Scheduling method of remote sensing data interpretation task

Abstract

The invention provides a scheduling method for remote sensing data interpretation tasks. The task processing method comprises the following steps: obtaining to-be-executed task data and to-be-executed task user behavior data; obtaining a first interpretation algorithm recommendation set and a second interpretation algorithm recommendation set based on pre-stored historical task features and historical task user behavior features; solving an intersection of the first interpretation algorithm recommendation set and the second interpretation algorithm recommendation set, and selecting an algorithm from the first interpretation algorithm recommendation set or the second interpretation algorithm recommendation set as a recommendation interpretation algorithm of the to-be-executed task; assuming to call at least one execution machine to execute a plurality of to-be-executed tasks, taking the recommendation interpretation algorithm as a factor influencing the execution duration estimated value of the corresponding to-be-executed task, and taking the minimum total execution duration estimated value of the plurality of to-be-executed tasks as an optimization target, and calculating to obtain a resource scheduling optimal solution; and completing the scheduling of the interpretation task according to the resource scheduling optimal solution.

Description

Technical Field

[0001] This invention relates to task processing methods, and more particularly to a scheduling method for remote sensing data interpretation tasks. Background Technology

[0002] Remote sensing interpretation refers to the analysis and understanding of information such as images and data acquired through various remote sensing technologies. Interpretation algorithms are a crucial task in the field of remote sensing, aiming to extract information about targets and the environment from remote sensing images, such as information on target activity in key areas, geological and hydrological environments, etc., which is of great significance to various remote sensing image analysis efforts. They provide key decision support for intelligent governance of cities and natural environments. With the modernization of urban and natural environment governance capabilities, high efficiency and high quality have become essential requirements for remote sensing interpretation capabilities in the execution phase. To ensure interpretation quality and improve interpretation efficiency, the following two prerequisites need to be met:

[0003] 1. Select the most suitable interpretation algorithm for the application scenario to be executed, realize automated interpretation work, and ensure interpretation quality.

[0004] The conventional approach to existing interpretation algorithm tasks is to give users the right to choose the interpretation algorithm; or users write manual rules to select the appropriate algorithm, which requires manual rule writing and has high requirements for the accuracy of the rules.

[0005] 2. In scenarios with multiple tasks to be decoded, the conventional approach for existing decoding algorithms is to statistically analyze the executor's resource capacity and average execution time, and then comprehensively consider both factors to provide a scheduling scheme for computing resources for the tasks to be decoded. However, this approach can only achieve a local optimum.

[0006] In the prior art, such as Chinese invention patent publication number CN 117687799 A, a distributed streaming acceleration method and computing terminal for remote sensing interpretation applications are disclosed. This invention provides a distributed streaming acceleration method and computing terminal for remote sensing interpretation applications, relating to the field of remote sensing data processing technology, and can solve the problems of lack of flexibility and inefficient use of hardware resources in existing remote sensing data processing. The method includes: performing virtualization operations on a neural network processor (NPU) to decompose the NPU into multiple dynamically configurable virtual NPU cores; acquiring multiple remote sensing data streams to be processed, determining the task attributes of each remote sensing data stream, including detailed survey tasks and general survey tasks; and processing the multiple remote sensing data streams in parallel by multiple virtual NPU cores according to the task attributes. This invention utilizes streaming computing and NPU virtualization technology to significantly improve the efficiency of remote sensing data interpretation, responding to different task requirements through dynamic resource scheduling. However, the aforementioned prior art does not disclose how to optimize the algorithm model for interpretation tasks and ensure globally optimal total execution time. Summary of the Invention

[0007] To address the technical problems existing in the prior art, the present invention aims to provide a scheduling method for remote sensing data interpretation tasks, which can optimize the algorithm model for the interpretation task and ensure that the total execution time is globally optimal.

[0008] To achieve the above-mentioned objectives, this invention provides a scheduling method for remote sensing data interpretation tasks, comprising the following steps:

[0009] Obtain data on tasks to be executed and user behavior data for those tasks;

[0010] The task data to be executed and the user behavior data of the task to be executed are respectively based on the pre-stored historical task features and historical task user behavior features to obtain the first interpretation algorithm recommendation set and the second interpretation algorithm recommendation set;

[0011] Find the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms, and select an algorithm from the first set of recommended interpretation algorithms or the second set of recommended interpretation algorithms as the recommended interpretation algorithm for the task to be executed, based on whether the intersection is empty and the preset algorithm selection conditions.

