Remote sensing data optimization method

Abstract

This application relates to the field of data optimization technology and provides a method for optimizing remote sensing data. The method includes: obtaining a preliminary screening dataset of remote sensing images based on user-input query conditions; inputting the query conditions and the preliminary screening dataset into a data optimization model to obtain a optimized dataset output by the data optimization model; the data optimization model is obtained by training a deep learning generative network using historical query conditions and corresponding historical preliminary screening datasets and historical optimized datasets; the historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method. This application can achieve efficient optimization of massive amounts of remote sensing data with large spatial range and long time series, meeting the time constraints of online services.

Description

Technical Field

[0001] This application relates to the field of data optimization technology, specifically to a method for optimizing remote sensing data. Background Technology

[0002] Currently, remote sensing satellite data is generally stored and organized on a scene-by-scene basis. By cataloging the metadata information of remote sensing data and inputting it into a spatial data engine for unified storage and management, users can then use human-computer interaction applications to input relevant attribute information such as satellite, sensor, time range, and spatial range. With the help of efficient attribute-spatial joint indexing technology, they can quickly filter out the cataloged data information that meets the conditions.

[0003] However, with the rapid development of satellite remote sensing technology, the amount of remote sensing data has exploded. Faced with long-term, multi-source remote sensing satellite data stored and organized by scene unit, traditional spatiotemporal attribute-based retrieval methods result in a massive number of retrieval results and a large amount of duplicate data. When users further optimize data based on data quality indicators, they often need to perform a lot of manual comparison and evaluation operations, which is very time-consuming. In some cases, multiple people may spend several weeks manually selecting data, resulting in low optimization efficiency.

[0004] While existing automated remote sensing data optimization methods have largely eliminated manual selection, their algorithms mostly revolve around a problem that is both computationally and data-intensive, requiring multiple recursive traversals. When optimizing massive amounts of remote sensing data with large spatial ranges and long time series, the execution efficiency of these algorithms still cannot meet the time constraints imposed by online services. Summary of the Invention

[0005] This application provides a remote sensing data optimization method to address the technical problem that existing automated remote sensing data optimization methods cannot meet the online service's time constraints when optimizing massive amounts of remote sensing data with large spatial ranges and long time series.

[0006] This application provides a method for optimizing remote sensing data, including: Based on the query conditions input by the user, a preliminary dataset of remote sensing images is obtained; Input the query conditions and the initial screening dataset into the data optimization model to obtain the optimized dataset output by the data optimization model; The data optimization model is obtained by training a deep learning generative network using historical query conditions and the corresponding historical preliminary screening dataset and historical optimized dataset; the historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method.

[0007] In one embodiment, the data optimization model is constructed based on the following: The historical query conditions and the historical preliminary screening dataset are input into the deep learning generative network to obtain the predicted preferred dataset output by the deep learning generative network. The deep learning generative network is trained with the goal of minimizing the loss value between the predicted preferred dataset and the historical preferred dataset to obtain the data optimization model.

[0008] In one embodiment, the step of inputting the historical query conditions and the historical preliminary screening dataset into the deep learning generative network to obtain the predicted preferred dataset output by the deep learning generative network includes: Vectorize the target area information in the historical query conditions and the spatial coverage information of each remote sensing image in the historical preliminary screening dataset to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image. The non-spatial information of each remote sensing image in the historical preliminary screening dataset is vectorized to obtain non-spatial features; The single-scene spatial coverage features and the non-spatial features are stitched together to obtain the single-scene full features of each remote sensing image. The set of single-scene full features in the historical preliminary screening dataset is taken as the full features of the historical preliminary screening dataset. The target region spatial features and the full features are input into the deep learning generation network to obtain the prediction optimization dataset output by the deep learning generation network.

[0009] In one embodiment, the step of vectorizing the target area information in the historical query conditions and the spatial coverage information of each remote sensing image in the historical preliminary screening dataset to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image includes: Obtain the spatial filtering range of the target area in the historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical preliminary screening dataset; both the spatial filtering range and the spatial coverage area are located within a preset global spatial area. The entire spatial region is divided into grids to generate a reference grid system; The spatial filtering range and the spatial coverage area are taken as geographic objects to be vectorized, and spatial intersection operations are performed with the baseline grid system to obtain the set of grid cells covered by the geographic object and the grid index list corresponding to the set of grid cells. Based on the grid index list, vector encoding is performed on the dimensions of the reference grid system to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image.

[0010] In one embodiment, the step of vectorizing the non-spatial information of each remote sensing image in the historical initial screening dataset to obtain non-spatial features includes: Based on the calculation method used in the multi-constraint combination optimization method, the time information and cloud cover information of each remote sensing image in the historical preliminary screening dataset are vectorized to obtain time features and cloud cover features.

[0011] In one embodiment, the deep learning generative network includes an encoder and a decoder; the step of inputting the spatial features of the target region and the full features into the deep learning generative network to obtain the prediction preferred dataset output by the deep learning generative network includes: The target region spatial features and the full features are input into the encoder to obtain the encoded features output by the encoder; The encoded features are input into the decoder to obtain the preferred prediction dataset output by the decoder.

[0012] In one embodiment, the encoder is used for: The spatial features of the target region and the single-scene spatial coverage features in the full features are compressed and reduced in dimensionality to obtain low-dimensional spatial features. Based on the attention mechanism, the spatial overlay relationship between the spatial filtering range of the target area in the historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical initial screening dataset is modeled to obtain the target overlay weighted feature of each remote sensing image. The target superimposed weighted features, the non-spatial features in the full features, and the low-dimensional spatial features corresponding to the target region spatial features are fused to obtain the comprehensive features; The integrated features are input into the scene context fusion feature extraction model to model the interdependence of each scene remote sensing image in terms of spatial and non-spatial attributes, and generate encoded features that fuse global context information.

