A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion

Abstract

This invention discloses a remote sensing image processing method supporting dynamic geographic alignment and metadata fusion, belonging to the field of remote sensing image processing. Addressing the issues of spatial misalignment, serialization deadlock, and high-depth stretching memory overflow during concurrent loading of large-scale multi-source images, this invention determines the tile viewport through the intersection and inverse mapping of physical bounding boxes of multi-source images, and generates a safety buffer zone based on receptive field expansion; it removes GIS spatial pointers, injects the data matrix and scalar spatial metadata independently into subprocesses, and reassembles the virtual dataset in memory; for high-depth tiles, after dynamic detection, it stretches and maps elements in situ in pre-allocated contiguous memory and forces truncation conversion. This invention establishes a closed-loop system from heterogeneous reading and lock-free scheduling to accurate tiling, significantly improving memory efficiency and computational throughput.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically a remote sensing image processing method that supports dynamic geographic alignment and metadata fusion. Background Technology

[0002] In modern deep learning processing pipelines for high-resolution remote sensing imagery, due to the enormous physical size of a single image, it must be sliced ​​into thousands of local grid tiles before being fed into the deep learning model. Especially when processing registration, fusion, or change detection tasks for multi-source sensor imagery, ensuring strict geographic alignment of data tiles sliced ​​from two different reference images in absolute physical space is a core prerequisite task of the entire pipeline. However, existing published patents and engineering solutions still have the following limitations at the underlying architectural level when handling distributed concurrent loading and dynamic tiling tasks for ultra-large-scale multi-source imagery: Firstly, for the block loading of large-scale remote sensing images, patent publication number CN114419191B explores a static image block partitioning strategy to overcome computer memory limitations. However, this method mainly relies on geometric grid rules for static partitioning when performing image cutting, failing to establish a dynamic inverse mapping mechanism based on the real geographic affine transformation matrix (GeoTransform), and thus cannot guarantee sub-pixel-level absolute geographic alignment of two heterogeneous image slices in real time during memory reading. Furthermore, this scheme fails to incorporate a morphological dilation buffer mechanism when extracting intersecting regions, taking into account the edge padding requirements (ContextPadding) of deep learning model convolution kernels. This results in the generation of numerous static binary masks, which can easily lead to damage to the receptive field of image slices at the boundaries due to context feature truncation.

[0003] Secondly, regarding multi-process concurrent scheduling, patent publication number CN113177956B proposes a parallel processing scheme for image semantic segmentation using a multi-process architecture. However, this system fails to completely decouple the underlying spatial projection parameters from the high-bit-depth image matrix during inter-process scheduling in a distributed concurrent environment. Because traditional GIS driver objects (such as GDALDataset) are highly bound to underlying C / C++ memory pointers, this scheme struggles to avoid the deadlock bottleneck in underlying resource allocation caused by the multi-process serialization (Pickle) mechanism across high-level programming languages ​​in modern deep learning high-concurrency DataLoader architectures. Forcibly reducing the image dimension to a regular multi-dimensional array for transmission to avoid deadlock fails to maintain zero-loss mapping of physical spatial information (such as GeoTransform metadata) while image data flows rapidly through memory.

[0004] Thirdly, regarding dimensionality reduction processing for high dynamic bit-depth data, such as the radiometric normalization correction steps for multi-source remote sensing images involved in patent publication CN109191503A, this type of method still uses conventional array copying and implicit data type conversion logic when performing high dynamic bit-depth (such as 16-bit or Float32) to 8-bit low bit-depth data mapping. When facing extremely large concurrent slice loading of tens of millions of pixels, conventional arithmetic stretching implicitly creates an equal-sized floating-point temporary copy matrix in memory. This method does not design in-place operations with pre-allocated single contiguous addresses and immediate forced truncation pipelines for the underlying computing engine, making it difficult to effectively suppress the severe memory spikes and out-of-memory (OOM) risks caused by redundant cross-type temporary array copies. Summary of the Invention