[0012] Assuming that at least one execution machine is invoked to execute multiple tasks to be executed, and the recommended interpretation algorithm is used as a factor affecting the estimated execution time of the corresponding tasks to be executed, the optimal solution for resource scheduling is calculated with the goal of minimizing the estimated total execution time of the multiple tasks to be executed;

[0013] The number of executors is less than or equal to the number of tasks to be executed;

[0014] Based on the optimal resource scheduling solution, multiple tasks to be executed are assigned to their respective execution machines to complete the scheduling of the interpretation tasks.

[0015] According to one technical solution of the present invention, the process of obtaining historical task characteristics is as follows:

[0016] Obtain the task target scene label, task size, image parameters, algorithm operation indicators, user ID, interpretation algorithm used by the user, and the number of times each interpretation algorithm was used from historical tasks;

[0017] Through feature engineering, the task target scene label, task size and image parameters are vectorized to obtain historical task features;

[0018] Furthermore, by using task target scenario labels and user IDs as dimension labels, and statistically analyzing the interpretation algorithms used by users and the number of times each interpretation algorithm was used, a statistical feature vector is formed as the user behavior features of historical tasks.

[0019] According to a technical solution of the present invention, the process of obtaining the first interpretation algorithm recommendation set and the second interpretation algorithm recommendation set is as follows:

[0020] Obtain the task target scene label, task size, image parameters, algorithm operation indicators, user ID, the interpretation algorithm used by the user, and the number of times each interpretation algorithm is used in the task to be executed;

[0021] Through feature engineering, the task target scene label, task size and image parameters are vectorized to obtain the features of the task to be executed.

[0022] In addition, the task target scene label and user ID are used as dimension labels, and the interpretation algorithms used by users and the number of times each interpretation algorithm is used are counted to form a statistical feature vector as the user behavior features of the task to be executed;

[0023] Based on the user behavior characteristics of the task to be executed, the number of times each interpretation algorithm is used under the same user ID in the same task target scenario is found from multiple historical task user behavior characteristics; interpretation algorithms with a usage count greater than or equal to a preset usage count threshold are selected to construct a first interpretation algorithm recommendation set.

[0024] Calculate the second similarity between the features of the task to be executed and the features of multiple historical tasks to obtain multiple similar historical tasks;

[0025] The similarity history task is a historical task whose second similarity with the features of the task to be executed is greater than a similarity threshold;

[0026] We summarize all the interpretation algorithms corresponding to the historical similarity tasks and construct a second set of interpretation algorithm recommendations.

[0027] According to one technical solution of the present invention, it further includes:

[0028] Based on the user behavior characteristics of the task to be executed, find the number of times the same user ID has executed the historical task in the same target scenario from multiple historical task user behavior characteristics;

[0029] If the number of times the historical task is executed is less than the preset execution number threshold, then the first interpretation algorithm recommendation set is set to be empty.

[0030] According to one technical solution of the present invention, in the first set of recommended interpretation algorithms, the interpretation algorithms are sorted according to the number of times they are used;

[0031] In the second set of recommended interpretation algorithms, the interpretation algorithms are ranked according to their performance metrics.

[0032] The algorithm's performance metrics include accuracy and / or false alarm rate.

[0033] According to one technical solution of the present invention, the algorithm selection conditions are as follows:

[0034] When the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms is empty, the interpretation algorithm ranked first in the second set of recommended interpretation algorithms is taken as the recommended interpretation algorithm for the corresponding task to be executed.

[0035] When the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms is not empty, the interpretation algorithm ranked first in the intersection is selected as the recommended interpretation algorithm for the corresponding task to be executed, based on the ranking of the recommended interpretation algorithms in the first set of recommended interpretation algorithms.

[0036] According to a technical solution of the present invention, the process of calculating the optimal solution for resource scheduling is as follows:

[0037] The total estimated execution time of all tasks to be executed is calculated by estimating the execution time of each task on the corresponding execution machine; and the resource scheduling objective function is constructed by minimizing the total estimated execution time.

[0038] By using the particle swarm optimization method, and with the constraint that each task to be executed is assigned to only one executor, the optimal solution for resource scheduling is obtained by solving the resource scheduling objective function.

[0039] According to a technical solution of the present invention, the estimated execution time of each task to be executed on the corresponding execution machine is obtained as follows:

[0040] Based on the task parameters of the task to be executed and the hardware parameters of the executor, a first linear equation including the task parameters and hardware parameters is constructed as a model for predicting the execution time of the task to be executed.

