A method and system for processing remote sensing information to calibrate terrain.

Abstract

This invention discloses a remote sensing information processing method and system for calibrating terrain, belonging to the field of remote sensing mapping technology. It addresses the technical problems of traditional remote sensing data processing methods, which limit the depth of utilization of dynamic features, and whose extensive resource allocation exacerbates processing efficiency bottlenecks, making it difficult to meet the comprehensive mapping needs of large-scale, high-timeliness, and multi-element applications. The method includes: extracting the geometric features of fixed markers and the temporal features of dynamic markers respectively; fusing the geometric and temporal features to obtain fused features; dividing the target area into terrain zones based on the fused features to obtain different types of calibrated terrain regions; processing remote sensing information using an adaptive processing algorithm according to the processing priority of the calibrated terrain regions to obtain terrain change information; constructing a dynamic terrain inversion model to obtain terrain inversion change information; and cross-validating the terrain change information with the terrain inversion change information to obtain accurate terrain change information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing mapping, and in particular to a remote sensing information processing method and system for calibrating terrain. BACKGROUND

[0002] In the technical field of remote sensing mapping, although certain progress has been made in the prior art, the traditional remote sensing data processing technology has long been limited by a static processing framework and a rough resource allocation mode, and still has certain technical bottlenecks, mainly embodied in the following three dimensions:

[0003] 1. Spatial adaptability defects caused by static processing paradigm: The existing technical system generally adopts a fixed ground feature labeling strategy based on prior knowledge, and simply defines towers, buildings and other artificial facilities as geometric reference points. This processing method has double limitations: first, the manual labeling process highly depends on the subjective experience of the operator, and is prone to position drift and attribute misjudgment under complex terrain conditions; second, the fixed ground features are only used as a spatial positioning reference for the terrain, and are not included in the dynamic zoning optimization process. Traditional methods usually use a regular grid (such as UTM zoning) or administrative boundaries for data segmentation, resulting in urban building groups and surrounding farmland being forced into the same processing unit, which not only destroys the spatial continuity of ground features, but also causes redundant calculations in urban areas due to over-division, and blurred linear features such as ridges in farmland areas due to insufficient resolution.

[0004] 2. Lack of value mining of time-dynamic characteristics: For ground feature types such as vegetation and water bodies that have significant temporal variation characteristics, the traditional technical system only extracts single-time spectral features, completely ignoring the following key dimensions: NDVI index fluctuations caused by vegetation phenological changes, seasonal changes in water body chlorophyll concentration, and the interaction between dynamic ground features and terrain elements (such as the spatial expansion law of flood inundation areas). This processing mode results in two major technical defects: first, there is no data verification mechanism under temporal constraints; second, it cannot invert the terrain evolution process according to the changes in dynamic ground features.

[0005] 3. Resource mismatch problem of global unified processing architecture: The traditional system adopts a "one-size-fits-all" processing strategy, implementing a unified processing flow for global terrain data: in terms of resolution configuration, the same spatial resolution is used for simple terrain (such as plains) and complex terrain (such as mountains); in terms of algorithm parameter setting, the same preset threshold combination is used, without considering the influence of terrain undulation on spectral response; in terms of computing resource allocation, the same processing priority is used for simple terrain and complex terrain. This mode results in double resource waste: redundant calculations in simple terrain areas, and feature loss in complex terrain areas.

[0006] The above technical defects form a chain reaction: the static processing framework limits the depth of utilization of dynamic features, and extensive resource allocation exacerbates the processing efficiency bottleneck, which together causes the traditional remote sensing data processing system to be difficult to cope with the comprehensive topographic mapping demand of large range, high timeliness and multiple elements. SUMMARY

[0007] The embodiment of the application provides a remote sensing information processing method and system for calibrating topography, which is used for solving the technical problems that the static processing framework of the traditional remote sensing data processing method limits the depth of utilization of dynamic features, and extensive resource allocation exacerbates the processing efficiency bottleneck, and it is difficult to cope with the comprehensive topographic mapping demand of large range, high timeliness and multiple elements.