[0005] The purpose of this invention is to provide a remote sensing image processing method that supports dynamic geographic alignment and metadata fusion, in order to solve the problems in the prior art mentioned in the background, such as missing cross-source geographic alignment and convolution edge context filling, inability to serialize and transfer GIS underlying driving objects across processes, and transient memory peak overflow caused by high dynamic bit depth data stretching operations.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion includes the following steps: Constructing a dynamically aligned geographic grid: Extracting geographic bounding boxes from multi-source images and performing physical intersection operations, using an inverse spatial mapping mechanism to map the effective area of ​​the physical intersection back to the pixel coordinate system of the images to dynamically define the slice extraction window, and performing morphological outward expansion of the effective slice area according to the receptive field size parameters of the deep learning model to generate a computational safety buffer zone. Memory-based cross-process routing that decouples execution data stream from spatial metadata: strips away the underlying geographic information system spatial pointer object, extracts the full imagery into a clean data matrix stream compatible with cross-process, and independently extracts and encapsulates lightweight spatial metadata of basic scalar type. The clean data matrix stream and the lightweight spatial metadata are independently injected into concurrent worker processes, and a virtual dataset carrying physical spatial information is reassembled in the child process memory. Perform bit-depth-aware intelligent interception and in-situ rapid normalization: dynamically detect the bit-depth status of the read slice data, and when it is identified as a high dynamic bit-depth data stream, pre-allocate a single continuous memory address segment at the bottom layer, and perform element-level stretching mapping and forced type truncation conversion in-situ within the pre-allocated single continuous memory address segment to output the target low bit-depth image matrix.

[0007] According to the above technical solution, the step of constructing a dynamically aligned geographic grid specifically includes the following physical bounding box intersection and inverse mapping process: The foreground affine mapping model is used to convert the limiting bounding boxes of the input source image and the reference image in their respective pixel spaces into physical bounding boxes in absolute geographic space. Calculate the effective area of ​​physical intersection of physical bounding boxes in absolute geographic space using polygon Boolean operations; For any geographic boundary vertex in the polygon of the physically intersecting effective region, the inverse affine transformation matrix parameters of the reference image are used to accurately inverse map it back to the pixel coordinate system of the reference image.

[0008] According to the above technical solution, the step of constructing a dynamically aligned geographic grid also includes the step of constructing a scaled-down mask for ultra-large-scale remote sensing images: A low-resolution, scaled binary mask is generated under a preset maximum physical side length threshold, and a dynamic scaling factor between the width and height of the real image and the width and height of the scaled mask is calculated accordingly. When scheduling slice tasks concurrently in a multi-threaded manner, the center coordinates of the high-speed sliding window are mapped to the scaled binary mask through the dynamic scaling factor to perform low-complexity extraction validity query verification.

[0009] According to the above technical solution, the step of morphologically expanding the effective slice region outward based on the receptive field size parameter of the deep learning model includes: Based on the theoretical maximum receptive field size and slice step size of the neural network, and combined with the dynamic scaling factor, the expansion kernel radius for the scaled binary mask is dynamically calculated. Based on the calculated radius of the expansion kernel, the corresponding morphological structural element is defined, and a morphological expansion operation is performed on the abbreviated binary mask to force the effective data extraction area to expand outward.

[0010] According to the above technical solution, the process of stripping the underlying geographic information system spatial pointer objects and extracting the entire imagery into a clean data matrix stream compatible with cross-process requirements includes: In the system memory full loading mode, the association between the image entity and the hard disk underlying read driver engine is actively severed; The underlying interface is called to extract all physical pixel blocks of the image entity into a pure array stream that is independent of the system environment and has the characteristics of multi-dimensional continuous memory space.

[0011] According to the above technical solution, the lightweight spatial metadata extraction and encapsulation mechanism has the following characteristics: Extract the underlying metadata necessary to maintain geospatial mapping. The underlying metadata includes at least spatial reference system projection string parameters, affine parameter matrix for absolute geographic anchoring, and invalid pixel value identifiers. The extracted underlying metadata is explicitly encapsulated into a state machine dictionary consisting only of native lightweight scalar data types.

[0012] According to the above technical solution, the process of reorganizing the virtual dataset carrying physical space information in the child process memory includes: After concurrently receiving the clean data matrix stream, the spatial deformer module running on the receiving end directly calls the high-speed virtual memory driver to instantly create a spatial virtual dataset object in the memory stack of the sub-worker process. The reverse calculation environment for physical coordinates is re-established using the parameters in the received state machine dictionary.

