Method and device for determining regional ecological environment restoration scheme

Abstract

The invention relates to a method and device for determining a regional ecological environment restoration scheme, and relates to the technical field of ecological environment restoration. The method comprises the following steps: preprocessing multi-modal data corresponding to a target geographic area to obtain standard input data, and inputting the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; determining a candidate ecological restoration scheme according to the ecological state recognition result, and performing a multi-target optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; performing simulation evaluation on the ecological evolution trend of the restoration scheme decision result based on a preset ecological space-time prediction model to obtain an ecological environment prediction result of the target geographic area in a future time period; and optimizing the restoration scheme decision result according to the ecological environment prediction result to obtain a target ecological restoration scheme. According to the invention, the accuracy of the obtained ecological restoration scheme is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of ecological environment restoration, in particular, to a method for determining a regional ecological environment restoration scheme and a device for determining a regional ecological environment restoration scheme. BACKGROUND

[0002] Traditional ecological environment restoration scheme formulation and effect evaluation are mainly based on small sample measurement and experience judgment; however, this method has problems such as poor timeliness, limited spatial coverage, and subjective scheme screening, which further results in low accuracy of the obtained ecological restoration scheme and low efficiency of the ecological restoration scheme generation.

[0003] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0004] The purpose of the present disclosure is to provide a method for determining a regional ecological environment restoration scheme and a device for determining a regional ecological environment restoration scheme, thereby at least partially overcoming the problem of low accuracy of the obtained ecological restoration scheme and low efficiency of the ecological restoration scheme determination due to the limitations and defects of the related art.

[0005] According to one aspect of the present disclosure, a method for determining an ecological environment restoration scheme is provided, including: pre-processing multi-modal data corresponding to a target geographical area to obtain standard input data, and inputting the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; determining a candidate ecological restoration scheme according to the ecological state recognition result, and performing multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; simulating and evaluating the ecological evolution trend of the restoration scheme decision result based on a preset ecological spatio-temporal prediction model to obtain an ecological environment prediction result of the target geographical area in a future time period; optimizing the restoration scheme decision result according to the ecological environment prediction result to obtain a target ecological restoration scheme, and performing ecological restoration on the target geographical area according to the target ecological restoration scheme.

[0006] According to one aspect of the present disclosure, a device for determining an ecological environment restoration scheme is provided, comprising: an ecological state recognition result determination module, configured to preprocess multi-modal data corresponding to a target geographical area to obtain standard input data, and input the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; a restoration scheme decision result determination module, configured to determine a candidate ecological restoration scheme according to the ecological state recognition result, and perform multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; an ecological environment prediction result determination module, configured to simulate and evaluate an ecological evolution trend of the restoration scheme decision result based on a preset ecological spatio-temporal prediction model, to obtain an ecological environment prediction result of the target geographical area in a future time period; and a target ecological restoration scheme determination module, configured to optimize the restoration scheme decision result according to the ecological environment prediction result, to obtain a target ecological restoration scheme, and perform ecological restoration on the target geographical area according to the target ecological restoration scheme.

[0007] The method for determining an ecological environment restoration scheme provided by the embodiments of the present disclosure, on one hand, preprocesses multi-modal data corresponding to a target geographical area to obtain standard input data, and inputs the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; then, determines a candidate ecological restoration scheme according to the ecological state recognition result, and performs multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; further, simulates and evaluates an ecological evolution trend of the restoration scheme decision result based on a preset ecological spatio-temporal prediction model, to obtain an ecological environment prediction result of the target geographical area in a future time period; finally, optimizes the restoration scheme decision result according to the ecological environment prediction result, to obtain a target ecological restoration scheme; since the target ecological restoration scheme can be determined from multiple different dimensions, the problems of poor timeliness, limited spatial coverage and subjective scheme screening in the prior art based on small sample measurement and experience judgment to determine the ecological restoration scheme are solved, and the accuracy of the obtained target ecological restoration scheme is improved; on the other hand, since the target ecological restoration scheme can be determined in an automated manner based on the multi-modal data of the target geographical area, the generation efficiency of the ecological restoration scheme is improved; on the other hand, since the target geographical area can be subjected to ecological restoration according to the target ecological restoration scheme, the ecological restoration efficiency of the target geographical area is improved.

[0008] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0009] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate implementations of the present disclosure and, together with the description, further serve to explain the principles of the present disclosure. It is apparent that the accompanying drawings, in which like numerals indicate corresponding parts, serve only for purposes of illustration of some embodiments of the present disclosure and other accompanying drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0010] Figure 1 A flow chart diagram of a method for determining an ecological environment restoration scheme is schematically shown according to an example embodiment of the present disclosure.

[0011] Figure 2 A structure example diagram of a determination model of an ecological environment restoration scheme is schematically shown according to an example embodiment of the present disclosure.

[0012] Figure 3 A structure example diagram of a remote sensing image segmentation model in a determination model of an ecological environment restoration scheme is schematically shown according to an example embodiment of the present disclosure.

[0013] Figure 4 A structure example diagram of a region feature extraction model in a remote sensing image segmentation model is schematically shown according to an example embodiment of the present disclosure.

[0014] Figure 5 A scene example diagram of a obtained region division result is schematically shown according to an example embodiment of the present disclosure.

[0015] Figure 6 An example diagram of a obtained graph structure is schematically shown according to an example embodiment of the present disclosure.

[0016] Figure 7 An example diagram of dividing a target geographical region into a plurality of ecological sub-regions according to an ecological state identification result is schematically shown according to an example embodiment of the present disclosure.

[0017] Figure 8 A scene example diagram of obtaining a set of all non-dominated solutions based on a Pareto front is schematically shown according to an example embodiment of the present disclosure.

[0018] Figure 9 A scene example diagram of a obtained target ecological restoration scheme is schematically shown according to an example embodiment of the present disclosure.

[0019] Figure 10 A structure example diagram of a determination apparatus of an ecological environment restoration scheme is schematically shown according to an example embodiment of the present disclosure. DETAILED DESCRIPTION

[0020] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations, however, can be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example implementations to those skilled in the art. The described features, structures, or characteristics can be combined in one or more implementations. In the following description, numerous specific details are provided to give a thorough understanding of implementations of the disclosure. One skilled in the relevant art will recognize, however, that the implementations of the disclosure can be practiced without one or more of the specific details, or with other methods, components, devices, steps, etc. In other instances, well-known structures have not been described in detail so as not to obscure aspects of the disclosure. Furthermore, the purpose of the drawings accompanying this disclosure in which like reference numerals designate similar structures is to provide illustrations so that the inventive process can be understood. Like components have been designated with like reference numerals throughout the various drawings and detailed descriptions. Some block diagrams are functional entities that do not necessarily have to correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0021] With the superimposed effects of global climate change and human activities, the degradation of ecosystems is gradually intensifying. At present, the degradation of ecosystems presents characteristics of cross-scale, strong heterogeneity, and nonlinearity, and thus ecological restoration has become an important measure to promote regional sustainable development.

[0022] The formulation and effect evaluation of traditional ecological restoration schemes mainly rely on small sample measurements and empirical judgments, which have problems such as poor timeliness, limited spatial coverage, and subjective scheme screening. At the same time, single static scheme design is difficult to capture the nonlinear changes and elastic recovery characteristics of ecosystems, and existing methods still lack effective means in quantifying the confidence of restoration effect and the system's anti-interference ability.

[0023] In recent years, the rapid development of remote sensing technology has provided multi-scale data support for ecological environment monitoring. Under the premise of remote sensing technology, the interpretation of remote sensing images combined with deep learning can automatically identify land cover types and degraded areas, making remote sensing images perform well in ecological assessment. Therefore, how to improve the scientificity and effectiveness of restoration schemes based on remote sensing image data is a problem that needs to be solved.