[0008] The embodiment of the application adopts the following technical scheme:

[0009] On one hand, the embodiment of the application provides a remote sensing information processing method for calibrating topography, and the method comprises the following steps:

[0010] Based on the double-flow neural network, a multi-modal feature fusion model is constructed, and the geometric features and the time sequence features are fused to obtain fusion features;

[0011] A dynamic topographic partition model is constructed, and the target area is topographically partitioned based on the fusion features to obtain different types of calibrated topographic regions;

[0012] Real-time remote sensing information of each calibrated topographic region is obtained, and a corresponding adaptive processing algorithm is called; remote sensing information is processed by the adaptive processing algorithm according to the processing priority of the calibrated topographic region to obtain topographic change information;

[0013] A dynamic topographic inversion model is constructed based on an LSTM network, and the real-time remote sensing information and the historical remote sensing information are input into the dynamic topographic inversion model to obtain topographic inversion change information;

[0014] The topographic change information and the topographic inversion change information are cross-validated to obtain accurate topographic change information.

[0015] In a feasible implementation manner, based on the historical remote sensing information of the target area, the geometric features of the fixed markers and the time sequence features of the dynamic markers are extracted, specifically comprising:

[0016] Historical remote sensing data collected in multiple historical periods of the target area are obtained; wherein the historical remote sensing data at least comprises SAR satellite remote sensing data, airborne laser radar point cloud data and multispectral image data;

[0017] inputting the multispectral image data into a DeepLabV3+ network for target recognition and classification to obtain fixed markers and dynamic markers in a target region; wherein the dynamic markers at least include vegetation and water bodies;

[0018] Based on the nearest point search algorithm, the SAR satellite remote sensing data collected in multiple historical periods and the airborne laser radar point cloud data are registered, and the geometric features of the fixed markers are obtained by feature extraction on the registered data through the random sample consensus (RANSAC) algorithm; wherein the geometric features at least include the outer contour feature and the inclination feature with the ground;

[0019] The normalized difference vegetation index (NDVI) and the normalized difference water index (NDWI) are extracted from each frame of multispectral image data of the historical remote sensing information, and the NDVI time series data and the NDWI time series data are constructed;

[0020] The time series features of the dynamic markers are extracted from the NDVI time series data and the NDWI time series data; wherein the time series features at least include the growth state feature and the area change feature of the vegetation, and the texture change feature and the area change feature of the water body.

[0021] In a feasible implementation, based on a dual-flow neural network, a multi-modal feature fusion model is constructed, and the geometric features and the time series features are fused to obtain fusion features, specifically including:

[0022] On the basis of a standard residual network, a coordinate attention module is inserted after each residual block, and each channel feature is adaptively weighted by an SE module to obtain a geometric flow neural network;

[0023] A Transformer encoder is constructed, and window attention and global attention are alternately used in the Transformer encoder to obtain a time series flow neural network;

[0024] The output end of the geometric flow neural network and the output end of the time series flow neural network are respectively connected with a cross-attention module to constitute a dual-flow neural network;

[0025] A multi-level feature pyramid is constructed, and different levels of features are fused by bilinear interpolation upsampling and element-wise addition;

[0026] The multi-level feature pyramid is connected with the output end of the cross-attention module to constitute the multi-modal feature fusion model together with the dual-flow neural network;

[0027] The geometric features and the time series features are input into the multi-modal feature fusion model to output fusion features.

[0028] In an implementable embodiment, a dynamic terrain partition model is constructed, specifically comprising:

[0029] A multi-level partition architecture comprising a coarse-grained administrative partition module, a medium-grained clustering partition module, and a fine-grained super voxel partition module is constructed to obtain the dynamic terrain partition model;

[0030] The specific construction method comprises:

[0031] Administrative division vector face data is obtained to construct the coarse-grained administrative partition module, which is used to divide administrative partitions and extract regional features of each administrative partition;

[0032] Based on the DBSCAN clustering algorithm, the medium-grained clustering partition module is constructed to perform clustering analysis on the fusion features, determine each clustering partition, and extract regional features;

[0033] Based on the super voxel segmentation algorithm, the fine-grained super voxel partition module is constructed to voxelize the airborne laser radar point cloud data in the historical remote sensing information to obtain voxel cloud data, and perform super voxel segmentation on the voxel cloud data to obtain each super voxel partition and extract regional features.

[0034] In an implementable embodiment, the target area is partitioned based on the fusion features to obtain different types of calibrated terrain regions, specifically comprising:

[0035] The fusion features are input into the dynamic terrain partition model to obtain a plurality of administrative partitions, clustering partitions, and super voxel partitions, and regional features corresponding to each partition; wherein the regional features at least include: building density, texture entropy, night light value, NDVI seasonal amplitude, plot regularity, elevation difference, NDWI, surface temperature, texture contrast, NIR reflectivity, red edge band slope, and elevation;

[0036] According to the terrain discrimination conditions satisfied by the regional features corresponding to each partition, the terrain type of each partition is determined; wherein the terrain type at least includes: urban building area, farmland area, water area, forest area, and mountain area;

[0037] According to the terrain type of each partition, each partition is calibrated to obtain the different types of calibrated terrain regions.