[0013] According to the above technical solution, before performing the bit-depth-aware intelligent interception and in-situ rapid normalization steps, a global statistical extreme value extraction strategy is also preset: During the initialization phase of the system slicing task, the pixel statistics at fixed quantiles in non-extreme edges are directly extracted from the thumbnail based on global downsampling, and these values ​​are used as the effective minimum and maximum radiometric extreme values ​​of the target high-resolution image in subsequent normalization processing.

[0014] According to the above technical solution, the step of performing bit-depth-aware intelligent interception and in-situ rapid normalization includes a zero-copy direct path routing mechanism: At the instant the concurrent reader acquires the local slice data to be processed, the system data stream status probe dynamically detects the underlying data type and bit depth characteristics of the currently acquired tensor; When the detection result indicates that the input data is in the standard low-bit deep format supported by the target deep learning model, the system automatically triggers the direct path routing channel and directly returns a zero-copy slice view containing memory pointers to the model.

[0015] According to the above technical solution, the process of performing element-level in-situ stretching mapping and forced type truncation conversion includes: The underlying system takes over the memory allocator to request a unique, contiguous pre-allocated address segment of strictly consistent size, and calls the underlying computing library environment to respond to the computation scheduling of that address segment; Within this unique contiguous pre-allocated address segment, scalar element-level in-situ floating-point arithmetic stretching is performed for high-dynamic bit-deep source data and radiation extrema, followed by immediate seamless safe truncation of the numerical range and forced conversion to the target low-bit-deep unsigned integer to overwrite the original data memory.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention uses a real geographic coordinate system to find the intersection of bounding boxes from multiple source images and employs inverse affine transformation to accurately map the effective area back to the source pixel system. This completely abandons the traditional pure pixel sliding window cutting mechanism, thus ensuring absolute spatial isomorphic alignment of the Earth's surface slices with heterogeneous resolutions and different projection origins, fundamentally eliminating the truncation and damage of edge features. At the same time, it combines the theoretical maximum receptive field dynamic calculation of the morphological expansion kernel radius of the deep learning model to extend the effective area outward into a computational safety buffer zone, fundamentally eliminating the phenomenon of contextual feature false killing caused by the lack of external neighboring pixels in boundary slices at the data input flow level.

[0017] Addressing the engineering pain point of incompatibility between high-level programming language multi-process serialization mechanisms and C / C++ low-level pointers, this invention proactively strips away the outer shell of GIS spatial objects, transforming high-concurrency data streams into pure matrix streams that support zero-copy sharing. Simultaneously, it explicitly extracts and encapsulates the spatial reference frame and affine parameters into a lightweight state machine dictionary containing only native scalars. After both are injected into subprocesses in parallel, the virtual dataset is reassembled in real time using high-speed virtual memory. While maintaining extremely high task distribution throughput, it achieves absolute zero loss and zero deadlock in the cross-process flow of image physical spatial coordinate parameters, completely breaking through the cross-process scheduling deadlock of underlying GIS objects and realizing zero-overhead fusion of physical spatial metadata.

[0018] Furthermore, to address the issue of implicit floating-point copy surges caused by interpreter broadcast operations during dimensionality reduction and stretching of high-dynamic bit-depth data, this invention innovatively introduces a bit-depth-aware intelligent I / O interception channel. This channel directly takes over the memory allocator through the underlying interface, pre-allocating a unique, fixed-size contiguous physical memory address segment for the high-bit-depth data stream. Within this single address segment, it directly calls the underlying compilation library, performing linear stretching in scalar element-level in-situ and immediately performing seamless forced truncation overwriting. This mechanism completely cuts off the generation path of intermediate temporary arrays, hard-compressing instantaneous memory consumption to the theoretical minimum physical limit. This greatly suppresses memory peak oscillations during heterogeneous bit-depth conversion, eliminates the risk of memory overflow at the underlying hardware level, and significantly improves the system stability of industrial-grade massive remote sensing slice concurrent loading pipelines. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the remote sensing image hybrid processing method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of dynamic geographic alignment grid and morphological dilation in this invention. Figure 3 This is a diagram of the pure memory high-speed routing and multi-process concurrent architecture of the present invention; Figure 4 This is the in-situ ultra-fast normalized interception diagram based on probes in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1:

[0022] like Figures 1 to 4 As shown, in one specific embodiment, the present invention provides a remote sensing image processing method that supports dynamic geographic alignment and metadata fusion. The method mainly consists of three core components: a lightweight Geo-Aligned dynamic grid construction module, a pure memory-based high-speed routing concurrency module, and a depth-aware in-situ high-speed normalization module.