[0024] Based on this, the example embodiment first provides a method for determining an ecological environment restoration scheme, which can run on a terminal device, a server, a server cluster, or a cloud server, etc. Of course, those skilled in the art can also run the method of the present disclosure on other platforms according to needs, which is not specially limited in the example embodiment. Referring to Figure 1 The method for determining an ecological environment restoration scheme can include the following steps: Step S110. Preprocess multi-modal data corresponding to a target geographical area to obtain standard input data, and input the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; Step S120. Determine a candidate ecological restoration scheme according to the ecological state recognition result, and make a multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; Step S130. Simulate and evaluate the ecological evolution trend of the restoration scheme decision result based on a preset ecological spatio-temporal prediction model to obtain an ecological environment prediction result of the target geographical area in a future time period; Step S140. Optimize the restoration scheme decision result according to the ecological environment prediction result to obtain a target ecological restoration scheme, and perform ecological restoration on the target geographical area according to the target ecological restoration scheme.

[0025] In the method for determining an ecological environment restoration scheme described above, on the one hand, by preprocessing multi-modal data corresponding to a target geographical area to obtain standard input data, and inputting the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; then determining a candidate ecological restoration scheme according to the ecological state recognition result, and making a multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; further simulating and evaluating the ecological evolution trend of the restoration scheme decision result based on a preset ecological spatio-temporal prediction model to obtain an ecological environment prediction result of the target geographical area in a future time period; finally optimizing the restoration scheme decision result according to the ecological environment prediction result to obtain a target ecological restoration scheme; since the target ecological restoration scheme can be determined from multiple different dimensions, the problems of poor timeliness, limited spatial coverage, and subjective scheme screening in the prior art based on small sample measurement and experience judgment to determine the ecological restoration scheme are solved, and the accuracy of the obtained target ecological restoration scheme is improved. On the other hand, since the target ecological restoration scheme can be determined in an automated manner based on the multi-modal data of the target geographical area, the generation efficiency of the ecological restoration scheme is improved. On the other hand, since the target geographical area can be ecologically restored according to the target ecological restoration scheme, the ecological restoration efficiency of the target geographical area is improved.

[0026] Hereinafter, the method for determining an ecological environment restoration scheme according to the embodiments of the present disclosure will be explained and described in detail with reference to the accompanying drawings.

[0027] Firstly, the technical implementation principle of the embodiments of the present disclosure will be explained and described. Specifically, the method for determining an ecological environment restoration scheme according to the embodiments of the present disclosure can drive the identification of ecological restoration needs and the intelligent matching of schemes based on remote sensing data. In the actual application process, firstly, an anisotropic graph containing terrain constraints is constructed, and a hybrid GNN (Graph Neural Network)-Transformer is used to realize the six-element and degradation fine segmentation of remote sensing images. Secondly, the regional and restoration scheme tokens are soft-aligned through entropy regular optimal transport, and cross-modal matching scores are obtained through joint contrast learning. Subsequently, a double-layer multi-objective optimization (outer layer NSGA-II / III (Non-dominated Sorting Genetic Algorithm II / III) searches for sub-area-scheme assignment, and the inner layer constraint reinforcement learning optimizes continuous parameters) is used to optimize among ecological benefits, costs and risks. A PDE (Partial Differential Equation)-constrained spatiotemporal prediction is combined with ensemble Kalman filtering to perform long-term distributed prediction and data assimilation, quantify uncertainty, and form a prediction-feedback closed loop. This method realizes the whole process from automatic diagnosis to intelligent matching and robust optimization. Compared with the traditional method relying on artificial experience, the method for determining an ecological environment restoration scheme according to the embodiments of the present disclosure not only can solve the problems of lack of large-scale automatic diagnosis, difficulty in cross-modal matching of optimal schemes, and lack of dynamic feedback optimization in the prior art, but also can automatically evaluate the long-term ecological benefits of different measures on a large scale, realize intelligent recommendation of ecological restoration schemes, greatly improve the efficiency and scientific basis of scheme screening, and is suitable for the field of regional ecological environment governance and planning.

[0028] Secondly, the determination model of the ecological environment restoration scheme according to the embodiments of the present disclosure will be explained and described. Specifically, as shown in Figure 2 , the determination model of the ecological environment restoration scheme can include a data preprocessing layer 210, a remote sensing image segmentation model 220, a candidate ecological restoration scheme matching model 230, a multi-objective optimization decision layer 240, an ecological spatiotemporal prediction model 250, and a decision result optimization layer 260. The roles of each model and / or functional layer in the ecological environment restoration scheme will be described in detail hereinafter, and will not be further described here. In an example embodiment, as shown in Figure 3As shown, the remote sensing image segmentation model described herein can include a region feature extraction model (i.e., an encoder) 310 and a region classification model (i.e., a mixing layer + a decoding head) 320; wherein the region feature extraction model described herein can be implemented based on Vision-Transformer, and the region classification model described herein can be implemented based on a graph neural network + a classification layer; at the same time, a specific example diagram of the region feature extraction model described herein can be referred to Figure 4 As shown.

[0029] In the following, the determination method of the ecological environment restoration scheme shown in Figures 2-4 The determination method of the ecological environment restoration scheme shown in Figure 1 will be further explained and described. Specifically: In step S110, the multi-modal data corresponding to the target geographic region is pre-processed to obtain standard input data, and the standard input data is input into a preset remote sensing image segmentation model to obtain an ecological state recognition result.

[0030] In the present example embodiment, first, the multi-modal data corresponding to the target geographic region is pre-processed to obtain standard input data; specifically, this can be achieved by the following way: obtaining multi-modal data corresponding to the target geographic region; wherein the multi-modal data described herein can include but is not limited to multi-source remote sensing data, geographic information data, laser detection data, and soil and meteorological data, etc.; performing grid division on the target geographic region to obtain a plurality of region division results, and determining region attribute data of each region division result according to the multi-modal data corresponding to each region division result; taking each region division result as a node, taking the region attribute data as the node attribute of each node, and taking the adjacency relationship between each region division result as an edge, an original region graph structure is constructed; determining the weight value of the edge according to the node attribute, and updating the original region graph structure according to the weight value to obtain a target region graph structure, so as to determine the standard input data according to the target region graph structure and multi-source remote sensing data.

[0031] In an example embodiment, determining the region attribute data of each region division result according to the multi-modal data corresponding to each region division result can be implemented by: performing radiation correction and / or geometric correction on multi-source remote sensing data corresponding to each region division result to obtain first standard remote sensing data; wherein the multi-source remote sensing data described herein can include but is not limited to multi-spectral remote sensing data, hyperspectral remote sensing data, and unmanned aerial vehicle remote sensing data, etc.; performing cloud and haze removal processing on the first standard remote sensing data to obtain second standard remote sensing data, and determining the region remote sensing index of each region division result according to the second standard remote sensing data; registering the first standard remote sensing data and geographic information data corresponding to each region division result to obtain an image alignment result, and determining the topographic parameter of each region division result according to the image alignment result; determining the laser detection derived layer of each region division result according to the laser detection data corresponding to each region division result, and determining the region attribute data of the region division result according to the region remote sensing index, the topographic parameter, the laser detection derived layer, and the soil and meteorological data corresponding to each region division result.

[0032] In the following, the specific determination process of the standard input data will be further explained and described. Specifically, first, multi-modal data can be obtained from a corresponding database based on the latitude and longitude coordinates or specific region name of the target geographic region; wherein the multi-modal data described herein can include but is not limited to multi-source remote sensing data, Digital Elevation Model (DEM) data, LiDAR (Light Laser Detection and Ranging) data, and soil and meteorological data, etc.; at the same time, the multi-source remote sensing data described herein can include multi-spectral satellite remote sensing data, hyperspectral satellite remote sensing data, and unmanned aerial vehicle remote sensing data, etc.