[0038] In an implementable embodiment, before obtaining real-time remote sensing information of each calibrated terrain region and calling the corresponding adaptive processing algorithm, the method further comprises:

[0039] Different adaptive processing algorithms are constructed for different calibration terrain regions, and a remote sensing information processing algorithm library is constructed; wherein, a high-precision geometric processing algorithm is used for urban building areas, a time-series spectral analysis algorithm is used for farmland areas, a sub-pixel decomposition algorithm is used for water areas, and a three-dimensional point cloud processing algorithm is used for forest areas and mountain areas;

[0040] According to the historical regional change rate of the calibration terrain region, the processing priority of each calibration terrain region is determined; wherein, the calibration terrain region with the largest regional change rate has the highest processing priority.

[0041] In a feasible implementation, real-time remote sensing information of each calibration terrain region is acquired, and a corresponding adaptive processing algorithm is called; according to the processing priority of the calibration terrain region, remote sensing information is processed through the adaptive processing algorithm to acquire terrain change information, specifically including:

[0042] The terrain type of the current calibration terrain region is acquired, and a corresponding adaptive processing algorithm is extracted from the remote sensing information processing algorithm library according to the terrain type;

[0043] The historical regional change rate of the current calibration terrain region is calculated, and compared and sorted with other calibration terrain regions in the target region to determine the processing priority of the current calibration terrain region;

[0044] After the calibration terrain region with the previous processing priority is processed, the real-time remote sensing information of the current calibration terrain region is processed through the adaptive processing algorithm to acquire the terrain change information.

[0045] In a feasible implementation, a dynamic terrain inversion model is constructed based on an LSTM network, specifically including:

[0046] A three-layer LSTM network structure is used, and a residual connection is added between the second-layer LSTM network structure and the third-layer LSTM network structure to construct a stacked LSTM network; wherein, the three-layer LSTM network structure is respectively used to capture short-term time-series dependency, medium-term time-series dependency and long-term time-series dependency;

[0047] An attention enhancement module is constructed and connected with the output end of the stacked LSTM network, which is used to dynamically adjust the importance weight of each time step and the contribution weight of the input feature;

[0048] The output end of the attention enhancement module is connected with the output layer to complete the dynamic terrain inversion model, and the model is trained through a pre-constructed NASA GEDI laser elevation data set.

[0049] In an implementable embodiment, the topographic change information and the topographic inversion change information are cross-verified to obtain accurate topographic change information, specifically comprising:

[0050] A first change feature set in the topographic change information is extracted, and a second change feature set in the topographic inversion change information is extracted;

[0051] The feature consistency of each change feature in the first change feature set and the corresponding change feature in the second change feature set is calculated;

[0052] The change feature with feature consistency lower than a preset threshold is determined as an abnormal feature, and a warning is sent and the position information of the abnormal feature is sent to a maintenance terminal;

[0053] After receiving the confirmation feature returned by the maintenance terminal, the non-abnormal features are combined to obtain the accurate topographic change information.

[0054] In another aspect, the embodiment of the present application also provides a remote sensing information processing system for calibrating topography, comprising:

[0055] A topography calibration module is configured to extract geometric features of fixed markers and time sequence features of dynamic markers based on historical remote sensing information of a target region, construct a multi-modal feature fusion model based on a double-flow neural network, fuse the geometric features and the time sequence features to obtain fusion features, construct a dynamic topography partition model, and partition the target region based on the fusion features to obtain different types of calibrated topographic regions;

[0056] A remote sensing information processing module is configured to acquire real-time remote sensing information of each calibrated topographic region, and call a corresponding adaptive processing algorithm; and perform remote sensing information processing by the adaptive processing algorithm according to the processing priority of the calibrated topographic region to obtain topographic change information;

[0057] A topographic inversion module is configured to construct a dynamic topographic inversion model based on an LSTM network, and input the real-time remote sensing information and historical remote sensing information into the dynamic topographic inversion model to obtain topographic inversion change information;

[0058] A cross-verification module is configured to cross-verify the topographic change information and the topographic inversion change information to obtain accurate topographic change information.