[0023] Specifically, to address the issues of geographic alignment and edge truncation across cross-source images, the system performs multi-source physical bounding box (GeoBox) intersection and sub-pixel grid inverse mapping. Let the geographic affine transformation matrices (GeoTransform) of the input source image A and the reference image B be respectively... and The system first extracts the limiting bounding boxes (0,0) of the two images in their respective pixel spaces to... The forward affine mapping model is applied to convert it into a physical bounding box in absolute geographic space. and Then, the effective area of ​​physical intersection between the two is calculated using polygon Boolean operations:

[0024] , Furthermore, for intersecting polygons For any geographic boundary vertex in the image, use the inverse affine transformation matrix of the reference image B. This precisely inversely maps the image back to the pixel coordinate system of image B, dynamically defining a strictly physically aligned slice extraction window. The underlying inverse mapping formula is as follows:

[0025] , In the formula: , These represent the horizontal and vertical pixel coordinates in the pixel coordinate system inverted onto the reference image B, respectively. This indicates the pixel resolution (pixel width) of reference image B in the X-axis direction. This represents the rotation parameters of reference image B in the X-axis direction; This represents the rotation parameters of reference image B in the Y-axis direction; This indicates the pixel resolution (pixel height) of reference image B in the Y-axis direction. , These represent the longitude (or horizontal physical) coordinates and latitude (or vertical physical) coordinates of any geographic boundary vertex in the intersecting polygons, respectively. This indicates the coordinates of the origin of the X-axis of reference image B in the physical coordinate system; This represents the origin coordinates of the Y-axis of the reference image B in the physical coordinate system. This mechanism completely abandons pure pixel sliding windows, ensuring absolute spatial isomorphism on the Earth's surface.

[0026] Preferably, to address the memory consumption of ultra-large-scale images with tens of thousands of pixels, the system operates at extremely low resolutions (e.g., by setting a maximum physical side length threshold). Generate a scaled binary mask at pixel level Specifically, the system calculates the high-precision dynamic scaling factor (ScaleFactor) using the following formula: , , In the formula: Represents the dynamic scaling factor in the X-axis direction (horizontal direction); Represents the dynamic scaling factor in the Y-axis direction (vertical direction); This represents the width of the thumbnail mask in pixels. This represents the actual width in pixels of the original image; This represents the height in pixels of the thumbnail mask; This represents the actual height in pixels of the original image. In multi-threaded task scheduling, it indicates the center coordinates of the high-speed sliding window. Dynamically scaled and mapped to coordinates And in the mask Execution Validity queries with varying complexity.

[0027] Furthermore, such as Figure 2 As shown, to prevent edge feature truncation caused by deep learning convolution kernels, the system introduces a morphological dilation safety buffer expansion with receptive field linkage. Specifically, the size is defined as... Morphological structural elements For lightweight masks The specific formula for performing the expansion operation is as follows:

[0028] ,

[0029] , In the formula: This represents the dynamically calculated radius of the expansion kernel. This represents the round-up operator; This represents the maximum receptive field size in neural network theory; This represents the function that takes the maximum value. This indicates the coordinates of the mask after morphological dilation operation. The value at; This represents a scaled binary mask before the dilation operation is performed. This represents the morphological dilation operator; Represents the morphological structural element used for the dilation operation; Indicates the center pixel coordinates of the target currently being processed; Representing morphological structural elements Local offset coordinates within.