[0033] In an example embodiment, for the above-mentioned multispectral satellite remote sensing data and hyperspectral satellite remote sensing data, they can be obtained through public or commercial platforms, such as Landsat series (multispectral, 30m resolution), Sentinel-2 (multispectral, 10-60m), Hyperion (hyperspectral, 30m), Gaofen series satellites (domestic, multispectral / hyperspectral), etc. The sensor needs to be selected according to the range and accuracy requirements of the study area (such as hyperspectral suitable for fine feature classification, multispectral suitable for large-scale monitoring); for unmanned aerial vehicle remote sensing data, a multi-spectral camera (such as Parrot Sequoia) or a hyperspectral sensor can be used on an unmanned aerial vehicle, and the flight route is planned according to the terrain of the study area (overlapping degree ≥ 70%), high-resolution (0.1-1m) images are obtained, and camera parameters (focal length, image principal point) and flight logs (GPS position, attitude angle) are recorded simultaneously to obtain. For digital elevation model data, if it is a public data source (such as SRTM data (30m / 90m, global coverage), ASTER GDEM data (30m, with errors in some areas), ALOS World3D data (12.5m, higher accuracy)), it can be downloaded through USGS, NASA Earthdata, etc. If the data is of high accuracy requirement, the local DEM of the study area can be generated by combining unmanned aerial vehicle LiDAR or ground measurement data (the resolution can be below 0.5m). For laser detection and ranging data, it can be obtained through airborne LiDAR systems (such as Riegl, Leica) or terrestrial laser scanning (TLS), and the data format is point cloud (.las / .laz), which contains three-dimensional coordinates (X, Y, Z), echo intensity, classification information (ground points, vegetation points, etc.), and the point cloud density needs to be noted (≥ 10 points / m2 in key areas of the study area). For soil and meteorological data, soil type, organic matter content, pH value, etc. can be obtained from soil survey databases (such as HWSD global soil database, China soil data platform) or supplemented by field sampling and testing; air temperature, precipitation, radiation, etc. can be obtained from meteorological station observation data (China Meteorological Data Network), reanalysis data (such as ERA5, spatial resolution 0.25) or unmanned aerial vehicle-mounted meteorological sensor real-time acquisition.

[0034] In an example embodiment, the pre-processing of data can be implemented based on the following manners: on the one hand, the multi-source remote sensing data is radiometrically corrected to solve the problem of brightness distortion; for example, for satellite images (i.e. satellite remote sensing data), system radiometric correction or relative radiometric correction can be performed; the specific implementation process of the system radiometric correction is to remove the influence of atmospheric scattering / absorption by using sensor parameters (gain, bias) and atmospheric models (such as MODTRAN) to generate a surface reflectance product; the specific implementation process of the relative radiometric correction is to eliminate the inconsistency of radiation caused by the difference in lighting conditions by using pseudo-invariant features (PIF, Pseudo-Invariant Features) or histogram matching on multi-period images of the same area; for example, for unmanned aerial vehicle images (i.e. unmanned aerial vehicle remote sensing data), since they are less affected by the atmosphere, the sensor noise is mainly removed by camera calibration parameters (dark current, exposure time), and the DN (Digital Number, original integer value without unit) is converted into reflectivity by combining ground reflectivity calibration panels (such as gray scale panels).

[0035] On the other hand, the multi-source remote sensing data is geometrically corrected to solve the problem of geometric deformation; for example, for satellite images (i.e. satellite remote sensing data), based on ground control points (GCPs, such as Google Earth high-resolution images or field GPS points), polynomial fitting (2-3 terms) or RPC model (high-resolution satellite) can be used to correct the images to the target coordinate system (such as WGS84 / UTM) with an error controlled within 1 pixel; for example, for unmanned aerial vehicle images (i.e. unmanned aerial vehicle remote sensing data), the flight log (POS data) and camera internal parameters can be used to generate an initial orthographic image through a motion recovery structure algorithm, and then the GCPs are optimized to eliminate the projection difference caused by the terrain undulation (especially in mountainous areas, the orthographic correction needs to be combined with DEM).

[0036] In an example embodiment, the registration of the remote sensing data and the geographic information data (i.e. digital elevation model) can achieve the spatial alignment of multi-source data; the specific registration process is as follows: taking high-resolution images (such as unmanned aerial vehicle orthographic images) as the reference, other data (such as satellite images, DEM) as the data to be registered, the spatial alignment is achieved by using affine or projective transformation through feature point matching (SIFT / SURF algorithm) or manual selection of homonymic points (≥6 pairs), and the registration error is ≤0.5 pixels.

[0037] In an example embodiment, cloud haze removal can be used to eliminate cloud interference; for example, for satellite images (i.e., satellite remote sensing data), cloud detection and cloud repair can be used; the specific implementation process of cloud detection is to use the high reflectivity (visible light band) and low emissivity (thermal infrared band) of clouds to identify cloud areas in combination with cloud mask products (such as the QA60 band of Sentinel-2) or algorithms (such as Fmask); the specific implementation process of cloud repair is to use multi-period image time series interpolation (such as linear interpolation, harmonic analysis) or similar feature replacement (such as neighborhood similar pixel filling) for cloud-covered areas. For example, for unmanned aerial vehicle images (i.e., unmanned aerial vehicle remote sensing data), because the flight height is low, the influence of cloud and haze is small, and single-band threshold (such as blue band reflectivity > 0.8 to determine as cloud) or multi-band index (such as NDSI = (green-short wave infrared) / (green+short wave infrared)) can be used to identify, and small-area cloud areas can be directly cropped or repaired by interpolating the surrounding cloud-free areas.

[0038] In an example embodiment, regional remote sensing indices (such as NDVI, NDWI, and EVI) can be determined based on multi-spectral / hyper-spectral remote sensing data; the specific calculation process is as follows: first, NDVI (Normalized Difference Vegetation Index) can be used to reflect the coverage of vegetation (range -1 ~ 1, positive value for vegetation): the specific calculation formula is: NDVI = (NIR-Red) / (NIR+Red); where NIR is the reflectance value of the near-infrared band, and Red is the reflectance value of the red light band; second, NDWI (Normalized Difference Water Index) can be used to highlight water bodies (high value for water bodies): the specific calculation formula is: NDWI = (Green-NIR) / (Green+NIR); where Green is the reflectance value of the green light band; further, EVI (Enhanced Vegetation Index) can be used to reduce soil and atmospheric interference to improve vegetation monitoring accuracy; the specific calculation formula is: EVI = 2.5 (NIR-Red) / (NIR+6Red-7.5Blue+1); where Blue is the reflectance value of the blue light band.

[0039] In an example embodiment, the terrain parameters can be used to characterize the degree of inclination and the curvature of the ground unit; wherein the degree of inclination can be specifically calculated by the angle between the DEM normal vector and the vertical direction (unit: degree or percentage), and the formula is based on the elevation change rate in a 3x3 window; the curvature can include planar curvature (affecting water flow diffusion) and profile curvature (affecting water flow velocity), which can be specifically calculated by the second derivative of the DEM.

[0040] In an example embodiment, the LiDAR derived layer (i.e. laser detection derived layer) can include but is not limited to vegetation height, point cloud density, and average echo intensity, etc.; wherein the specific determination process of the vegetation height is to extract the ground points after point cloud classification to generate a high-precision DEM, and the non-ground points to generate a DSM (Digital Surface Model), and the difference between the two is CHM (Canopy Height Model, reflecting vegetation height); the point cloud density and the average echo intensity can be realized based on the corresponding algorithm, and this example does not make special restrictions.

[0041] In an example embodiment, the specific construction process of the region graph structure can be realized by the following way: first, the target geographical region is divided to obtain a plurality of region division results; wherein the obtained region division result can be represented as , N is the number of region division results, and a specific example graph can be referred to as shown in Figure 5 ; in the division process, the division can be based on a square of a predetermined size, or the division can be based on the actual geographical situation, and this example does not make special restrictions; the area of the obtained region division result can be consistent or inconsistent, and this example does not make special restrictions; second, taking the region division result as a node and taking a specific adjacency relationship as an edge, a corresponding graph structure is constructed; wherein the obtained graph structure can be referred to as shown in Figure 6 ; further, the weight value of the edge in the graph structure needs to be determined; specifically, the specific determination process of the weight value can be shown in the following formula (1): ; formula (1) wherein, is the weight value of the edge between the i-th region division result and the j-th region division result, is the region coordinate of the i-th region division result, is the region coordinate of the j-th region division result, is the Gaussian kernel bandwidth of the spatial distance (controlling the influence degree of the spatial distance on the weight); is the region attribute data of the i-th region division result; is the region attribute data of the j-th region division result; is the Gaussian kernel bandwidth of the feature similarity (controlling the influence degree of the feature difference on the weight); The elevation gradient between the i-th region division result and the j-th region division result; The Gaussian kernel bandwidth of the terrain slope (controls the degree of influence of terrain differences on the weights). Based on this, an anisotropic weight map with terrain constraints can be obtained. Then, it is also necessary to perform anisotropic weighting on the terrain-constrained graph. Normalization is performed; the specific normalization process can be based on the following formula (2): ; Formula (2) in, A is the normalized anisotropic weight graph, D is the degree matrix of the anisotropic weight graph A, and I is the identity matrix. After obtaining the normalized anisotropic weight graph, the weights of the edges in the original region graph structure can be updated based on the normalized anisotropic weight graph to obtain the target region graph structure.