[0059] Compared with the prior art, the remote sensing information processing method and system for calibrating topography provided by the embodiment of the present application have the following beneficial effects:

[0060] 1. The application realizes intelligent segmentation of terrain units through a dynamic terrain partitioning model, allowing different characteristic regions such as urban building clusters and farmland to be treated with differentiated granularity. This effectively solves the spatial mismatch problem caused by traditional grid division.

[0061] 2. The application establishes a terrain complexity-computational resource mapping model through an adaptive processing algorithm, which improves the spectral feature retention rate of complex terrain regions such as mountains, while reducing computational redundancy in flat regions. Through a dynamic priority scheduling mechanism, the application achieves precise matching of computational resources and terrain complexity.

[0062] 3. The geometric-temporal fusion feature space constructed by the double-flow neural network significantly enhances the spatio-temporal correlation analysis capability of ground objects, providing more rich feature dimensions for terrain evolution inversion. At the same time, the constructed LSTM dynamic terrain inversion model also maximizes the use of time series data.

[0063] 4. Finally, the bidirectional verification mechanism of terrain change information and inversion information in the application greatly improves the reliability of the final result. A complete evidence chain is formed from phenomenon observation to mechanism inversion. This mechanism effectively eliminates the uncertainty of single-source data and constructs a multi-dimensional verification closed-loop system.

[0064] By constructing the innovative processing paradigm of "feature fusion-dynamic partitioning-intelligent processing-inversion verification", the application forms a revolutionary improvement to the traditional static processing framework, especially in the key areas of complex terrain processing, dynamic ground object monitoring, and resource optimization configuration, showing significant technical advantages. It lays a core methodological foundation for building a new generation of intelligent remote sensing processing system. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments described in the present application, and for those skilled in the art, other drawings can also be obtained without creative labor. In the drawings:

[0066] Figure 1 A flow chart of a remote sensing information processing method for calibrating terrain provided by an embodiment of the present application;

[0067] Figure 2 A structural schematic diagram of a remote sensing information processing system for calibrating terrain provided by an embodiment of the present application. DETAILED DESCRIPTION

[0068] In order for those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0069] The embodiment of the present application provides a remote sensing information processing method for calibrating terrain, as shown in the figure, the remote sensing information processing method for calibrating terrain specifically includes steps S101-S105: Figure 1

[0070] S101, based on the historical remote sensing information of the target area, the geometric features of the fixed markers and the time sequence features of the dynamic markers are extracted respectively.

[0071] Specifically, historical remote sensing data collected in multiple historical periods of the target area are acquired; wherein the historical remote sensing data at least includes SAR satellite remote sensing data, airborne laser radar point cloud data and multispectral image data.

[0072] Further, the multispectral image data is input into the DeepLabV3+ network for target recognition and classification, and the fixed markers and dynamic markers in the target area are obtained; wherein the dynamic markers at least include vegetation and water.

[0073] In one embodiment, the fixed markers include fixed objects on the ground such as buildings, towers and dams. The dynamic markers include changing objects such as vegetation and water that change with the seasons.

[0074] Further, based on the nearest point search algorithm, the SAR satellite remote sensing data and the airborne laser radar point cloud data collected in multiple historical periods are registered, and the geometric features of the fixed markers are obtained by feature extraction on the registered data through the random sample consensus RANSAC algorithm; wherein the geometric features at least include the outer contour feature and the inclination feature with the ground.

[0075] Further, the normalized vegetation index NDVI and the normalized water index NDWI are extracted in each frame of multispectral image data of the historical remote sensing information, and the NDVI time sequence data and the NDWI time sequence data are constructed. The time sequence features of the dynamic markers are extracted in the NDVI time sequence data and the NDWI time sequence data; wherein the time sequence features at least include the growth state features of the vegetation, the area change features, and the texture change features of the water, the area change features.

[0076] ​S102, based on the dual-flow neural network, a multi-modal feature fusion model is constructed, and the geometric features and the time sequence features are fused to obtain the fused features.

[0077] Specifically, on the basis of the standard residual network, a coordinate attention module is inserted after each residual block, and each channel feature is adaptively weighted by an SE module to obtain a geometric flow neural network.

[0078] Then a Transformer encoder is constructed, and window attention and global attention are alternately used in the Transformer encoder to obtain a time sequence flow neural network. The output ends of the geometric flow neural network and the time sequence flow neural network are respectively connected with a cross-attention module to form a dual-flow neural network.