[0041] Based on the task parameters of historical tasks and the hardware parameters of the executor, a second linear equation including the task parameters and hardware parameters is constructed as a model for predicting the execution time of historical tasks.

[0042] The task parameters include the sample size corresponding to the task to be executed, the parameter size of the corresponding execution time prediction model, the spatial resolution of the sample image, and the spectral resolution of the sample image; the hardware parameters include the memory, CPU clock frequency, CPU core count, GPU clock frequency, and number of GPU cards used in the execution machine.

[0043] The loss function of the historical task is calculated using the actual execution time of the historical task and the execution time prediction model of the historical task.

[0044] Based on multiple loss functions, the partial derivatives of different coefficients of the loss functions are calculated to obtain all coefficients in the historical task execution time prediction model;

[0045] Substitute all coefficients into the execution time prediction model of the task to be executed, and then calculate the estimated execution time of each task on the corresponding execution machine.

[0046] The present invention also provides an electronic device, comprising: one or more CPUs, one or more memories, and one or more computer programs; wherein the CPU is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the CPU executes the one or more computer programs stored in the memory to enable the electronic device to perform the scheduling method for the remote sensing data interpretation task described above.

[0047] The present invention also provides a computer-readable storage medium for storing computer instructions, which, when executed by a CPU, implement the scheduling method for the remote sensing data interpretation task described above.

[0048] The scheduling method for remote sensing data interpretation tasks of the present invention has the following advantages compared with the prior art:

[0049] 1. This invention intelligently analyzes the interpretation task scenario and user behavior, employing intelligent recommendation technology and multiple recommendation strategies to optimize the algorithm model for the interpretation task, providing the best interpretation algorithm. Compared to human experience and rules, it considers multiple perspectives to adapt a better algorithm for the interpretation task, ensuring interpretation quality.

[0050] 2. This invention employs an operations research approach, aiming to minimize the total execution time. It utilizes the particle swarm optimization algorithm to seek the globally optimal computing resource matching scheme for concurrent execution scheduling scenarios involving multiple interpretation tasks. The step-by-step approach effectively finds the globally optimal solution for computing resource matching, reducing task queuing time by an average of 20% and task inference computation time by 11% compared to existing task scheduling methods, resulting in a total saving of 12% of the overall task execution time and improving global interpretation efficiency. Furthermore, the solution process is highly efficient, supports parallelization, and is less prone to getting trapped in local optima, allowing for simultaneous optimal matching searches for different tasks. It reduces the queuing time and inference time of interpretation tasks, ensuring a globally optimal total execution time. This reduces the overall execution time of interpretation tasks while maintaining the quality of remote sensing image interpretation. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating a scheduling method for remote sensing data interpretation tasks according to an embodiment of the present invention;

[0053] Figure 2 The flowchart illustrates the process of calculating the optimal resource scheduling solution in a remote sensing data interpretation task scheduling method according to an embodiment of the present invention. Detailed Implementation

[0054] The description of the embodiments in this specification should be taken in conjunction with the accompanying drawings, which should form part of the complete specification. In the drawings, the shape or thickness of the embodiments may be exaggerated and may be indicated in a simplified or convenient manner. Furthermore, parts of the various structures in the drawings will be described separately; it is worth noting that elements not shown in the figures or not described in words are in a form known to those skilled in the art.

[0055] The descriptions of the embodiments herein, including any references to directions and orientations, are for ease of description only and should not be construed as limiting the scope of the invention. The following description of preferred embodiments involves combinations of features, which may exist independently or in combination; the invention is not particularly limited to the preferred embodiments. The scope of the invention is defined by the claims.

[0056] like Figures 1-2 As shown in the figure, a method for scheduling remote sensing data interpretation tasks according to this embodiment includes the following steps:

[0057] S1. Obtain data on tasks to be executed and user behavior data for tasks to be executed;

[0058] S2. Using the data of tasks to be executed and the user behavior data of tasks to be executed, based on the pre-stored historical task features and historical task user behavior features, respectively, the first interpretation algorithm recommendation set and the second interpretation algorithm recommendation set are obtained;

[0059] S3. Find the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms, and select an algorithm from the first set of recommended interpretation algorithms or the second set of recommended interpretation algorithms as the recommended interpretation algorithm for the task to be executed, based on whether the intersection is empty and the preset algorithm selection conditions.

[0060] S4. Assuming that at least one execution machine is called to execute multiple tasks to be executed, and the recommended interpretation algorithm is used as a factor affecting the estimated execution time of the corresponding tasks to be executed, the optimal solution for resource scheduling is calculated with the goal of minimizing the estimated total execution time of the multiple tasks to be executed.