[0042] Secondly, the standard input data is input into a preset remote sensing image segmentation model to obtain the ecological status identification result. Specifically, this can be achieved as follows: based on the regional feature extraction model, multiple key features at different scales are extracted from the second standard remote sensing data corresponding to the multi-source remote sensing data in the standard input data to obtain feature extraction results corresponding to each of the regional division results; based on the regional classification model, clustering and reasoning are performed on the feature extraction results and the target region map structure in the standard input data to obtain the ecological status identification result; wherein, the ecological status identification result is used to characterize the distribution of degraded plots and / or vegetation coverage in the target geographic region. Specifically, in practical applications, preprocessed multimodal data can be input into a remote sensing image segmentation model for ecological status identification. This model integrates graph neural networks (GNNs) and Transformer mechanisms to fully extract deep features and spatial correlation information from remote sensing images. In practice, firstly, a visual Transformer can be used to extract multi-scale depth features from the remote sensing image. Then, a graph neural network is constructed to represent image pixels or superpixels as graph nodes and proximity relationships as graph edges, thereby clustering and reasoning about deep features to achieve semantic segmentation of the image. This segmentation process can accurately identify different land cover types and degradation levels within a region, generating ecological diagnostic maps, such as the distribution of degraded plots, vegetation cover, and soil erosion patches.

[0043] In an example embodiment, the region feature extraction model described above can include an image blocking layer, a first linear transformation layer, an embedding layer, a position encoding module, a rotation feature encoding module, a feature decoding module, a second linear transformation layer, and a classification layer; under this premise, based on the region feature extraction model, the second standard remote sensing data corresponding to the multi-source remote sensing data in the standard input data is subjected to key feature extraction of multiple different scales, and the feature extraction result corresponding to each region division result is obtained, which can be realized in the following way: based on the image blocking layer, the second standard remote sensing data is subjected to image segmentation to obtain a plurality of image blocks, and based on the first linear transformation layer, the image blocks are subjected to linearization processing to obtain a first input feature sequence; based on the embedding layer, the first input feature sequence is subjected to embedding processing to obtain a first embedding vector, and based on the position encoding module, the first input feature sequence is subjected to encoding processing to obtain a first position vector; the first embedding vector and the first position vector are superimposed to obtain a first input vector, and based on the rotation feature encoding module, the first input vector is subjected to encoding processing to obtain a first encoding result; based on the feature decoding module, the first encoding result is subjected to decoding processing to obtain a first encoding matrix, and based on the second linear transformation layer, the first encoding matrix is converted to obtain a first logic matrix; based on the classification layer, the first logic matrix is subjected to mapping processing to obtain the feature extraction result corresponding to each region division result.

[0044] In an example embodiment, the rotation feature encoding module described above includes a first multi-head self-attention module, a first residual connection and normalization module, a first feedforward neural network, a second residual connection and normalization module; wherein, based on the rotation feature encoding module, the first input vector is subjected to encoding processing to obtain the first encoding result, which can be realized in the following way: based on the first multi-head self-attention module, the first attention mechanism of the first input vector is calculated, and based on the first residual connection and normalization module, the residual connection and normalization processing of the first attention mechanism is performed to obtain the first normalization processing result; based on the first feedforward neural network, the first normalization processing result is subjected to linearization processing to obtain the first linearization processing result, and based on the second residual connection and normalization module, the residual connection and normalization processing of the first linearization processing result is performed to obtain the first encoding result.

[0045] In an example embodiment, the first attention mechanism of the first input vector based on the first multi-head self-attention module can be implemented in the following manner: linear fusion is performed on the first input vector to obtain a first query vector, a first key vector and a first value vector, and the relative angle of the first query vector and the first key vector is adjusted based on a preset first rotation matrix to obtain an adjusted first query vector and an adjusted first key vector; a first outer product vector of the adjusted first query vector and the adjusted first key vector is calculated to obtain a first similarity between the adjusted first query vector and the adjusted first key vector; the first similarity is normalized to obtain a first weight matrix, and a second outer product vector of the first weight matrix and the first value vector is calculated to obtain the first attention mechanism.

[0046] In an example embodiment, the specific determination process of the ecological state recognition result can be implemented by the following formula (3)-formula (6): , , ; formula (3) ; formula (4) , ; formula (5) , ; formula (6) wherein, is the hidden feature of the i-th layer (i is the layer number of the network) is the number of network layers , and are the weight matrix of the query (Q), the weight matrix of the key (K) and the weight matrix of the value (V) in the Transformer (i.e. the regional feature extraction model) is the query vector, is the key vector, is the value vector is the feature dimension of the query (Q), the key (K) and the value (V), which is used to scale the attention score to avoid too large numerical value is the relative position bias based on the spatial coordinates and DEM of the regional division result, which is used to improve the sensitivity of the attention mechanism to the terrain is the output of the global attention mechanism in the Transformer (i.e. the regional feature extraction model), which is used to capture the feature correlation between long distances is the output of the graph neural network, which is used to capture the local spatial correlation is the Chebyshev polynomial of the k-th order, which is used to approximate the graph convolution to reduce the calculation amount is a learnable parameter for the coefficient of graph convolution, is a fusion weight of attention and graph convolution, is a sigmoid activation function, is a weight matrix. Meanwhile, the output of the decoding head of the graph neural network is which can be used to represent the probability that the i-th regional division result belongs to the c-th degradation level in the "mountain, water, forest, farmland, and grassland"; based on this, the ecological diagnosis map Y of the target geographical area can be obtained, and a homogeneous sub-region set can be formed based on the probability of the degradation level of each regional division result . Among them, the ecological diagnosis map can be used to mark the distribution of degraded plots and vegetation coverage; the homogeneous sub-region set can be used to divide regional division results with similar ecological characteristics into the same sub-region, so as to facilitate subsequent repair scheme matching.

[0047] Further, in order to strengthen the boundary and continuity of the remote sensing image segmentation model, a compound loss can be used in the process of model training; wherein the specific loss function can be shown in the following formulas (7)-(12): ; formula (7) ; formula (8) ; formula (9) ; formula (10) ; formula (11) ; formula (12) wherein, is the total segmentation loss, in the process of actual application, in order to improve the segmentation accuracy, five kinds of loss can be fused in the total segmentation loss; , , , and are the weight coefficients of each kind of loss, which can be used to balance the contribution degree of different losses; is the focus loss, which can solve the problem of class imbalance (for example, there are few degraded areas but many normal areas); is the weight of the c-th class, which can be used to balance the difference in sample size of different classes; is the focus parameter of the focus loss, the larger the focus parameter, the more attention is paid to the difficult-to-distinguish samples; is the probability that the i-th regional division result is predicted as the c-th class; N is the number of sub-regional division results; is the Dice loss (also known as F1 loss) function, which can be used to improve the accuracy of the segmentation boundary; The true label for whether the i-th region partitioning result belongs to class c; The total variable loss can be used to maintain the spatial smoothness of the segmentation results; E is the set of adjacent edges between the segmentation results of each region. The coordinates of the i-th region are the result of the region division. The coordinates of the region resulting from the division of the j-th region; Let be the weight value of the edge between the i-th region partitioning result and the j-th region partitioning result; Boundary loss can be used to enhance the accuracy of boundary recognition in degraded regions; The gradient of the prediction result, The gradient of the true label can be used to describe the transformation of the boundary. The Lovars loss function can be used to further optimize the IoU performance of the segmentation; The total number of categories represents the total number of categories. In this example embodiment of the disclosure, the specific value can be 6; c represents the c-th category. For the prediction result of the c-th category, The true label for the c-th category; This represents the overall regional division results encompassing the entire geographic region; For intersection calculation, This is for calculating the union.

[0048] In step S120, candidate ecological restoration schemes are determined based on the ecological state identification results, and multi-objective optimization decisions are made on the candidate ecological restoration schemes to obtain restoration scheme decision results.