[0079] As a feasible implementation, the Transformer encoder adopts an 8-head self-attention mechanism with a hidden layer dimension of 512. Two attention modes are alternately applied: window attention: attention weights are calculated within an 8x8 local window; global attention: attention is calculated within a complete 128x128 space.

[0080] Further, a multi-level feature pyramid is constructed, and different levels of features are fused by bilinear interpolation upsampling and element-wise addition. The multi-level feature pyramid is connected with the output end of the cross-attention module to form a multi-modal feature fusion model together with the dual-flow neural network.

[0081] Finally, the geometric features and the time sequence features are input into the multi-modal feature fusion model to output the fused features.

[0082] As a feasible implementation, the cross-attention module extracts a 256-dimensional query vector from the geometric flow features, and simultaneously generates a 512-dimensional key-value pair based on the time sequence flow features. Then the query vector is divided into 8 heads (each head is 32-dimensional) for multi-head attention calculation, and a similarity matrix is calculated for each head and key. Finally, the attention weights are generated by softmax normalization, and the fused features are obtained by weighted aggregation of the value vector.

[0083] S103, a dynamic terrain partitioning model is constructed, and the target region is partitioned based on the fused features to obtain different types of calibrated terrain regions.

[0084] Specifically, a multi-level partitioning architecture including a coarse-grained administrative partitioning module, a medium-grained clustering partitioning module, and a fine-grained super voxel partitioning module is constructed to obtain a dynamic terrain partitioning model.

[0085] As a feasible implementation, the specific construction method includes:

[0086] The administrative division vector surface data is acquired, a coarse-grained administrative division module is constructed, is used for dividing administrative division and extracting the regional characteristics of each administrative division. Based on the DBSCAN clustering algorithm, a medium-grained clustering division module is constructed, which is used for clustering analysis of the fusion features, determining each clustering division and extracting the regional characteristics. Based on the super voxel segmentation algorithm, a fine-grained super voxel division module is constructed, which is used for voxelizing the airborne laser radar point cloud data in the historical remote sensing information to obtain voxel cloud data, and performing super voxel segmentation on the voxel cloud data to obtain each super voxel division and extracting the regional characteristics.

[0087] Further, the fusion features are input into the dynamic terrain division model to obtain a plurality of administrative divisions, clustering divisions and super voxel divisions, and the regional characteristics corresponding to each division; wherein the regional characteristics at least include: building density, texture entropy, night light value, NDVI seasonal amplitude, plot regularity, elevation difference, NDWI, surface temperature, texture contrast, NIR reflectivity, red edge band slope and elevation.

[0088] Further, according to the terrain discrimination conditions met by the regional characteristics corresponding to each division, the terrain type of each division is determined; wherein the terrain type at least includes: urban building area, farmland area, water area, forest area and mountain area.

[0089] Finally, according to the terrain type of each division, each division is labeled to obtain different types of labeled terrain regions.

[0090] As a feasible implementation, a feature-type mapping table is constructed in advance, and the fields of the table include terrain type and discrimination condition. For example, for urban building area, if a regional characteristic meets the three discrimination conditions of "building density > 15 / km 2 , texture entropy > 6.5, night light value > 50" at the same time, the region is urban building area. For farmland area, if a regional characteristic meets at least two of "NDVI seasonal amplitude > 0.3, plot regularity > 0.7, elevation difference < 2m", the region is farmland area. For water area, if a regional characteristic meets the main item + any auxiliary item of "NDWI > 0.2 (main item), surface temperature < surrounding 2℃ (auxiliary item), texture contrast < 25 (auxiliary item)", the region is water area. For forest area, if a regional characteristic meets at least two of "NIR reflectivity > 0.4, red edge band slope > 0.05, elevation > 300m", the region is forest area.

[0091] S104, acquiring real-time remote sensing information of each labeled terrain region, and calling corresponding adaptive processing algorithm; according to the processing priority of the labeled terrain region, the remote sensing information is processed by the adaptive processing algorithm to obtain terrain change information.

[0092] Specifically, different adaptive processing algorithms are constructed for different calibrated terrain regions respectively, and a remote sensing information processing algorithm library is constructed; wherein, a high-precision geometric processing algorithm is used for urban building areas, a time-series spectral analysis algorithm is used for farmland areas, a sub-pixel decomposition algorithm is used for water areas, and a three-dimensional point cloud processing algorithm is used for forest areas and mountain areas.

[0093] Further, the processing priority of each calibrated terrain region is determined according to the historical regional change rate of the calibrated terrain region; wherein, the processing priority of the calibrated terrain region with the largest regional change rate is the highest.