[0013] In one embodiment, the target overlay weighted features of each remote sensing image are obtained based on the following method: The low-dimensional spatial features corresponding to the spatial features of the target region are extended along the sequence dimension of the historical preliminary screening dataset to obtain extended features that match the sequence length of the historical preliminary screening dataset. The extended feature is combined with the low-dimensional spatial feature corresponding to the single-scene spatial coverage feature to obtain the spatial splicing feature. The spatial stitching features are input into a learnable attention network, which outputs the spatial overlay relationship weights between the spatial coverage area of ​​each remote sensing image in the historical preliminary screening dataset and the spatial filtering range of the target area in the historical query conditions. Based on the spatial overlay relationship weights, the low-dimensional spatial features corresponding to the spatial coverage features of each single scene are weighted and scaled to obtain the target overlay weighted features of each remote sensing image.

[0014] In one embodiment, the decoder is used for: Simulating the multi-constraint combinatorial optimization strategy employed by the aforementioned multi-constraint combinatorial optimization method to generate the predicted optimal dataset includes: The multi-constraint combinatorial optimization strategy includes sequential decision-making optimization strategy and optimization strategy based on global solution representation; In simulating the sequential decision-making optimization strategy, the encoded features are progressively and recursively decoded using an autoregressive approach. At each time step, based on the context of the remote sensing images that have been selected into the historical preferred dataset in the historical preliminary screening dataset, the ID of the next remote sensing image to be selected into the historical preferred dataset is predicted, and the predicted preferred dataset is generated sequentially. In simulating the optimization strategy based on the global solution representation, the remote sensing data optimization task is modeled as a binary classification problem using a non-autoregressive approach. Based on the encoded features, the probability of each remote sensing image in the historical preliminary screening dataset being selected into the historical optimization dataset is calculated in parallel. Based on the relationship between the probability and the probability threshold, the predicted optimization dataset is generated.

[0015] In one embodiment, inputting the variable-length historical preliminary screening dataset into the deep learning generative network includes: Based on the output length threshold of the deep learning generative network, the historical preliminary screening dataset is length-aligned to obtain the aligned historical preliminary screening dataset. A corresponding location mask is set for each data position in the aligned historical preliminary screening dataset to obtain a location mask array; the location mask array includes a valid location mask for identifying valid images and an invalid location mask for identifying invalid images; The aligned historical preliminary screening dataset and the location mask array are input into the deep learning generative network.

[0016] The remote sensing data optimization method provided in this application obtains a preliminary screening dataset of remote sensing images based on user-input query conditions. The query conditions and the preliminary screening dataset are then input into a data optimization model to obtain the optimized dataset output by the model. This model is trained on a deep learning generative network using historical query conditions and corresponding historical preliminary screening and optimized datasets. The historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method. In this application, the deep learning generative network, during training, fully utilizes its powerful learning capabilities to learn the optimization patterns of the multi-constraint combinatorial optimization method. It performs function simulation and behavioral approximation of the set coverage solution process under multiple constraints. Furthermore, when optimizing the preliminary screening dataset, it leverages the massively parallel computing capabilities of GPUs to achieve large-scale parallel neural inference for the originally computationally intensive and data-intensive problem, thereby effectively improving data optimization efficiency. This method enables efficient optimization of massive amounts of remote sensing data with large spatial ranges and long time series, meeting the time constraints of online services. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the remote sensing data optimization method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the data optimization model construction process provided in the embodiments of this application; Figure 3 This is a schematic diagram of the process of using data vectorization representation to output a predicted preferred dataset, provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the prediction of the optimal data subset from a historical preliminary screening dataset, provided in an embodiment of this application. Figure 5 This is a schematic diagram of the spatial information vectorization process provided in the embodiments of this application; Figure 6 This is a flowchart illustrating the process of a deep learning-generated network outputting a preferred dataset for the encoder-decoder architecture provided in this application embodiment. Figure 7 This is a schematic diagram of the process of inputting the variable-length historical initial screening dataset into the deep learning generation network after masking, as provided in the embodiments of this application. Figure 8This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] It should be noted that in the description of the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The terms "upper," "lower," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, or integral connections; they can be mechanical connections or electrical connections; they can be direct connections or indirect connections through an intermediate medium; and they can be internal connections between two elements. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0021] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects have an "or" relationship.

[0022] Figure 1 This is a flowchart illustrating the remote sensing data optimization method provided in an embodiment of this application. (Refer to...) Figure 1This application provides a method for optimizing remote sensing data, which may include: Step 101: Based on the query conditions input by the user, obtain the initial screening dataset of remote sensing images; Step 102: Input the query conditions and the initial screening dataset into the data optimization model to obtain the optimized dataset output by the data optimization model.

[0023] The data optimization model is obtained by training a deep learning generative network using historical query conditions and the corresponding historical preliminary screening datasets and historical optimized datasets. The historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method.

[0024] When optimizing remote sensing data, the optimization problem can be transformed into a set coverage problem and then solved using multi-constraint combinatorial optimization methods. However, practice has shown that this solution process revolves around a high-time-complexity problem with both computationally intensive and data-intensive aspects. When optimizing massive amounts of remote sensing data with large spatial ranges and long time series, it is difficult to meet the online service's time constraints.

[0025] In this embodiment, the data optimization model is more efficient than the multi-constraint combinatorial optimization method. If a greedy algorithm is used for remote sensing data optimization, it is typically implemented through an iterative strategy, resulting in a time complexity of polynomial level (e.g., ...). If integer programming methods are used, they rely on implicit enumeration strategies such as branch and bound, and their worst-case time complexity increases exponentially with the size of the initial dataset. Both of these methods struggle to effectively utilize parallel computing hardware such as GPUs for acceleration. In scenarios with large spatial scales and long time series, response latency often reaches the minute level, failing to meet the second-level response requirements of online services. Data optimization models, however, can fully leverage the massively parallel computing capabilities of networks using GPUs to achieve large-scale parallel neural inference for problems with both computationally intensive and data-intensive computations, thus effectively improving the data optimization efficiency of data networks. This enables efficient optimization of massive amounts of remote sensing data with large spatial scales and long time series, meeting the second-level response requirements of online services.