[0049] In this exemplary embodiment, firstly, candidate ecological restoration schemes are determined based on the ecological state identification results; specifically, this is achieved by dividing the target geographical area into multiple ecological sub-regions based on the ecological state identification results (see details...). Figure 7 As shown), a preset original ecological restoration scheme is obtained; based on a preset cross-modal matching model, cross-modal matching is performed between the preset original ecological restoration scheme and the ecological identification results of each ecological sub-region to obtain a matching score; based on the matching score, candidate ecological restoration schemes required for each ecological sub-region are determined from the preset original ecological restoration schemes. The preset cross-modal matching model described here is obtained as follows: a feature extraction network for image modality and text modality is trained using contrastive learning or triplet loss under preset constraints to obtain a preset vector embedding model; wherein the preset constraint is that the first distance between the region division results of the same category and the original ecological restoration scheme in the embedding space is less than the second distance between the region division results of different categories and the original ecological restoration scheme in the embedding space.

[0050] The following section will further explain and illustrate the specific process for determining candidate ecological restoration schemes. Specifically, in practical applications, the determination of candidate ecological restoration schemes can be achieved through cross-modal ecological restoration scheme matching. Under this premise, the aforementioned ecological diagnosis results can first be transformed into vector representations characterizing regional ecological features (e.g., combining remotely sensed ecological indicators and degradation types to form feature vectors), which are then input into the cross-modal scheme matching module. This module stores a variety of pre-collected ecological restoration schemes and their feature descriptions (e.g., text descriptions, parameter configurations, etc.). A deep learning model projects regional features and scheme features into a unified high-dimensional embedding space, calculating a matching similarity score. Preferably, contrastive learning or triplet loss can be used to train the feature extraction networks for image and text modalities to ensure that regions and schemes of the same category are closer in the embedding space, while those that do not match are farther apart. This yields several candidate restoration schemes adapted to the target region and their matching scores, solving the problem of semantic alignment between remote sensing information and restoration measures.

[0051] In one example embodiment, during the matching of candidate ecological restoration schemes, firstly, the set of homoproton regions (i.e., the results of identifying the same ecological state) is analyzed. any ecological subregion Feature aggregation is performed to obtain region embeddings. Any of the pre-set original ecological restoration plans Encode the action using a triplet or text / parameter graph based on the "measure-object-constraint" structure. ;in, This is a pre-defined vector embedding model, also known as a feature aggregation function; Let d be the feature embedding (i.e., the first embedding mapping vector) of the r-th ecological sub-region, where d is the feature dimension; For the first An original ecological restoration plan; The feature embedding for the i-th original ecological restoration scheme; The preset vector embedding model, also known as the scheme encoding function, may include, but is not limited to, text vector embedding models and / or image vector embedding models, etc.

[0052] Secondly, token-level soft alignment and scoring (OT) are performed; specifically, the set of regional tokens for the region partitioning results is defined. And the set of scheme tokens for the original ecological restoration plan. ;in, The number of regional tokens can be used to break down ecological sub-region features into fine-grained tokens; The number of scheme tokens can be used to split the characteristics of the repair scheme into fine-grained tokens; on this basis, a cost function between the regional division result and the original ecological restoration scheme is defined ; wherein, is the cost between the regional token t and the scheme token u, which can be determined by the square of the Euclidean distance, and the smaller the distance, the more matched. At the same time, the entropy regular OT is defined; wherein, the specific expression of OT can be shown in the following formula (13): ; formula (13) wherein, is the entropy regular optimal transport (OT) matrix, which can be used to describe the optimal matching relationship between the regional token and the scheme token; is an entropy regularization parameter, which can be used to control the smoothness of the OT matrix to avoid the matching result being too extreme; is a set of transport matrices that satisfy the marginal constraint, a is the marginal distribution of the regional token, and b is the marginal distribution of the scheme token; is the token-level matching score between the region r and the scheme , which can be specifically represented by the sum of the cost, and the larger the value, the more matched; is the global fusion score, is a fusion function, which can combine the embedding similarity and the token-level score to determine the global fusion score, so as to improve the accuracy of the obtained global fusion score.

[0053] Further, the loss function used by the above-mentioned preset cross-modal matching model in the model training process can be specifically shown in the following formula (14): ; formula (14) wherein, is the cross-modal matching loss, which can be used to ensure the accuracy of the matching result; is a temperature parameter, which can be used to control the smoothness of the similarity distribution; is the positive example scheme (i.e. the adaptive repair scheme) matched with the ecological sub-region r, is the negative example scheme (i.e. the non-adaptive repair scheme) not matched with the ecological sub-region r; is the weight coefficient of the triplet loss; is the interval parameter of the triplet loss, which can be used to ensure that the positive example score is at least m higher than the negative example score.

[0054] ​Secondly, the candidate ecological restoration scheme is subjected to multi-objective optimization decision to obtain a restoration scheme decision result; specifically, the multi-objective optimization decision can be realized by constructing a multi-objective optimization decision corresponding to the candidate ecological restoration scheme; wherein the multi-objective optimization decision includes multiple optimization decisions of ecological benefit, economic cost, implementation cycle and social impact; the candidate ecological restoration scheme is input into a first scheme optimization model to dynamically adjust the candidate ecological restoration scheme to obtain a first optimization result; the multi-objective optimization decision and the first optimization result are input into a second scheme optimization model to screen the candidate ecological restoration scheme to obtain a second optimization result; and the second optimization result is input into a third scheme optimization model to predict the risk of the second optimization result to obtain the restoration scheme decision result. That is, in the actual application process, the selected candidate ecological restoration scheme can be subjected to double-layer multi-objective optimization decision; specifically, for the candidate restoration scheme, a multi-objective optimization model can be established to comprehensively balance multiple indexes such as ecological benefit, economic cost, implementation cycle and social impact to make an optimization decision on the candidate scheme; wherein the optimization process can be realized by combining a reinforcement learning (RL) and a non-inferior solution sorting genetic algorithm (NSGA-II); the specific implementation process is as follows: the reinforcement learning agent adjusts the optimization direction or weight parameter according to the current environment state, and the NSGA-II maintains the diversity and Pareto optimality of the solution set in population evolution. The strategy network of the reinforcement learning module can dynamically adjust the crossover and mutation probability of the genetic algorithm according to the real-time performance index to realize the balance between global search and local optimization; the NSGA-II module iteratively evolves the candidate scheme population, and filters out the Pareto frontier solution set according to the multi-objective dominance relationship. On this basis, the optimization algorithm can adaptively explore the multi-objective solution space and converge to a high-quality solution through the synergistic effect of the two, and finally output an optimal restoration scheme suggestion meeting the predetermined priority or balanced strategy.

[0055] In an example embodiment, in the process of determining the restoration scheme decision result, first, the problem is described; for example, for each ecological sub-region Selecting a scheme and continuous parameters wherein, represents whether the sub-region r selects a scheme ; if yes, the value is 1; if no, the value is 0; is a continuous parameter of the scheme on the ecological sub-region r, such as restoration area and input intensity; secondly, the target vector is defined; specifically, the target vector can be shown in the following formula (15): ; Equation (15) wherein, is a target vector, , , , and are five objectives to be optimized, the negative sign represents “maximization”, and the non-negative sign represents “minimization”; is an ecological benefit expectation value, for example, vegetation coverage growth; is an economic cost expectation value, for example, repair fund input; is a conditional expectation of risk value, is a confidence level, is a loss, which can be used to measure the risk of the scheme; is a carbon sink expectation value, which is a preset ecological index; is a water conservation expectation value, which is a preset ecological index; the constraint condition for defining the target vector is ; wherein, k is the number of constraint conditions, such as budget upper limit, time limit upper limit, accessibility, ecological red line, and land use upper limit, etc.

[0056] Then, the inner constraint reinforcement learning (Lagrange-RL) is performed; specifically, the continuous parameter (that is, the aforementioned ) and the time sequence action are regarded as the output of the strategy , and the environment is an ecological evolution simulation / proxy model; wherein, is a reinforcement learning strategy network, is a network parameter, which can be used to output the continuous parameter and the time sequence action; on this basis, the multi-objective optimization decision and the candidate ecological restoration scheme are input into the first scheme optimization model to dynamically adjust the candidate ecological restoration scheme to obtain the first optimization result; the specific implementation process can be shown in the following formula (16): ; Equation (16) wherein, is the state of the environment at the t time, for example, the current ecological index and the constraint satisfaction condition, etc.; is the action at the t time, if the restoration intensity is adjusted, etc. is the immediate reward of the action, such as the reward brought by the ecological benefit improvement; is the discount factor at t time, which can be used to measure the weight of future rewards; is the upper limit of the kth constraint, such as the budget upper limit, etc.

