder is used to encode the remote sensing image to generate a fine-scale feature map for calculating semantic similarity and a macro-scale feature map for obtaining the macro-visual feature vector.

10. A remote sensing image surveying and mapping feature extraction system based on deep learning, used to execute the method according to any one of claims 1 to 9, characterized in that: include: A query acquisition module is used to obtain natural language query instructions; A query encoding module, configured to encode the natural language query instruction into a compound query vector; An image encoding module is used to encode the remote sensing image and obtain a visual feature vector of a preset area in the remote sensing image; a preliminary matching module, configured to calculate the semantic similarity between the composite query vector and the visual feature vector, and generate a preliminary semantic similarity heat map; A verification and fusion module, configured to determine candidate regions based on the preliminary semantic similarity heat map, perform spatial semantic consistency verification on the candidate regions to obtain a contextual consistency score, and fuse the semantic similarity and the contextual consistency score to generate a final confidence map; A feature extraction module is used to extract the ground feature corresponding to the natural language query instruction based on the final confidence map.

Images