[0026] Figure 2 This is a schematic diagram illustrating the process of constructing the data optimization model provided in an embodiment of this application. (Refer to...) Figure 2 In one embodiment, the data optimization model can be constructed based on the following: Step 201: Input the historical query conditions and the historical preliminary screening dataset into the deep learning generative network to obtain the predicted optimal dataset output by the deep learning generative network. Step 202: With the goal of minimizing the loss between the predicted preferred dataset and the historical preferred dataset, train the deep learning generative network to obtain the data optimization model.

[0027] In step 201, historical query conditions can be pre-constructed and stored in a spatial database to generate the corresponding initial screening dataset. Specifically, the construction method can involve randomly setting spatial ranges, satellite payloads, time ranges, cloud cover thresholds, etc., to form a query condition. Taking province A in China as an example: A municipal administrative region in province A is randomly selected as the spatial range filtering condition. One to five Gaofen series satellite-sensor combinations are randomly selected as the satellite sensor filtering condition. The time span of the query is randomly set to three months, six months, one year, one and a half years, or two years. The start time of the query is randomly set within the valid data acquisition time interval, and the end time of the query is calculated based on the time span. The time between the start time and the end time is the time range filtering condition. 20% or 30% is randomly set as the cloud cover threshold filtering condition. These filtering conditions together form a historical query condition and are stored in the spatial database as a record.

[0028] Based on the generated historical query conditions, the corresponding historical preliminary screening dataset is generated through the spatial database engine.

[0029] Furthermore, a deep learning generative network can be a combination architecture of encoder and decoder. Historical query conditions and historical preliminary screening datasets can be input into the encoder for encoding to obtain the encoded features output by the encoder. The encoded features are then input into the decoder for decoding to obtain the predicted optimal dataset output by the decoder.

[0030] In step 202, the multi-constraint combinatorial optimization method for generating the historical optimal dataset can include two multi-constraint combinatorial optimization strategies: one is a sequential decision-making optimization strategy, such as a greedy algorithm, and the other is an optimization strategy based on global solution representation, such as integer programming algorithms or genetic algorithms. Using these two multi-constraint combinatorial optimization methods, based on the same historical initial screening dataset, we can obtain a historical optimal dataset with sequential dependencies and a globally optimal historical optimal dataset. Pairing this historical initial screening dataset with the two sets of historical optimal datasets yields two paired datasets. Training the deep learning generative network using these two paired datasets allows the network to fully learn the optimization patterns of the two multi-constraint combinatorial optimization methods, thereby generating a model that can simulate the efficient optimization capabilities of these two optimization strategies.

[0031] Furthermore, the paired dataset and historical query conditions can be used as sample datasets, divided into training and test datasets according to a certain ratio, and stored in the corresponding tables of the spatial database. Specifically, for sample datasets generated based on the optimization strategy of global solution representation, the sample dataset table needs to store the ID information of remote sensing images in the historical preliminary screening dataset, the record ID of the corresponding historical query conditions, and the label information of whether the remote sensing image exists in the corresponding historical optimal dataset, which can be represented by 0 to indicate non-existence and 1 to indicate existence; for sample datasets generated based on the sequential decision-making optimization strategy, in addition to storing the above information, the sample dataset table also stores the optimization order information of the remote sensing images in each historical preliminary screening dataset in the corresponding historical optimal dataset.

[0032] With the above information, we can use the training dataset and the test dataset to efficiently train the deep learning generative network and obtain a data-optimized model with better performance.

[0033] Furthermore, the historical preferred dataset can be input into the decoder to calculate the loss value between the predicted preferred dataset and the historical preferred dataset. If the loss value is greater than or equal to the threshold, the encoder and / or decoder parameters are adjusted, and the process returns to step 201 to regenerate the predicted preferred dataset and calculate the loss value until the loss value is less than the threshold or the preset number of training iterations is reached. The deep learning generative network at this point is then determined as the data preferred model.

[0034] In this embodiment, a deep learning generative network with an encoder-decoder architecture is used to encode and decode the training samples. The network is trained with the goal of minimizing the loss between the output predicted optimal dataset and the pre-calculated historical optimal dataset. This allows the network to fully learn the optimization patterns of the historical optimal dataset, making its output infinitely close to the historical optimal dataset. This transforms the multi-constraint combinatorial optimization method into a neural network forward inference process. The trained model can significantly reduce the response latency of the optimization task and meet the stringent real-time requirements of online services.

[0035] Figure 3 This is a schematic diagram illustrating the process of using data vectorization representation to output a predicted preferred dataset, as provided in an embodiment of this application. (Refer to...) Figure 3 In one embodiment, step 201 may include: Step 301: Vectorize the target area information in the historical query conditions and the spatial coverage information of each remote sensing image in the historical preliminary screening dataset to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image. Step 302: Vectorize the non-spatial information of each remote sensing image in the historical preliminary screening dataset to obtain non-spatial features; Step 303: Stitch together the spatial coverage features and non-spatial features of a single scene to obtain the full features of each remote sensing image. Step 304: Take the set of full features of a single scene in the historical preliminary screening dataset as the full features of the historical preliminary screening dataset, input the spatial features of the target region and the full features into the deep learning generative network, and obtain the prediction optimization dataset output by the deep learning generative network.

[0036] In this embodiment, in order to facilitate the training of the encoder and decoder, the input data information needs to be converted into a vector form suitable for network computation in order to improve the efficiency and accuracy of network computation; the main conversions include spatial information and non-spatial information.

[0037] In step 301, for historical query conditions, the spatial information needs to be vectorized. This spatial information is the spatial filtering range information input by the user, that is, the target area information. For the historical preliminary screening dataset, the spatial information (i.e., spatial coverage information) of each remote sensing image contained therein needs to be vectorized so that the spatial relationship between the target area and each remote sensing image can be captured later.