[0094] Further, the terrain type of the current calibrated terrain region is obtained, and the corresponding adaptive processing algorithm is extracted in the remote sensing information processing algorithm library according to the terrain type. The historical regional change rate of the current calibrated terrain region is calculated, and compared and sorted with other calibrated terrain regions in the target region to determine the processing priority of the current calibrated terrain region.

[0095] After the processing of the calibrated terrain region with the previous processing priority is completed, the real-time remote sensing information of the current calibrated terrain region is processed by the adaptive processing algorithm to obtain terrain change information.

[0096] S105, based on the LSTM network, a dynamic terrain inversion model is constructed, and the real-time remote sensing information and the historical remote sensing information are input into the dynamic terrain inversion model to obtain terrain inversion change information; the terrain change information and the terrain inversion change information are cross-validated to obtain accurate terrain change information.

[0097] Specifically, a three-layer LSTM network structure is used, and a residual connection is added between the second layer LSTM network structure and the third layer LSTM network structure to construct a stacked LSTM network; wherein, the three-layer LSTM network structure is used to capture short-term time sequence dependency, medium-term time sequence dependency and long-term time sequence dependency respectively.

[0098] Further, an attention enhancement module is constructed and connected with the output end of the stacked LSTM network, which is used to dynamically adjust the importance weight of each time step and the contribution weight of the input features. The output end of the attention enhancement module is connected with the output layer to construct the dynamic terrain inversion model, and the model is trained through the pre-constructed NASA GEDI laser elevation data set.

[0099] Further, a first change feature set in the terrain change information is extracted, and a second change feature set in the terrain inversion change information is extracted. The feature consistency of each change feature in the first change feature set and the corresponding change feature in the second change feature set is calculated.

[0100] The change feature with the feature consistency lower than the preset threshold is determined as an abnormal feature, and an early warning is sent to a maintenance terminal with position information of the abnormal feature. After receiving the confirmation feature returned by the maintenance terminal, the non-abnormal features are combined to obtain accurate terrain change information. The accurate terrain change information can be displayed on a large screen subsequently.

[0101] As a feasible implementation, if the abnormal features in one surveying and mapping are too many, an algorithm iteration early warning is sent to check whether the current model and algorithm are inconsistent with the characteristics of the current target area, prompting the maintenance personnel to adjust the parameters and supplement the training.

[0102] In addition, the embodiment of the present application also provides a remote sensing information processing system for calibrating terrain, as shown in the figure, the remote sensing information processing system 200 for calibrating terrain specifically comprises: Figure 2

[0103] The terrain calibration module 210 is configured to extract geometric features of fixed markers and time sequence features of dynamic markers based on historical remote sensing information of the target area; construct a multi-modal feature fusion model based on a double-flow neural network, and fuse the geometric features and the time sequence features to obtain fused features; construct a dynamic terrain partition model, and partition the terrain of the target area based on the fused features to obtain different types of calibrated terrain regions;

[0104] The remote sensing information processing module 220 is configured to acquire real-time remote sensing information of each calibrated terrain region, and call a corresponding adaptive processing algorithm; perform remote sensing information processing by the adaptive processing algorithm according to the processing priority of the calibrated terrain region to obtain terrain change information;

[0105] The terrain inversion module 230 is configured to construct a dynamic terrain inversion model based on an LSTM network, and input the real-time remote sensing information and historical remote sensing information into the dynamic terrain inversion model to obtain terrain inversion change information;

[0106] The cross-validation module 240 is configured to cross-validate the terrain change information and the terrain inversion change information to obtain accurate terrain change information.

[0107] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments. Especially, for the device, equipment and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0108] ​The above-described embodiments of the application have special structure and can achieve particular advantages. It is to be understood that various equivalents, changes, and modifications can be made by those skilled in the art without departing from the spirit and scope of the present application. Therefore, the described embodiments of the present application are to be considered in all respects as illustrative and not restrictive, and all changes coming within the meaning and equivalency range of the appended claims are intended to be embraced therein.

[0109] The above-described embodiments of the application have special structure and can achieve particular advantages. It is to be understood that various equivalents, changes, and modifications can be made by those skilled in the art without departing from the spirit and scope of the present application. Therefore, the described embodiments of the present application are to be considered in all respects as illustrative and not restrictive, and all changes coming within the meaning and equivalency range of the appended claims are intended to be embraced therein.