[0057] Meanwhile, the Lagrange relaxation is constructed; specifically, it can be shown by the following formula (17): ; formula (17) is the Lagrange multiplier of the kth constraint, which can be used to balance the weights of the objective and constraints. On this basis, the Actor-Critic update and (the Lagrange multiplier of the constraint), is obtained ; wherein, is the optimal continuous parameter after selecting x for the given scheme, that is, the first optimization result.

[0058] Further, external NSGA-II / NSGA-III search is performed to determine the second optimization result ; specifically, in the actual application process, the first optimization result is evaluated to perform non-dominated sorting and crowding degree maintenance to obtain the Pareto front; wherein, is the objective vector evaluated by the optimal parameter , which can be used to perform NSGA-II / III sorting; then based on the Pareto front, the set of all non-dominated solutions (that is, the comprehensive optimal scheme set, which can be specifically referred to as shown in Figure 8 ) is obtained, which is taken as the second optimization result; for example, RL-guided initialization and adaptive crossover / mutation (adjusted by ) can be introduced.

[0059] Finally, the robustness and risk assessment of the second optimization is performed to obtain the repair scheme decision result; specifically, the and chance constraint can be introduced; specifically, it can be referred to as shown in the following formula (18): , ; formula (18) wherein, is the repair scheme decision result; is the second optimization result; is the upper limit of CVaR, which can be used to control the maximum acceptable risk of the scheme; is the upper limit of the constraint violation probability, for example, the probability of budget overruns needs to be less than or equal to ; used to represent the probability of meeting the constraint should be at least, based on which the feasibility of the scheme can be guaranteed.

[0060] In step S130, the ecological evolution trend of the restoration scheme decision result is simulated and evaluated based on the preset ecological spatiotemporal prediction model to obtain the ecological environment prediction result of the target geographical region in the future time period.

[0061] Specifically, the specific determination process of the ecological environment prediction result can be implemented in the following manner: constructing physical equation parameters and discrete and state space parameters, and training a long sequence time series prediction model to be trained based on the physical equation parameters and the discrete and state space parameters to obtain the preset ecological spatiotemporal prediction model; inputting the restoration scheme decision result into the preset ecological spatiotemporal prediction model for simulation and evaluation of the ecological evolution trend to obtain an ecological evolution region evaluation result; and correcting the ecological evolution region evaluation result based on a preset data assimilation algorithm to obtain the ecological environment prediction result of the target geographical region in the future time period. That is, in the actual application process, the specific determination process of the ecological environment prediction result can be implemented in the manner of PDE-constrained spatiotemporal ecological index prediction + data assimilation. On this basis, for the selected restoration scheme, the ecological evolution trend after the implementation of the scheme can be simulated and evaluated by using the ecological spatiotemporal prediction model; wherein the ecological spatiotemporal prediction model recorded herein is a deep time series network Informer, which can be optimized for long-term sequence prediction; in the actual prediction process, the Informer model introduces sparse self-attention mechanism (Prob-Sparse Self-attention) and self-attention distillation technology, effectively reduces the computational complexity of long sequence modeling, and can efficiently predict the changes of ecological indexes in a long time span while ensuring the accuracy.

[0062] In the example embodiments of the present disclosure, the ecological spatiotemporal prediction model takes the historical ecological index sequence (such as vegetation index NDVI, soil moisture, biodiversity index, etc.) obtained by remote sensing inversion and auxiliary data such as climate and soil as input, and outputs the ecological index prediction values at multiple future time points through time embedding and multi-head self-attention mechanism; therefore, the model can evaluate the improvement effect and trend change of the selected restoration scheme on the key indicators of the ecological system in a certain future period, such as the growth rate of vegetation coverage, the recovery curve of water conservation capacity, etc. The specific process is as follows: first, a PDE (Partial Differential Equation) of reaction-diffusion is constructed; wherein the specific construction process is as follows: first, assume that the vegetation coverage at position x at time t is , the soil moisture content at position x at time t is , and the water quality / nutrient salt concentration at position x at time t is ; On this basis, the partial differential equation PDE obtained can be shown as formula (19): ; Formula (19) Wherein, is the first derivative of vegetation coverage, is the first derivative of soil moisture content, is the first derivative of water quality / nutrient salt concentration; is the diffusion coefficient of vegetation, is the diffusion coefficient of soil moisture, is the diffusion coefficient of water quality, and each diffusion coefficient can vary with the spatial position x, which is used to describe the spatial propagation ability of the index; is the vegetation growth rate; is the vegetation environmental carrying capacity, which can vary with the spatial position x, and can also be used to represent the maximum vegetation coverage of the region; is the difference between the vegetation diffusion coefficients of two adjacent time points, is the difference between the soil moisture diffusion coefficients of two adjacent time points, is the difference between the vegetation diffusion coefficients of two adjacent time points; is the vegetation decay rate, which can vary with the spatial position x, and which can be used to represent the decrease of vegetation in this space due to pests and degradation; is the promotion coefficient of the vegetation by the repair measure, which can vary with the spatial position x, and u is the repair field intensity; is the vegetation consumption coefficient of soil moisture, and the more vegetation, the more water consumption; is the rainfall at time t, which is input data; is the evaporation at position x at time t, which can vary with space and time; is the decay coefficient of water quality / nutrient salt, such as the decrease caused by natural purification; is the intensity of the pollutant emission source at position x at time t, which is input data; x is the boundary condition, which can be referred to as the Neumann boundary, that is, the diffusion flux of the index on the boundary is 0, avoiding edge effects; is the repair intensity field intensity (decided by , ).

[0063] Secondly, discretize (graph Laplace L) on the graph G and backward Euler time step Δt; Wherein, the specific expression can be shown as formula (20)-formula (22): ; Formula (20) ​; Equation (21) ; Equation (22) wherein, is the graph Laplacian of vegetation, is the graph Laplacian of soil moisture; is the graph Laplacian of water quality, which can be used to characterize the graph discrete process of PDE; is the time step, which can be used to characterize the time interval that needs to be predicted, for example, it can be 1 month; is the vegetation state at time t, is the soil moisture state at time t, is the water quality state at time t, all of which are discrete values; is the dot product process; is the repair field intensity at time t, which can be determined by the position x and the contact parameter θ in the repair scheme; , and is the process noise at time t, which can be used to simulate natural fluctuations and model errors.

[0064] Then, the observation model is constructed to perform remote sensing inversion; the observation model can be specifically as shown in the following formula (23): , ; Equation (23) wherein, is the observation value of remote sensing inversion, C is the observation operator, is the observation noise; indicates that the observation noise obeys the normal distribution with mean 0 and covariance R.

[0065] Further, the physical-data joint training is performed; specifically, in the process of training, the PINN / weak constraint mode can be used to realize; for example, the neural network approximation , , minimizes to determine the physical loss; wherein the specific expression can be as shown in the following formula (24): ; Equation (24) wherein, , and are the vegetation state, soil moisture state and water quality state predicted by the Informer model; is the physical consistency loss, which needs to ensure that the predicted vegetation state, soil moisture state and water quality state satisfy the PDE rule; , and The reaction-diffusion term on the right side of the PDE equation (such as vegetation growth, soil water consumption, and water pollution, etc.); at the same time, the observation loss is minimized; wherein the specific expression of the observation loss can be shown in the following formula (25): ; Formula (25) Wherein, The observation loss needs to ensure that the prediction is consistent with the remote sensing observation data; The remote sensing observation data; The weighted Euclidean norm, the weight is the inverse of the observation covariance R, which can highlight the contribution of high-precision observations. On this basis, based on the physical consistency loss and the observation loss, the Informer model is predicted to obtain the preset ecological spatiotemporal prediction model; during the training process, the observation loss can also be configured with a weight value, and the specific weight value size can be set according to actual needs, and this example does not make special restrictions.