[0030] In one specific implementation, to overcome the serialization deadlock of GIS objects during concurrent loading, this invention performs pure matrix stream stripping and lightweight state machine metadata injection. Specifically, the system actively strips the I / O binding between GIS spatial objects and the hard disk system, and calls the underlying interface to extract the physical pixel blocks of image entities into pure, environment-independent, multidimensional contiguous memory arrays. This allows the system to penetrate multi-process boundaries. Furthermore, the system precisely extracts relevant parameters and explicitly encapsulates them into a lightweight state machine dictionary containing only native scalar data types. Its core structure is:

[0031] , In the formula: This represents a dictionary of the encapsulated lightweight state machine. This represents the extracted spatial reference frame projection string parameter; This represents the affine parameter matrix used for absolute geographic anchoring; This indicates an invalid pixel value in the image.

[0032] Preferably, the deformer module receives the above concurrently. Parallel injection Then, the high-speed virtual memory driver is used to reassemble the virtual dataset in the memory stack of the child process in real time, and the physical coordinate calculation environment is re-established using dictionary parameters to achieve zero-overhead fusion and lock-free reconstruction of physical coordinates.

[0033] In another specific implementation, to address the memory overflow problem caused by high-bit-depth conversion, the system constructs an intelligent I / O interception and in-place ultra-fast normalization channel based on data bit-depth awareness. Specifically, during the task initialization phase, the system uses a global downsampling thumbnail to extract the 1% and 99th percentile values ​​as effective extrema. and Furthermore, when the data stream status probe detects that the input is a standard 8-bit image, it triggers a direct route to return to the memory pointer slice view; if it is identified as high dynamic range data (such as 16-bit), it will return to the memory pointer slice view. If so, the system directly takes over the underlying memory allocator and pre-allocates a single, contiguous memory address segment of completely consistent and fixed size.

[0034] Specifically, the system call binds to a low-level C language-optimized computation library, responds to the address by performing a scalar element-wise stretch in place, and immediately and seamlessly truncates and forces a type cast to 8-bit to overwrite and write back. The formula for forcibly cutting off the pipeline is:

[0035] , In the formula: This indicates that the output target 8-bit low-depth image matrix is ​​in the [missing information - likely a specific location or region]. Line number The pixel values ​​of the column; This represents the type casting operator to an unsigned 8-bit integer; This function represents the larger of the two values. This function represents the smaller of the two values. This indicates that the input high dynamic range source data matrix is ​​at the th... Line number The original pixel values ​​of the column; This represents the effective minimum radiometric extreme value of the pre-extracted image; This represents the maximum effective radiation extremum of the pre-extracted image. This fundamentally cuts off the generation path of the intermediate implicit copy, suppressing instantaneous memory consumption to the theoretical minimum limit.

[0036] Example 2:

[0037] To further illustrate the collaborative application of the pure memory high-speed routing architecture and dynamic alignment grid of the present invention in actual deep learning distributed training, this embodiment provides a specific workflow for dynamic change detection of heterogeneous dual-temporal remote sensing images (such as 8-bit optical images and Float32-bit deep SAR radar images).

[0038] Specifically, during the distributed training startup phase based on a multi-process data loader such as PyTorch or TensorFlow, the system first reads the optical and radar data of two heterogeneous images. Due to the significant differences in resolution and projection origin between the radar and optical images, the system extracts the geographical bounding boxes of both images in the main process and performs physical intersection calculations. Subsequently, it accurately maps the data back to the source pixel system through inverse affine transformation, completely eliminating the heterogeneous spatial misalignment caused by direct pixel cropping.

[0039] Furthermore, when processing edge slices, if the object detection network uses a deep convolutional architecture with a large receptive field (such as the Feature Pyramid Network, FPN), the system will base the calculation on the pre-calculated dilation kernel radius. For low-resolution scaled binary masks Perform a morphological dilation operation. Preferably, this operation forces the effective data extraction area to overflow outward in memory by a corresponding computational safety buffer, so that each slice edge output from the data pipeline contains sufficient neighborhood context information, thereby physically blocking feature truncation and misjudgment of edge targets from the input end.