[0038] Step 302 may specifically include: Based on the computational method used in the multi-constraint combinatorial optimization method, the temporal information and cloud cover information of each remote sensing image in the historical preliminary screening dataset are vectorized to obtain temporal features and cloud cover features.

[0039] Since the multi-constraint combinatorial optimization method used in this embodiment considers temporal and cloud cover information for non-spatial information, the process of vectorizing the non-spatial information of each remote sensing image in the historical initial screening dataset also needs to follow the calculation method for temporal and cloud cover information in the multi-constraint combinatorial optimization method. By using the non-spatial information in the historical query conditions, the temporal and cloud cover features can be obtained so that the multi-constraint combinatorial optimization method can be simulated and learned more accurately during subsequent network training.

[0040] In this embodiment, the multi-constraint combination optimization method defines the median value of the time filtering range in the historical query conditions as the target time, and prioritizes remote sensing images closer to the target time as preferred data. That is, the closer the imaging time of the remote sensing image is to the target time, the higher the probability of it being selected as preferred data. The time information of each remote sensing image in the historical preliminary screening dataset can be vectorized in the following way: No. The imaging time of the remote sensing image is Calculate its time relative to the target time Time difference : ; The time difference is converted into a time vectorized value using a linear normalization method to obtain the time feature. : ; in, and These represent the maximum and minimum time differences between the imaging time and the target time for each remote sensing image in the historical initial screening dataset. (Normalized) The larger the value, the further away it is from the target time.

[0041] Similarly, cloud cover information for each remote sensing image in the historical initial screening dataset can be vectorized to obtain cloud cover features using the following method. : ; in, Let be the original cloud cover value of the i-th remote sensing image. and These represent the maximum and minimum cloud cover values ​​for each remote sensing image in the historical initial screening dataset, respectively. After normalization... A larger value indicates more clouds.

[0042] It should be noted that, as mentioned above, the multi-constraint combinatorial optimization method includes two types. Taking the greedy algorithm and the integer programming algorithm as examples, since the two algorithms implemented in this embodiment use the same calculation method for time information and the same calculation method for cloud cover information, this embodiment can follow the same calculation method to vectorize the time information and cloud cover information. The resulting time features are the same for both the greedy algorithm and the integer programming algorithm, and the resulting cloud cover features are also the same for both the greedy algorithm and the integer programming algorithm. If the multi-constraint combinatorial optimization method uses other specific calculation methods for time information or cloud cover information, it is necessary to follow the specific calculation methods to vectorize the time information and cloud cover information to obtain the corresponding time features and cloud cover features.

[0043] It should be noted that if there are other indicators that need to be vectorized, they should also follow the calculation method used in the simulation of multi-constraint combinatorial optimization.

[0044] This step follows the calculation methods used for time information and cloud cover information in various multi-constraint combinatorial optimization methods. The time information and cloud cover information of each remote sensing image in the historical preliminary screening dataset are vectorized, so that the obtained time features and cloud cover features have the same pattern as various multi-constraint combinatorial optimization methods, so that the network can more accurately simulate and learn various multi-constraint combinatorial optimization methods during subsequent training.

[0045] In step 303, the single-scene spatial coverage features and non-spatial features of each remote sensing image are stitched together to obtain the single-scene full features of each remote sensing image in the historical preliminary screening dataset. The single-scene full features include rich information including spatial features and non-spatial features.

[0046] In step 304, the spatial features of the target area and the single-scene full features of each remote sensing image in the historical preliminary screening dataset are input into a deep learning network. The network captures the refined spatial correlation between the spatial features of the target area and each remote sensing image, and simultaneously captures the refined spatial and non-spatial correlation between the remote sensing images in the historical preliminary screening dataset. This outputs a predicted optimal dataset that closely approximates the historical optimal dataset. This predicted optimal dataset is the optimal subset of data from the historical preliminary screening dataset that has complete target area coverage, excellent data quality, and low redundancy. Figure 4 As shown.

[0047] This embodiment vectorizes the target area information and vectorizes and stitches together the spatial coverage information and non-spatial information of each remote sensing image in the historical preliminary screening dataset. This allows for the acquisition of spatial features of the target area and full features of each remote sensing image. These features are then input into a deep learning generative network, which quickly captures the fine-grained correlation between the spatial features of the target area and the spatial and non-spatial features of each remote sensing image. Finally, it predicts and outputs the optimal data subset that fully covers the target area.

[0048] Figure 5 This is a schematic diagram illustrating the spatial information vectorization process provided in an embodiment of this application. (Refer to...) Figure 5 In one embodiment, step 301 may include: Step 501: Obtain the spatial filtering range of the target area in the historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical preliminary screening dataset; Both the spatial filtering range and the spatial coverage area are located within the preset global spatial area; Step 502: Divide the entire spatial region into grids to generate a baseline grid system; Step 503: Take the spatial filtering range and spatial coverage area as geographic objects to be vectorized, and perform spatial intersection operation with the baseline grid system to obtain the set of grid cells covered by the geographic object and the grid index list corresponding to the set of grid cells. Step 504: Based on the grid index list, perform vector encoding on the dimensions of the benchmark grid system to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image.

[0049] In step 501, the spatial filtering range is the range covered by the spatial range filtering conditions in the historical query conditions, and the spatial coverage area of ​​each remote sensing image is the area covered by each remote sensing image; neither the spatial filtering range nor the spatial coverage area of ​​each remote sensing image can exceed the preset global spatial area.

[0050] In step 502, to ensure that the spatial screening range and the spatial coverage area of ​​each remote sensing image are vectorized according to a unified standard, the entire spatial area needs to be divided into grids of a certain size to obtain a reference grid system under a regular grid framework. This reference grid system includes regularly arranged grid cells and their division and encoding methods. Each grid cell has a unified spatial resolution and geometric shape. These characteristics make the reference grid system a benchmark reference for subsequent spatial indexing and vectorization expression. For example, under an isometric cylindrical projection based on WGS84 (World Geodetic System 1984), such as the Plate Carrée projection, the area of ​​province A is divided into grids, where each grid cell is 0.025°×0.025° in size. Each grid cell is numbered starting from 0 according to a certain numbering order to generate the reference grid system.