[0066] Further, the ecological evolution trend simulation evaluation result of the ecological evolution region evaluation result is obtained by inputting the repair scheme decision result into the preset ecological spatiotemporal prediction model; on this basis, based on the preset data assimilation algorithm, the ecological evolution region evaluation result is corrected to obtain the ecological environment prediction result of the target geographical region in the future time period; wherein the data assimilation algorithm recorded herein is the ensemble Kalman filter (EnKF, Ensemble Kalman Filter) data assimilation algorithm, which can be used for prediction-correction of the ecological evolution region evaluation result; wherein the prediction-correction process recorded herein can be based on the following formula (26)-formula (28): ; Formula (26) ; Formula (27) ; Formula (28) Wherein, The Kalman gain (control the correction amplitude of the observation data to the prediction) at time t; The prediction error covariance (describe the uncertainty of the prediction) at time t; The analysis error covariance (uncertainty after assimilation) at time t; The predicted state (before assimilation) at time t; The analysis state (after assimilation, more accurate prediction result) at time t; to obtain the distributed prediction (that is, the ecological environment prediction result), which can be used for E and CVaR evaluation in robust optimization.

[0067] In step S140, the repair scheme decision result is optimized according to the ecological environment prediction result to obtain a target ecological repair scheme, and the target geographical region is repaired according to the target ecological repair scheme.

[0068] In the example embodiment, first, the repair scheme decision result is optimized according to the ecological environment prediction result to obtain a target ecological repair scheme. Specifically, the ecological environment expected result of the target geographical region in a future time period under the action of the repair scheme decision result can be determined, and the ecological environment expected result and the ecological environment prediction result are compared to obtain an ecological environment comparison result. If the ecological environment comparison result meets a preset condition, the repair scheme decision result is taken as the target ecological repair scheme of the target geographical region. If the ecological environment comparison result does not meet the preset condition, the repair scheme decision result is optimized until the ecological environment comparison result meets the preset condition, and the optimized repair scheme decision result is taken as the target ecological repair scheme of the target geographical region. That is, in actual application, after obtaining the ecological environment prediction result, a closed-loop feedback and adaptive adjustment process needs to be performed to form a diagnosis-matching-optimization-prediction closed-loop technical link to realize adaptive decision optimization. That is, the prediction evaluation result (i.e., the ecological environment prediction result) and the target expected value (i.e., the ecological environment expected result) are compared and analyzed. If it is found that the selected scheme is insufficient for improving some indicators or new bottleneck problems occur, the target weight or constraint condition of the prediction feedback adjustment optimization module can be adjusted to select or generate an improved scheme in the candidate scheme again until the requirement is met. The closed-loop feedback mechanism enables the system to continuously correct the decision according to the effect of simulation prediction, thereby improving the reliability and robustness of the scheme making. In practice, the latest ecological changes are obtained by continuously monitoring remote sensing data, which are integrated into the next round of diagnosis analysis to realize the rolling iteration of adaptive repair strategy optimization.

[0069] In an example embodiment, in the process of comparison and analysis, the multi-dimensional time sequence indicators can be normalized according to the target period to construct a utopia point The time sequence integrated vector of the scheme is calculated again, and the weighted distance is calculated. The specific calculation process of the weighted distance can be shown in the following formula (29): Formula (29) Wherein, is the weighted distance, is the utopia point, which can be used to represent the ideal optimal value of each target (such as budget, construction period, etc.) as a benchmark for scoring; is the time sequence integrated vector of the scheme (i.e., the repair scheme decision result), and has: ; is a multi-dimensional time sequence vector index of the scheme (i.e., the restoration scheme decision result) at time t, for example, an ecological prediction result (i.e., an ecological environment prediction result) at time t, and T is the overall prediction length. It should be further noted that the weighted distance can be used as an order reference in the Pareto frontier; on this basis, the final target ecological restoration scheme can be determined; wherein the obtained target ecological restoration scheme can refer to Figure 9 ; finally, the target geographical area can be restored based on the obtained target ecological restoration scheme.

[0070] It should be further noted that a total loss function can also be set to train the overall model. The total loss function recorded herein can be as shown in the following formula (30): ; formula (30) wherein, is the total loss function, is the parameter norm and sparse regularization, is the full model parameter; and is a specific weight value.

[0071] At this point, the ecological environment restoration scheme determination method recorded in the example embodiments of the present disclosure has been fully implemented. In the following, the ecological environment restoration scheme determination method will be further explained and described in combination with specific embodiments. Specifically, the first step is data and preprocessing: obtaining multi-source remote sensing (multi-spectral / hyperspectral satellite, unmanned aerial vehicle), DEM, LiDAR, soil and meteorological data, completing radiation / geometric correction, registration and cloud and haze removal, calculating NDVI / NDWI / EVI, slope / curvature and other derived layers; constructing graph weight and sparsification (KNN / threshold). The second step is to train the segmentation model; using the "encoder-mixed layer-decoding head" structure, the number of mixed layers L = 3-6, the Chebyshev order K = 2-3, and the multi-head attention H = 4-8; the loss seg in , , , , The parameters are adjusted by the validation set to take into account the boundaries and small patches; the class imbalance uses the ∈[1.5, 2.5]. Step 3: Cross-modal matching: Region embedding is achieved by concatenating patch tokens, graph readout, and statistical indicators (patch area, shape, connectivity, historical trends); Scheme embedding is generated jointly by a scheme text Transformer and a structured parametric encoder. Sinkhorn iterative solution. , Use a value of 0.05–0.5; assign a score. Joint optimization with InfoNCE / ranking loss. Step 4: Two-layer optimization: Outer layer NSGA-II (population 200–500, generations 50–100), the objective F is estimated from the distribution output of the PDE / prediction module, along with the expected value and CVaR; the inner constraint RL uses PPO / Lagrange-PPO, with constraints starting with soft penalties and switching to hard thresholds after convergence, and the feasible domain of policy θ is limited by engineering specifications and construction conditions. Step 5: PDE-PINN + EnKF prediction: PDE coefficient fields Dv, Dw, Dq, rv, μ, κ can be set as functions of terrain / soil / vegetation type; PINN is trained with weak constraint loss Lpred and observation consistency; the assimilation period (e.g., monthly / seasonal) is fused with the latest remote sensing inversion values ​​using EnKF, outputting the mean and covariance to provide uncertainty quantification for optimization. Step 6: Results Output and Loop Closure: Develop recommended solutions (single or a set) and indicator prediction curves (mean ± confidence interval). When the deviation between observation and prediction exceeds the threshold, automatic re-optimization is triggered, and the loop converges.

[0072] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the embodiments of the method disclosed herein.

[0073] This disclosure also provides an apparatus for determining an ecological environment restoration plan. Specifically, refer to... Figure 10 As shown, the device for determining the ecological environment restoration plan may include an ecological state identification result determination module 1010, a restoration plan decision result determination module 1020, an ecological environment prediction result determination module 1030, and a target ecological restoration plan determination module 1040. Wherein: The ecological state recognition result determination module 1010 can be configured to preprocess the multi-modal data corresponding to the target geographical region to obtain standard input data, and input the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result. The ecological restoration scheme decision result determination module 1020 can be configured to determine a candidate ecological restoration scheme according to the ecological state recognition result, and perform multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result. The ecological environment prediction result determination module 1030 can be configured to simulate and evaluate the ecological evolution trend of the restoration scheme decision result based on a preset ecological space-time prediction model, to obtain an ecological environment prediction result of the target geographical region in a future time period. The target ecological restoration scheme determination module 1040 can be configured to optimize the restoration scheme decision result according to the ecological environment prediction result, to obtain a target ecological restoration scheme, and perform ecological restoration on the target geographical region according to the target ecological restoration scheme.

[0074] The specific details of each module in the above ecological environment restoration scheme determination apparatus have been described in detail in the corresponding ecological environment restoration scheme determination method, and therefore will not be described here.

[0075] It should be noted that although several modules or units of the device for action execution are mentioned in the above detailed description, such division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into several modules or units for embodiment. In addition, although the steps of the method in the present disclosure are described in a specific order in the drawings, this does not require or imply that the steps must be performed in this specific order, or that all the steps shown must be performed to achieve the desired results. In addition or alternatively, some steps can be omitted, a plurality of steps can be combined into one step, or one step can be divided into a plurality of steps, etc.