Claims

1. A method of processing remote sensing information for calibrating a terrain, characterized in that, The method comprises: Based on the historical remote sensing information of the target area, the geometric features of the fixed markers and the time sequence features of the dynamic markers are extracted respectively; Based on the double-flow neural network, a multi-modal feature fusion model is constructed, and the geometric features and the time sequence features are fused to obtain fusion features; A dynamic terrain partitioning model is constructed, and the target area is partitioned based on the fusion features to obtain different types of calibration terrain areas; Real-time remote sensing information of each calibration terrain area is obtained, and a corresponding adaptive processing algorithm is called; according to the processing priority of the calibration terrain area, the remote sensing information is processed by the adaptive processing algorithm to obtain terrain change information, which specifically includes: Obtain the terrain type of the current calibration terrain area, and extract the corresponding adaptive processing algorithm in the remote sensing information processing algorithm library according to the terrain type; Calculate the historical area change rate of the current calibration terrain area, and compare and sort it with other calibration terrain areas in the target area to determine the processing priority of the current calibration terrain area; After the calibration terrain area with the previous processing priority is processed, the real-time remote sensing information of the current calibration terrain area is processed by the adaptive processing algorithm to obtain the terrain change information; Based on the LSTM network, a dynamic terrain inversion model is constructed, and the real-time remote sensing information and the historical remote sensing information are input into the dynamic terrain inversion model to obtain terrain inversion change information; Cross-validation is performed on the terrain change information and the terrain inversion change information to obtain accurate terrain change information.

2. The method of claim 1, wherein Based on the historical remote sensing information of the target area, the geometric features of the fixed markers and the time sequence features of the dynamic markers are extracted respectively, specifically including: Obtain historical remote sensing data collected in multiple historical periods of the target area; wherein the historical remote sensing data at least includes: SAR satellite remote sensing data, airborne laser radar point cloud data and multispectral image data; The multispectral image data is input into the DeepLabV3+ network for target recognition and classification to obtain fixed markers and dynamic markers in the target area; wherein the dynamic markers at least include vegetation and water body; Based on the nearest point search algorithm, the SAR satellite remote sensing data and the airborne laser radar point cloud data collected in multiple historical periods are registered, and the feature extraction is performed on the registered data by the random sample consensus RANSAC algorithm to obtain the geometric features of the fixed markers; wherein the geometric features at least include the outer contour feature and the inclination feature with the ground; In each frame of multispectral image data of the historical remote sensing information, the normalized vegetation index NDVI and the normalized water index NDWI are extracted, and the NDVI time sequence data and the NDWI time sequence data are constructed; The time sequence features of the dynamic markers are extracted in the NDVI time sequence data and the NDWI time sequence data; wherein the time sequence features at least include the growth state feature of the vegetation, the area change feature of the vegetation, and the texture change feature of the water body, the area change feature of the water body.

3. The method of claim 1, wherein Based on the dual-flow neural network, a multi-modal feature fusion model is constructed, and the geometric features and the time sequence features are fused to obtain fused features, specifically including: On the basis of the standard residual network, a coordinate attention module is inserted after each residual block, and each channel feature is adaptively weighted by the SE module to obtain a geometric flow neural network; A Transformer encoder is constructed, and window attention and global attention are alternately used in the Transformer encoder to obtain a time sequence flow neural network; The output ends of the geometric flow neural network and the time sequence flow neural network are respectively connected with a cross-attention module to form a dual-flow neural network; A multi-level feature pyramid is constructed, and different levels of features are fused by bilinear interpolation upsampling and element-wise addition; The multi-level feature pyramid is connected with the output end of the cross-attention module to form the multi-modal feature fusion model together with the dual-flow neural network; The geometric features and the time sequence features are input into the multi-modal feature fusion model to output fused features.

4. The method of claim 1, wherein A dynamic terrain partitioning model is constructed, specifically including: A multi-level partitioning architecture including a coarse-grained administrative partitioning module, a medium-grained clustering partitioning module, and a fine-grained super voxel partitioning module is constructed to obtain the dynamic terrain partitioning model; The specific construction method includes: Obtain administrative division vector face data to construct the coarse-grained administrative partitioning module for dividing administrative divisions and extracting regional features of each administrative division; Based on the DBSCAN clustering algorithm, the medium-grained clustering partitioning module is constructed for clustering analysis of the fused features to determine each clustering partition and extract regional features; Based on the super voxel segmentation algorithm, the fine-grained super voxel partitioning module is constructed for voxelization of airborne laser radar point cloud data in the historical remote sensing information to obtain voxel cloud data, and super voxel segmentation of the voxel cloud data to obtain each super voxel partition and extract regional features.