[0051] In step 503, the spatial filtering range and the spatial coverage area of ​​each remote sensing image are spatially intersected with the reference grid system to obtain the set of grid cells covered by the spatial filtering range, the set of grid cells covered by the spatial coverage area of ​​each remote sensing image, and the grid index list corresponding to each grid cell in the set.

[0052] In step 504, for the set of grid cells covered by the spatial filtering range and the set of grid cells covered by the spatial coverage area of ​​each remote sensing image, vector encoding can be performed on each grid cell in the dimension of the reference grid system according to its corresponding grid index list, so as to obtain the vector encoding of all grids covered by the spatial filtering range, that is, the target area spatial features corresponding to the target area information, and the vector encoding of all grids covered by the spatial coverage area of ​​each remote sensing image, that is, the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image.

[0053] The vector encoding method can be selected according to actual needs and is not limited here. In this embodiment, sparse Boolean vector encoding can be used. That is, for each grid cell, if it is in the grid index list, the corresponding encoding bit is assigned a first preset value (such as "1"), otherwise it is assigned a second preset value (such as "0"). Then the spatial features of the target area and the spatial coverage features of a single scene are both high-dimensional sparse vectors.

[0054] This embodiment limits the spatial filtering range and the spatial coverage area of ​​each remote sensing image to a preset global spatial area. The global spatial area is then divided into grids to generate a reference grid system, providing a unified standard vectorization basis for the spatial filtering range and the spatial coverage area of ​​each remote sensing image. Then, the spatial filtering range and the spatial coverage area of ​​each remote sensing image are spatially intersected with the reference grid system to obtain the set of grid cells and a grid index list covered by both. Based on this grid index list, the grid cells covered by both can be traversed and encoded, achieving efficient and standardized unified spatial information vectorization of target area information and the spatial coverage information of each remote sensing image.

[0055] Figure 6 This is a schematic diagram illustrating the process of a deep learning-based generative network outputting a predicted dataset using the encoder-decoder architecture provided in this application embodiment. (Refer to...) Figure 6 In one embodiment, step 304 may include: Step 601: Input the target region features and full features into the encoder to obtain the encoded features output by the encoder; Step 602: Input the encoded features into the decoder to obtain the optimal prediction dataset output by the decoder.

[0056] The encoder is responsible for extracting the spatial features of the target region and the latent representation of all features in the latent space, while the decoder is responsible for generating and outputting the predicted optimal remote sensing image dataset based on these latent representations.

[0057] The encoder is specifically used for: Step 601a: Compress and reduce the spatial features of the target region and the single-scene spatial coverage features in the full features to obtain low-dimensional spatial features; Step 601b: Based on the attention mechanism, model the spatial overlay relationship between the spatial filtering range of the target area in the historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical initial screening dataset, and obtain the target overlay weighted feature of each remote sensing image. Step 601c: Fuse the target overlay weighted features, the non-spatial features in the full features, and the low-dimensional spatial features corresponding to the target region spatial features to obtain the comprehensive features; Step 601d: Input the integrated features into the scene context fusion feature extraction model to model the inter-scene remote sensing images in terms of spatial and non-spatial attributes, and generate encoded features that fuse global context information.

[0058] In step 601a, since the spatial features of the target region and the single-scene spatial coverage features in the full features are both high-dimensional sparse vectors, the spatial grid feature dimension in this embodiment can reach 32472. Therefore, a two-stage spatial projection strategy can be adopted to compress and reduce the dimensionality of these features, mapping them to a low-dimensional space. Specifically: The first stage involves using a linear layer to map both the spatial features of the target region and the spatial coverage features of a single scene to an intermediate high-dimensional space, resulting in an intermediate high-dimensional representation. The dimension of this intermediate high-dimensional space can be hidden_size×2, where hidden_size is the size of the hidden layer.

[0059] The intermediate high-dimensional representation can then be further processed. For example, nonlinearity can be introduced through the ReLU activation function, followed by a Dropout layer to prevent overfitting. The dropout probability of this layer can be 0.3. Finally, it can be normalized through the LayerNorm layer.

[0060] The second stage: Map the intermediate high-dimensional representation obtained in the first stage to the target low-dimensional space to obtain the low-dimensional space features. The dimension of the target low-dimensional space can be hidden_size. The low-dimensional features can then be further processed, for example, by using LayerNorm layers to maintain the stability of the feature distribution.

[0061] The design of this two-stage spatial projection strategy effectively solves the computational bottleneck of high-dimensional sparse feature processing. It can compress the original 32472-dimensional features to the hidden_size dimension, such as 512 dimensions, which significantly reduces the computational complexity while retaining key feature information. This allows the model to maintain its feature expressive power while greatly reducing the consumption of computational resources.

[0062] In step 601b, specifically, the low-dimensional spatial features corresponding to the spatial features of the target area can be extended along the sequence dimension of the historical preliminary screening dataset to obtain extended features that match the sequence length of the historical preliminary screening dataset. The extended features are then concatenated with the low-dimensional spatial features corresponding to the spatial coverage features of a single scene to obtain spatial concatenation features. The spatial concatenation features are then input into a learnable attention network, which outputs the spatial overlay relationship weights between the spatial coverage area of ​​each remote sensing image in the historical preliminary screening dataset and the spatial filtering range of the target area in the historical query conditions. Based on the spatial overlay relationship weights, the low-dimensional spatial features corresponding to the spatial coverage features of each single scene are weighted and scaled to obtain the target overlay weighted features of each remote sensing image.

[0063] In particular, by extending the low-dimensional spatial features corresponding to the spatial features of the target area along the sequence dimension of the historical preliminary screening dataset, extended features matching the sequence length of the historical preliminary screening dataset are obtained, which can provide a foundation for subsequent spatial association modeling with remote sensing images.