[0076] Further, the above-described diagrams are merely schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, and are not intended for limiting purposes. It is readily understood that the processes shown in the above-described diagrams do not indicate or limit the time sequence of the processes. In addition, it is also readily understood that the processes can be executed, for example, synchronously or asynchronously in a plurality of modules. Other embodiments of the present disclosure will be readily apparent to those skilled in the art in view of the specification and practice of the inventive concepts disclosed herein. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure following the general principles thereof and including the general knowledge or custom of the art not specifically mentioned in the present disclosure. The specification and examples are to be regarded as exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

Claims

1. A method for determining an ecological environment remediation scheme, characterized in that, The method comprises the following steps: preprocessing multi-modal data corresponding to a target geographical area to obtain standard input data, and inputting the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; determining a candidate ecological restoration scheme according to the ecological state recognition result, and performing multi-objective optimization decision on the candidate ecological restoration scheme to obtain a restoration scheme decision result; based on a preset ecological space-time prediction model, simulating and evaluating the ecological evolution trend of the restoration scheme decision result to obtain an ecological environment prediction result of the target geographical area in a future time period; optimizing the restoration scheme decision result according to the ecological environment prediction result to obtain a target ecological restoration scheme, and performing ecological restoration on the target geographical area according to the target ecological restoration scheme.

2. The method of claim 1, wherein the ecological environment restoration scheme is determined based on the following steps: The preprocessing of the multi-modal data corresponding to the target geographical area to obtain the standard input data comprises: ​ obtaining multi-modal data corresponding to the target geographical area; wherein the multi-modal data comprises at least one of multi-source remote sensing data, geographic information data, laser detection data, and soil and meteorological data; dividing the target geographical area into grids to obtain a plurality of region division results, and determining region attribute data of each region division result according to the multi-modal data corresponding to each region division result; constructing an original region graph structure with each region division result as a node, the region attribute data as the node attribute of each node, and the adjacency relationship between each region division result as an edge; determining the weight value of the edge according to the node attribute, and updating the original region graph structure according to the weight value to obtain a target region graph structure, so as to determine the standard input data according to the target region graph structure and multi-source remote sensing data.

3. The method for determining the ecological environment restoration plan according to claim 2, characterized in that, The determination of the region attribute data of each region division result according to the multi-modal data corresponding to each region division result comprises: performing radiation correction and / or geometric correction on the multi-source remote sensing data corresponding to each region division result to obtain first standard remote sensing data; wherein the multi-source remote sensing data comprises at least one of multi-spectral remote sensing data, hyperspectral remote sensing data, and unmanned aerial vehicle remote sensing data; performing cloud and haze removal processing on the first standard remote sensing data to obtain second standard remote sensing data, and determining a regional remote sensing index of each region division result according to the second standard remote sensing data; registering the first standard remote sensing data and the geographic information data corresponding to each region division result to obtain an image alignment result, and determining a topographic parameter of each region division result according to the image alignment result; determining a laser detection derived layer of each region division result according to the laser detection data corresponding to each region division result, and determining the region attribute data of each region division result according to the regional remote sensing index, the topographic parameter, the laser detection derived layer, and the soil and meteorological data corresponding to each region division result.

4. The method for determining an ecological environment restoration plan according to claim 1, characterized in that, The preset remote sensing image segmentation model comprises a region feature extraction model and a region classification model; wherein the standard input data is input into the preset remote sensing image segmentation model to obtain an ecological state recognition result, comprising: Based on the region feature extraction model, the second standard remote sensing data corresponding to the multi-source remote sensing data in the standard input data is subjected to key feature extraction at multiple different scales to obtain feature extraction results corresponding to each region division result; Based on the region classification model, the feature extraction results and the target region graph structure in the standard input data are clustered and reasoned to obtain an ecological state recognition result; wherein the ecological state recognition result is used to represent the degraded plot distribution and / or vegetation coverage of the target geographical region.

5. The method for determining an ecological environment restoration plan according to claim 1, characterized in that, According to the ecological state recognition result, a candidate ecological restoration scheme is determined, comprising: According to the ecological state recognition result, the target geographical region is divided into multiple ecological sub-regions, and a preset original ecological restoration scheme is obtained; Based on the preset cross-modal matching model, the preset original ecological restoration scheme and the ecological recognition result of each ecological sub-region are cross-modally matched to obtain a matching score; Based on the matching score, the required candidate ecological restoration scheme for each ecological sub-region is determined from the preset original ecological restoration scheme.

6. The method of claim 5, wherein the ecological environment restoration scheme is determined based on the following steps: The preset cross-modal matching model is obtained by: ​ Using contrast learning or triple loss to train the feature extraction network of the image modality and the text modality under preset constraints to obtain a preset vector embedding model; wherein the preset constraints are that the first distance of the region division result and the original ecological restoration scheme of the same category in the embedding space is less than the second distance of the region division result and the original ecological restoration scheme of different categories in the embedding space.

7. The method of claim 1, wherein the ecological environment restoration scheme is determined based on the following conditions: The candidate ecological restoration scheme is subjected to multi-objective optimization decision to obtain a restoration scheme decision result, comprising: ​ A multi-objective optimization decision corresponding to the candidate ecological restoration scheme is constructed; wherein the multi-objective optimization decision comprises multiple of ecological benefit optimization decision, economic cost optimization decision, implementation cycle optimization decision and social influence optimization decision; The candidate ecological restoration scheme is input into a first scheme optimization model to dynamically adjust the candidate ecological restoration scheme to obtain a first optimization result; The multi-objective optimization decision and the first optimization result are input into a second scheme optimization model to screen the candidate ecological restoration scheme to obtain a second optimization result; The second optimization result is input into a third scheme optimization model to predict the risk of the second optimization result to obtain a restoration scheme decision result.

8. The method of claim 1, wherein the ecological environment restoration scheme is determined based on the following conditions: Based on the preset ecological spatio-temporal prediction model, the ecological evolution trend of the restoration scheme decision result is simulated and evaluated to obtain an ecological environment prediction result of the target geographical region in a future time period, comprising: ​ Physical equation parameters and discrete and state space parameters are constructed, and a long sequence time series prediction model to be trained is trained based on the physical equation parameters and the discrete and state space parameters to obtain the preset ecological spatio-temporal prediction model; inputting the repair scheme decision result into the preset ecological spatiotemporal prediction model to simulate and evaluate an ecological evolution trend, to obtain an ecological evolution region evaluation result; correcting the ecological evolution region evaluation result based on a preset data assimilation algorithm to obtain an ecological environment prediction result of the target geographical region in a future time period.

9. The method of claim 1, wherein the ecological environment restoration scheme is determined based on the following conditions: optimizing the repair scheme decision result according to the ecological environment prediction result to obtain a target ecological repair scheme, including: ​ determining an ecological environment expected result of the target geographical region in the future time period under the action of the repair scheme decision result, and comparing the ecological environment expected result with the ecological environment prediction result to obtain an ecological environment comparison result; if the ecological environment comparison result meets a preset condition, taking the repair scheme decision result as the target ecological repair scheme of the target geographical region; if the ecological environment comparison result does not meet the preset condition, optimizing the repair scheme decision result until the ecological environment comparison result meets the preset condition, and taking the optimized repair scheme decision result as the target ecological repair scheme of the target geographical region.

10. A device for determining an ecological environment restoration plan, characterized in that, including: an ecological state recognition result determination module, configured to pre-process multi-modal data corresponding to a target geographical region to obtain standard input data, and input the standard input data into a preset remote sensing image segmentation model to obtain an ecological state recognition result; a repair scheme decision result determination module, configured to determine a candidate ecological repair scheme according to the ecological state recognition result, and perform multi-objective optimization decision on the candidate ecological repair scheme to obtain a repair scheme decision result; an ecological environment prediction result determination module, configured to simulate and evaluate an ecological evolution trend of the repair scheme decision result based on a preset ecological spatiotemporal prediction model, to obtain an ecological environment prediction result of the target geographical region in a future time period; a target ecological repair scheme determination module, configured to optimize the repair scheme decision result according to the ecological environment prediction result to obtain a target ecological repair scheme, and perform ecological repair on the target geographical region according to the target ecological repair scheme.

Images