5. The method of claim 4, wherein, Based on the fused features, the target area is partitioned into different types of calibrated terrain regions, specifically including: The fused features are input into the dynamic terrain partitioning model to obtain several administrative partitions, clustering partitions, and super voxel partitions, as well as regional features corresponding to each partition; wherein the regional features at least include: building density, texture entropy, night light value, NDVI seasonal amplitude, plot regularity, elevation difference, NDWI, surface temperature, texture contrast, NIR reflectivity, red edge band slope, and elevation; According to the terrain discrimination conditions met by the regional features corresponding to each partition, the terrain type of each partition is determined; wherein the terrain type at least includes: urban building area, farmland area, water area, forest area, and mountain area; According to the terrain type of each partition, each partition is calibrated to obtain the different types of calibrated terrain regions.

6. The method of claim 1, wherein Before obtaining real-time remote sensing information of each calibrated terrain region and calling the corresponding adaptive processing algorithm, the method further includes: Different adaptive processing algorithms are constructed for different calibration terrain regions to construct a remote sensing information processing algorithm library; wherein, a high-precision geometric processing algorithm is used for urban building areas, a time-series spectral analysis algorithm is used for farmland areas, a sub-pixel decomposition algorithm is used for water areas, and a three-dimensional point cloud processing algorithm is used for forest areas and mountain areas; The processing priority of each calibration terrain region is determined according to the historical regional change rate of the calibration terrain region; wherein, the calibration terrain region with the largest regional change rate has the highest processing priority.

7. The method of claim 1, wherein A dynamic terrain inversion model is constructed based on an LSTM network, specifically including: A three-layer LSTM network structure is used, and a residual connection is added between the second and third layers of the LSTM network structure to construct a stacked LSTM network; wherein, the three-layer LSTM network structure is used to capture short-term, medium-term and long-term time-series dependencies; An attention enhancement module is constructed and connected to the output end of the stacked LSTM network to dynamically adjust the importance weight of each time step and the contribution weight of the input features; The output end of the attention enhancement module is connected to the output layer to complete the dynamic terrain inversion model, and the model is trained through a pre-constructed NASA GEDI laser elevation data set.

8. The method of claim 1, wherein, The terrain change information and the terrain inversion change information are cross-validated to obtain accurate terrain change information, specifically including: The first change feature set in the terrain change information is extracted, and the second change feature set in the terrain inversion change information is extracted; The feature consistency of each change feature in the first change feature set and the corresponding change feature in the second change feature set is calculated; Change features with feature consistency lower than a preset threshold are determined as abnormal features, and the location information of the abnormal features is sent to a maintenance terminal with a warning; After receiving the confirmation features returned by the maintenance terminal, the non-abnormal features are combined into the accurate terrain change information.

9. A system for processing remote sensing information for mapping a terrain, characterized by The system includes: A terrain calibration module is used to extract the geometric features of fixed markers and the time-series features of dynamic markers based on the historical remote sensing information of a target region; a multi-modal feature fusion model is constructed based on a double-flow neural network, and the geometric features and the time-series features are fused to obtain fused features; a dynamic terrain partitioning model is constructed, and the target region is partitioned into different types of calibration terrain regions based on the fused features; A remote sensing information processing module is used to obtain real-time remote sensing information of each calibration terrain region and call corresponding adaptive processing algorithms; remote sensing information is processed by the adaptive processing algorithms according to the processing priority of the calibration terrain region to obtain terrain change information, specifically including: The terrain type of the current calibration terrain region is obtained, and the corresponding adaptive processing algorithm is extracted from the remote sensing information processing algorithm library according to the terrain type; The historical regional change rate of the current calibration terrain region is calculated and compared and sorted with other calibration terrain regions in the target region to determine the processing priority of the current calibration terrain region; After the last processing priority of the calibration terrain area is completed, real-time remote sensing information of the current calibration terrain area is processed by the adaptive processing algorithm to obtain the terrain change information; The terrain inversion module is configured to construct a dynamic terrain inversion model based on an LSTM network, input the real-time remote sensing information and historical remote sensing information into the dynamic terrain inversion model, and obtain terrain inversion change information. The cross-validation module is configured to cross-validate the terrain change information and the terrain inversion change information to obtain accurate terrain change information.

Images