[0064] Since the extended features match the sequence length of the historical initial screening dataset and have the same dimension as the single-scene spatial coverage features of the remote sensing images, the extended features and the low-dimensional spatial features corresponding to the single-scene spatial coverage features can be sequentially concatenated to obtain spatially concatenated features with dimensions [B, L, 2]. [hidden_size], where B is the sample batch size and L is the sequence length of the historical preliminary screening dataset, i.e. the number of remote sensing images in the historical preliminary screening dataset.

[0065] Specifically, a learnable attention network can be used to calculate the spatial overlay relationship weights between the low-dimensional spatial features of single-scene spatial coverage features and the low-dimensional spatial features of the target region in the spatial stitching features. That is, the low-dimensional spatial features of single-scene spatial coverage features in the spatial stitching features are used as key vectors, and the low-dimensional spatial features of the target region in the spatial stitching features are used as query vectors. The spatial relevance score between the key vector and the query vector is calculated, and this score is converted into spatial overlay relationship weights through a normalization function. The dimension of this weight is [B, L, 1], and its value directly reflects the overlapping area of ​​the corresponding single-scene spatial coverage area and the target region in geographic space. The larger the weight value, the larger the overlapping area, which can enable the model to automatically focus on remote sensing images with high spatial relevance to the target region, significantly improving the accuracy of data selection.

[0066] Specifically, based on the spatial overlay relationship weight, the low-dimensional spatial features corresponding to the single-scene spatial coverage features are weighted and scaled to obtain the target overlay weighted features. The feature dimension is [B, L, hidden_size], thereby adaptively strengthening the single-scene spatial coverage features with a large overlay area with the target region, while suppressing the interference of single-scene spatial coverage features with a small overlay area or no overlap with the target region, thus achieving feature enhancement.

[0067] In step 601c, the target superimposed weighted features and extended features can be concatenated and fused through the feature fusion layer to obtain features with dimensions [B, L, hidden_size]. Then, these features are fused across dimensions with non-spatial features from all features, such as time features and cloud cover features, to obtain comprehensive features that include spatial correlation, temporal dynamics, and cloud cover conditions.

[0068] In step 601d, the scene context fusion feature extraction model can adopt a sequence neural network architecture, such as LSTM, GRU or Transformer sequence models, to extract highly discriminative encoded features such as spatial and non-spatial relationships between scenes from the historical initial screening dataset.

[0069] The decoder is specifically used for: Simulate the multi-constraint combinatorial optimization strategy employed by the multi-constraint combinatorial optimization method to generate a predictive optimization dataset, including: Multi-constraint combinatorial optimization strategies include sequential decision-making optimization strategies and optimization strategies based on global solution representation; Step 602a: Under the simulated sequence decision-making optimization strategy, the coded features are progressively and recursively decoded using an autoregressive approach. At each time step, based on the context of the remote sensing images that have been selected into the historical preferred dataset in the historical preliminary screening dataset, the ID of the next remote sensing image to be selected into the historical preferred dataset is predicted, and the predicted preferred dataset is generated in sequence. Step 602b: Under the simulation of the optimization strategy based on the global solution representation, the remote sensing data optimization task is modeled as a binary classification problem using a non-autoregressive approach. Based on the coding features, the probability of each remote sensing image in the historical preliminary screening dataset being selected into the historical optimization dataset is calculated in parallel. Based on the relationship between the probability and the probability threshold, the predicted optimization dataset is generated.

[0070] In step 602a, for the historical optimization dataset generated based on a sequential decision-making optimization strategy such as a greedy algorithm, there is a significant dependency between the various remote sensing images; that is, the earlier selected remote sensing image will affect the evaluation of subsequent remote sensing images. For this type of context-dependent optimization task, this embodiment uses an autoregressive approach to progressively and recursively decode the encoded features.

[0071] This mechanism maintains a dynamic selection state sequence. In each decoding step, the previously decoded preferred data is input back into the decoder as context information. Context-aware feature updates are used to model the conditional probabilities of remote sensing images in the historical initial screening dataset and progressively output the IDs of the selected remote sensing images. Specifically, at step t, the decoder receives the preferred data from the previous t-1 steps and their corresponding encoded features. A state transition is then performed to generate the hidden state for step t. An attention mechanism is used to calculate the conditional selection probability of remote sensing images in the historical initial screening dataset. Based on this, the ID of the selected remote sensing image in the historical preferred dataset is output at step t, progressively obtaining the predicted preferred dataset.

[0072] In step 602b, for historical optimization datasets generated using optimization strategies based on global solution representation, such as integer programming algorithms and genetic algorithms, there are no explicit dependencies between the various remote sensing images. For this type of dependency-free optimization task, this embodiment uses a non-autoregressive approach to model the remote sensing data optimization task as a binary classification problem. A classifier layer directly generates the probability of each remote sensing image in the initial historical screening dataset being selected for the historical optimization dataset. Based on a probability threshold, such as 0.5, remote sensing images with a probability greater than 0.5 are selected as preferred remote sensing images for the historical optimization dataset, thus obtaining the predicted optimization dataset.

[0073] This mechanism employs a single-layer linear transformation to map encoded features from the latent space to a one-dimensional output space. Subsequently, a nonlinear activation function unit, such as the sigmoid function, normalizes the output values ​​to the [0, 1] interval, forming a selection probability distribution. This probability characterizes the likelihood of each remote sensing image in the historical preliminary screening dataset being selected for the historical preferred dataset, thereby enabling parallel scoring and decision-making for each remote sensing image in the historical preliminary screening dataset, ultimately outputting a predicted preferred dataset.