[0040] Specifically, such as Figure 3 As shown, when entering the cross-process data distribution stage, to avoid the memory allocation deadlock inevitably caused by multi-process serialization of GIS objects containing C pointers in high-level languages, the main process actively separates radar and optical imagery from the underlying driver environment. Furthermore, the system extracts all of this data into a clean multidimensional array stream that supports shared memory transfer, and simultaneously extracts the spatial reference frame projection string and affine parameter matrix, explicitly encapsulating them into a lightweight state machine dictionary containing only scalar types. Preferably, after each working subprocess receives the pure data stream, the spatial deformer calls the underlying virtual memory driver to instantly reassemble a virtual dataset with absolute physical spatial information within an extremely short CPU clock cycle. This embodiment achieves zero-overhead fusion of spatial metadata while maintaining a high concurrency throughput of millions of requests, through the decoupling and reassembly of the pure data stream and the metadata state machine.

[0041] Example 3:

[0042] To further illustrate the anti-out-of-memory (OOM) mechanism of the bit-depth-aware intelligent I / O interception and in-place rapid normalization module in the face of extremely harsh hardware environments, this embodiment takes the concurrent loading scenario of 16-bit multispectral satellite image slices with tens of millions of pixels as an example for in-depth supplementary explanation.

[0043] Specifically, such as Figure 4 As shown, in traditional high-concurrency slice loading pipelines, when converting highly dynamic bit-depth data into model-readable 8-bit unsigned integers (uint8), the underlying interpreter's broadcast mechanism implicitly allocates multiple temporary Float64 copies of memory equal in size to the original image, which can easily lead to instantaneous RAM crashes on computing nodes. Furthermore, to completely block this redundant path, this system places a data flow status probe before the concurrent reader; when the probe identifies the input data matrix... When the dynamic range is 16-bit, the standard arithmetic operation logic is skipped directly, triggering hardware-level interception.

[0044] Preferably, the system's underlying layer actively takes over the memory allocator, using C-API to precisely locate and pre-allocate a single, continuous, and fixed address segment in the memory pool that is strictly identical in size to the original image. Further, it invokes the Just-In-Time (JIT) optimized underlying computation library to directly respond to the specified single memory address, subtracting the effective minimum radiometric extremum from the pixel value at the scalar element level. And divide by the difference in extreme values.

[0045] Specifically, within the same clock cycle during which the floating-point mapping is performed, the system immediately performs a seamless, safe truncation of the numerical range according to a set threshold (0, 255) and forces it to be converted to the target 8-bit low-order deep unsigned integer. The data is directly overwritten and written back to the pre-allocated address. Through the above-mentioned real-time calculation and forced type casting mechanism, which is strictly limited to a single physical memory address, this embodiment hard-compresses the instantaneous peak memory consumption during the heterogeneous high bit depth image conversion process to the theoretically lowest physical limit, ensuring extremely high system stability and I / O throughput in industrial-grade large-scale remote sensing processing scenarios.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion, characterized in that: Includes the following steps: Constructing a dynamically aligned geographic grid: Extracting geographic bounding boxes from multi-source images and performing physical intersection operations, using an inverse spatial mapping mechanism to map the effective area of ​​the physical intersection back to the pixel coordinate system of the images to dynamically define the slice extraction window, and performing morphological outward expansion of the effective slice area according to the receptive field size parameters of the deep learning model to generate a computational safety buffer zone. Memory-based cross-process routing that decouples execution data stream from spatial metadata: strips away the underlying geographic information system spatial pointer object, extracts the full imagery into a clean data matrix stream compatible with cross-process, and independently extracts and encapsulates lightweight spatial metadata of basic scalar type. The clean data matrix stream and the lightweight spatial metadata are independently injected into concurrent worker processes, and a virtual dataset carrying physical spatial information is reassembled in the child process memory. Perform bit-depth-aware intelligent interception and in-situ rapid normalization: dynamically detect the bit-depth status of the read slice data, and when it is identified as a high dynamic bit-depth data stream, pre-allocate a single continuous memory address segment at the bottom layer, and perform element-level stretching mapping and forced type truncation conversion in-situ within the pre-allocated single continuous memory address segment to output the target low bit-depth image matrix.

2. The remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1, characterized in that: The step of constructing a dynamically aligned geographic grid specifically includes the following physical bounding box intersection and inverse mapping process: The foreground affine mapping model is used to convert the limiting bounding boxes of the input source image and the reference image in their respective pixel spaces into physical bounding boxes in absolute geographic space. Calculate the effective area of ​​physical intersection of physical bounding boxes in absolute geographic space using polygon Boolean operations; For any geographic boundary vertex in the polygon of the physically intersecting effective region, the inverse affine transformation matrix parameters of the reference image are used to accurately inverse map it back to the pixel coordinate system of the reference image.

3. A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1 or 2, characterized in that: The step of constructing a dynamically aligned geographic grid also includes the step of constructing a scaled-down mask for ultra-large-scale remote sensing images: A low-resolution, scaled binary mask is generated under a preset maximum physical side length threshold, and a dynamic scaling factor between the width and height of the real image and the width and height of the scaled mask is calculated accordingly. When scheduling slice tasks concurrently in a multi-threaded manner, the center coordinates of the high-speed sliding window are mapped to the scaled binary mask through the dynamic scaling factor to perform low-complexity extraction validity query verification.

4. The remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 3, characterized in that: The step of morphologically expanding the effective slice region according to the receptive field size parameters of the deep learning model includes: Based on the theoretical maximum receptive field size and slice step size of the neural network, and combined with the dynamic scaling factor, the expansion kernel radius for the scaled binary mask is dynamically calculated. Based on the calculated radius of the expansion kernel, the corresponding morphological structural element is defined, and a morphological expansion operation is performed on the abbreviated binary mask to force the effective data extraction area to expand outward.

5. The remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1, characterized in that: The process of stripping away the underlying geographic information system spatial pointer objects and extracting the entire imagery into a clean, cross-process compatible data matrix stream includes: In the system memory full loading mode, the association between the image entity and the hard disk underlying read driver engine is actively severed; The underlying interface is called to extract all physical pixel blocks of the image entity into a pure array stream that is independent of the system environment and has the characteristics of multi-dimensional continuous memory space.

6. The remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1, characterized in that: The lightweight spatial metadata extraction and encapsulation mechanism has the following characteristics: Extract the underlying metadata necessary to maintain geospatial mapping. The underlying metadata includes at least spatial reference system projection string parameters, affine parameter matrix for absolute geographic anchoring, and invalid pixel value identifiers. The extracted underlying metadata is explicitly encapsulated into a state machine dictionary consisting only of native lightweight scalar data types.

7. A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1, characterized in that: The process of reorganizing the virtual dataset carrying physical space information in the child process memory includes: After concurrently receiving the clean data matrix stream, the spatial deformer module running on the receiving end directly calls the high-speed virtual memory driver to instantly create a spatial virtual dataset object in the memory stack of the sub-worker process. The reverse calculation environment for physical coordinates is re-established using the parameters in the received state machine dictionary.

8. The remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1, characterized in that: Before performing the bit-depth-aware intelligent interception and in-situ rapid normalization steps, a global statistical extreme value extraction strategy is also preset: During the initialization phase of the system slicing task, the pixel statistics at fixed quantiles in non-extreme edges are directly extracted from the thumbnail based on global downsampling, and these values ​​are used as the effective minimum and maximum radiometric extreme values ​​of the target high-resolution image in subsequent normalization processing.

9. A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1, characterized in that: The execution of the bit-depth-aware intelligent interception and in-situ rapid normalization steps includes a zero-copy direct path routing mechanism: At the instant the concurrent reader acquires the local slice data to be processed, the system data stream status probe dynamically detects the underlying data type and bit depth characteristics of the currently acquired tensor; When the detection result indicates that the input data is in the standard low-bit deep format supported by the target deep learning model, the system automatically triggers the direct path routing channel and directly returns a zero-copy slice view containing memory pointers to the model.

10. A remote sensing image processing method supporting dynamic geographic alignment and metadata fusion according to claim 1 or 8, characterized in that: The process of performing element-level in-situ stretching mapping and forced type truncation conversion includes: The underlying system takes over the memory allocator to request a unique, contiguous pre-allocated address segment of strictly consistent size, and calls the underlying computing library environment to respond to the computation scheduling of that address segment; Within this unique contiguous pre-allocated address segment, scalar element-level in-situ floating-point arithmetic stretching is performed for high-dynamic bit-deep source data and radiation extrema, followed by immediate seamless safe truncation of the numerical range and forced conversion to the target low-bit-deep unsigned integer to overwrite the original data memory.

Images