[0074] In this embodiment, the encoder constructs a multi-source heterogeneous feature fusion mechanism. First, spatial features are compressed and reduced in dimensionality. Then, based on an attention mechanism, the spatial overlay relationship between the spatial filtering range of the target region in historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical initial screening dataset is modeled. The comprehensive features are then input into the inter-scene context fusion feature extraction model to model the interdependencies between each remote sensing image in spatial and non-spatial attributes. The attention mechanism enables the modeling of the spatial overlay relationship between each remote sensing image in the historical initial screening dataset and the target region, as well as the modeling of the interdependencies between each remote sensing image in the historical initial screening dataset. Dependency-based modeling not only preserves the independent semantics of each constraint but also achieves implicit interaction modeling between constraints through nonlinear transformations. This enables end-to-end learning of complex optimization strategy logic, avoiding problems such as high rule coupling, difficult parameter tuning, and weak generalization ability caused by explicit rule nesting in traditional methods. The decoder employs a dual-modal decision generation mechanism: for optimization tasks without prior dependencies, a one-time parallel probabilistic selection mechanism directly outputs the selection probability of each remote sensing image in the historical preliminary screening dataset; for optimization tasks with prior dependencies, an autoregressive iterative decoding strategy is used to gradually generate the optimization list, supporting flexible adaptation to different optimization logics. This embodiment fundamentally solves the technical challenge of inefficient execution of traditional optimization algorithms in "large-area, long-time-series" scenarios in remote sensing data optimization, achieving a leap from algorithm-level optimization to architecture-level acceleration.

[0075] Figure 7 This is a schematic diagram illustrating the process of inputting the variable-length historical initial screening dataset, after masking, into the deep learning generative network, as provided in the embodiments of this application. (Refer to...) Figure 7In one embodiment, step 201 may include: Step 701: Based on the output length threshold of the deep learning generative network, perform length alignment on the historical preliminary screening dataset to obtain the aligned historical preliminary screening dataset; Step 702: Set the corresponding position mask for each data position in the aligned historical preliminary screening dataset to obtain the position mask array; The location mask array includes a valid location mask for identifying valid images and an invalid location mask for identifying invalid images; Step 703: Input the aligned historical preliminary screening dataset and the location mask array into the deep learning generative network.

[0076] In step 701, the number of remote sensing images in the historical preliminary screening datasets of different samples may not be the same. Before vectorizing their information, the length of each historical preliminary screening dataset can be extended to the output length threshold of the deep learning generation network to achieve length alignment of each historical preliminary screening dataset.

[0077] In step 702, for each aligned historical preliminary screening dataset, a valid location mask is set for the original valid image locations, for example, 1, and an invalid location mask is set for the invalid image locations used for length padding, for example, 0, so that a location mask array composed of 1 and 0 can be obtained.

[0078] Furthermore, the number of valid images in the aligned historical preliminary screening dataset can be calibrated to ensure that the number of valid images is the same as the number of real images.

[0079] In step 703, the aligned historical preliminary screening dataset and the location mask array are input into the deep learning generator network. This allows the deep learning generator network to use the location mask array to dynamically identify and mask invalid images during the encoding stage, suppressing the interference of invalid images on the encoded features. During the decoding and loss calculation stages, the location mask array is further used to supervise the learning of valid images, thereby ensuring that model training and evaluation focus on valid images, allocating model computing resources to valid images, and achieving optimal resource allocation.

[0080] The above steps apply to any architecture that can dynamically adjust the computation based on the sequence length, including RNN variants such as GRU and LSTM, and attention architectures such as Transformer.

[0081] This embodiment addresses the dynamically changing size of the historical initial screening dataset by designing a variable-length sequence adaptive processing mechanism. Combined with mask awareness and dynamic packing techniques, it effectively ignores padding data during both training and inference phases, avoiding wasted computing resources and significantly improving GPU utilization. Simultaneously, end-to-end mask propagation ensures effective control throughout the input-output process, preventing padding data from interfering with attention weights and the final decision, resulting in higher stability and predictive consistency in real-world business scenarios.

[0082] Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call a computer program stored in the memory 830 to execute steps of the remote sensing data optimization method, such as: Based on the query conditions input by the user, a preliminary dataset of remote sensing images is obtained; Input the query conditions and the initial screening dataset into the data optimization model to obtain the optimized dataset output by the data optimization model; The data optimization model is obtained by training a deep learning generative network using historical query conditions and the corresponding historical preliminary screening dataset and historical optimized dataset; the historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method.

[0083] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the remote sensing data optimization method provided in the above embodiments, such as including: Based on the query conditions input by the user, a preliminary dataset of remote sensing images is obtained; Input the query conditions and the initial screening dataset into the data optimization model to obtain the optimized dataset output by the data optimization model; The data optimization model is obtained by training a deep learning generative network using historical query conditions and the corresponding historical preliminary screening dataset and historical optimized dataset; the historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method.

[0085] On the other hand, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, the computer program being used to cause a processor to execute the steps of the remote sensing data optimization method provided in the above embodiments, for example including: Based on the query conditions input by the user, a preliminary dataset of remote sensing images is obtained; Input the query conditions and the initial screening dataset into the data optimization model to obtain the optimized dataset output by the data optimization model; The data optimization model is obtained by training a deep learning generative network using historical query conditions and the corresponding historical preliminary screening dataset and historical optimized dataset; the historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method.

[0086] The non-transitory computer-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for optimizing remote sensing data, characterized in that, include: Based on the query conditions input by the user, a preliminary dataset of remote sensing images is obtained; Input the query conditions and the initial screening dataset into the data optimization model to obtain the optimized dataset output by the data optimization model; The data optimization model is obtained by training a deep learning generative network using historical query conditions and the corresponding historical preliminary screening dataset and historical optimized dataset; the historical optimized dataset is obtained by optimizing the historical preliminary screening dataset using a multi-constraint combinatorial optimization method.

2. The method for optimizing remote sensing data according to claim 1, characterized in that, The data optimization model is constructed based on the following method: The historical query conditions and the historical preliminary screening dataset are input into the deep learning generative network to obtain the predicted preferred dataset output by the deep learning generative network. The deep learning generative network is trained with the goal of minimizing the loss value between the predicted preferred dataset and the historical preferred dataset to obtain the data optimization model.

3. The method for optimizing remote sensing data according to claim 2, characterized in that, The step of inputting the historical query conditions and the historical preliminary screening dataset into the deep learning generative network to obtain the predicted optimal dataset output by the deep learning generative network includes: Vectorize the target area information in the historical query conditions and the spatial coverage information of each remote sensing image in the historical preliminary screening dataset to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image. The non-spatial information of each remote sensing image in the historical preliminary screening dataset is vectorized to obtain non-spatial features; The single-scene spatial coverage features and the non-spatial features are stitched together to obtain the single-scene full features of each remote sensing image. The set of single-scene full features in the historical preliminary screening dataset is taken as the full features of the historical preliminary screening dataset. The target region spatial features and the full features are input into the deep learning generation network to obtain the prediction optimization dataset output by the deep learning generation network.

4. The method for optimizing remote sensing data according to claim 3, characterized in that, The step of vectorizing the target area information in the historical query conditions and the spatial coverage information of each remote sensing image in the historical preliminary screening dataset to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image includes: Obtain the spatial filtering range of the target area in the historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical preliminary screening dataset; both the spatial filtering range and the spatial coverage area are located within a preset global spatial area. The entire spatial region is divided into grids to generate a reference grid system; The spatial filtering range and the spatial coverage area are taken as geographic objects to be vectorized, and spatial intersection operations are performed with the baseline grid system to obtain the set of grid cells covered by the geographic object and the grid index list corresponding to the set of grid cells. Based on the grid index list, vector encoding is performed on the dimensions of the reference grid system to obtain the target area spatial features corresponding to the target area information and the single-scene spatial coverage features corresponding to the spatial coverage information of each remote sensing image.

5. The method for optimizing remote sensing data according to claim 3, characterized in that, The process of vectorizing the non-spatial information of each remote sensing image in the historical initial screening dataset to obtain non-spatial features includes: Based on the calculation method used in the multi-constraint combination optimization method, the time information and cloud cover information of each remote sensing image in the historical preliminary screening dataset are vectorized to obtain time features and cloud cover features.

6. The method for optimizing remote sensing data according to claim 3, characterized in that, The deep learning generative network includes an encoder and a decoder; the step of inputting the spatial features of the target region and the full features into the deep learning generative network to obtain the prediction optimization dataset output by the deep learning generative network includes: The target region spatial features and the full features are input into the encoder to obtain the encoded features output by the encoder; The encoded features are input into the decoder to obtain the preferred prediction dataset output by the decoder.

7. The method for optimizing remote sensing data according to claim 6, characterized in that, The encoder is used for: The spatial features of the target region and the single-scene spatial coverage features in the full features are compressed and reduced in dimensionality to obtain low-dimensional spatial features. Based on the attention mechanism, the spatial overlay relationship between the spatial filtering range of the target area in the historical query conditions and the spatial coverage area of ​​each remote sensing image in the historical initial screening dataset is modeled to obtain the target overlay weighted feature of each remote sensing image. The target superimposed weighted features, the non-spatial features in the full features, and the low-dimensional spatial features corresponding to the target region spatial features are fused to obtain the comprehensive features; The integrated features are input into the scene context fusion feature extraction model to model the interdependence of each scene remote sensing image in terms of spatial and non-spatial attributes, and generate encoded features that fuse global context information.

8. The method for optimizing remote sensing data according to claim 7, characterized in that, The target overlay weighted features of each remote sensing image are obtained based on the following method: The low-dimensional spatial features corresponding to the spatial features of the target region are extended along the sequence dimension of the historical preliminary screening dataset to obtain extended features that match the sequence length of the historical preliminary screening dataset. The extended feature is combined with the low-dimensional spatial feature corresponding to the single-scene spatial coverage feature to obtain the spatial splicing feature. The spatial stitching features are input into a learnable attention network, which outputs the spatial overlay relationship weights between the spatial coverage area of ​​each remote sensing image in the historical preliminary screening dataset and the spatial filtering range of the target area in the historical query conditions. Based on the spatial overlay relationship weights, the low-dimensional spatial features corresponding to the spatial coverage features of each single scene are weighted and scaled to obtain the target overlay weighted features of each remote sensing image.

9. The method for optimizing remote sensing data according to claim 6, characterized in that, The decoder is used for: Simulating the multi-constraint combinatorial optimization strategy employed by the aforementioned multi-constraint combinatorial optimization method to generate the predicted optimal dataset includes: The multi-constraint combinatorial optimization strategy includes sequential decision-making optimization strategy and optimization strategy based on global solution representation; In simulating the sequential decision-making optimization strategy, the encoded features are progressively and recursively decoded using an autoregressive approach. At each time step, based on the context of the remote sensing images that have been selected into the historical preferred dataset in the historical preliminary screening dataset, the ID of the next remote sensing image to be selected into the historical preferred dataset is predicted, and the predicted preferred dataset is generated sequentially. In simulating the optimization strategy based on the global solution representation, the remote sensing data optimization task is modeled as a binary classification problem using a non-autoregressive approach. Based on the encoded features, the probability of each remote sensing image in the historical preliminary screening dataset being selected into the historical optimization dataset is calculated in parallel. Based on the relationship between the probability and the probability threshold, the predicted optimization dataset is generated.

10. The method for optimizing remote sensing data according to claim 2, characterized in that, The variable-length historical initial screening dataset is input into the deep learning generative network, including: Based on the output length threshold of the deep learning generative network, the historical preliminary screening dataset is length-aligned to obtain the aligned historical preliminary screening dataset. A corresponding location mask is set for each data position in the aligned historical preliminary screening dataset to obtain a location mask array; the location mask array includes a valid location mask for identifying valid images and an invalid location mask for identifying invalid images; The aligned historical preliminary screening dataset and the location mask array are input into the deep learning generative network.

Images