[0061] The number of execution machines is less than or equal to the number of tasks to be executed;

[0062] S5. Based on the optimal resource scheduling solution, assign multiple tasks to be executed to their respective execution machines to complete the scheduling of the interpretation tasks.

[0063] In this embodiment, the general process of a scheduling method for remote sensing data interpretation tasks is as follows:

[0064] S1. Obtain the task data to be executed, including the task target scene label (target type, whether there is cloud or fog obstruction), task scale (image scale), image parameters (image size, spatial resolution, spectral resolution, satellite payload, projection method, signal-to-noise ratio, number and range of bands, shooting angle), and algorithm operation indicators.

[0065] And user behavior data for tasks to be executed, including the interpretation algorithm used by the user, user ID, and other information.

[0066] S2. A recommendation strategy based on the similarity of interpretation task features is adopted. Combining historical interpretation task features and the algorithms used, candidate recommendation results based on objective algorithm indicators are obtained, namely the first interpretation algorithm recommendation set.

[0067] It also employs a user-based collaborative filtering recommendation strategy to obtain candidate recommendation results from the perspective of user habits, which is the second interpretation algorithm recommendation set.

[0068] S3. Integrate the above subjective and objective recommendation results to formulate the final recommendation results.

[0069] S4. Computational Resource Scheduling. This scenario involves several interpretation tasks awaiting execution and several task execution machines with available computing resources. An operations research approach is adopted, aiming to minimize the total execution time, and the particle swarm optimization algorithm is used to quickly find the optimal resource scheduling solution.

[0070] S5. Start executing the interpretation task based on the optimal resource scheduling solution, then end.

[0071] In the scheduling method for remote sensing data interpretation tasks, the process of obtaining historical task features is as follows:

[0072] Obtain the task target scene label, task size, image parameters, algorithm operation indicators, user ID, interpretation algorithm used by the user, and the number of times each interpretation algorithm was used from historical tasks;

[0073] Through feature engineering, the task target scene label, task size and image parameters are vectorized to obtain historical task features;

[0074] Furthermore, by using task target scenario labels and user IDs as dimension labels, and statistically analyzing the interpretation algorithms used by users and the number of times each interpretation algorithm was used, a statistical feature vector is formed as the user behavior features of historical tasks.

[0075] This embodiment specifically includes:

[0076] Step 1: Historical mission data preparation. Historical mission data mainly includes mission target scene labels (target type, presence of cloud or fog obstruction), mission scale (image size), image parameters (image size, spatial resolution, spectral resolution, satellite payload, projection method, signal-to-noise ratio, number and range of bands, shooting angle), interpretation algorithm used by the user, algorithm performance indicators, user ID, etc., forming a historical interpretation mission dataset (historical dataset).

[0077] Step 2: Develop feature engineering based on historical datasets to construct historical task features and historical task user behavior features.

[0078] (1) Historical task features: Based on the target scene label information, task scale information and image parameter information in the historical dataset, each interpretation task feature in the historical dataset is vectorized, and the continuous value features and label features are combined to form a feature vector (where the label features are formed by one-hot encoding), thus forming the interpretation task features.

[0079] (2) Historical Task User Behavior Characteristics: Based on historical datasets, the selection of interpretation algorithms by each user in each target scenario is statistically analyzed. Using the target scenario label + user ID as dimension labels, the number of times each algorithm is used is counted to form a statistical feature vector.

[0080] In the scheduling method for remote sensing data interpretation tasks, the process of obtaining the first interpretation algorithm recommendation set and the second interpretation algorithm recommendation set is as follows:

[0081] Obtain the task target scene label, task size, image parameters, algorithm operation indicators, user ID, the interpretation algorithm used by the user, and the number of times each interpretation algorithm is used in the task to be executed;

[0082] Through feature engineering, the task target scene label, task size and image parameters are vectorized to obtain the features of the task to be executed.

[0083] In addition, the task target scene label and user ID are used as dimension labels, and the interpretation algorithms used by users and the number of times each interpretation algorithm is used are counted to form a statistical feature vector as the user behavior features of the task to be executed;

[0084] Based on the user behavior characteristics of the task to be executed, the number of times each interpretation algorithm is used under the same user ID in the same task target scenario is found from the user behavior characteristics of multiple historical tasks; interpretation algorithms with a usage count greater than or equal to a preset usage count threshold are selected to construct the first interpretation algorithm recommendation set.

[0085] Calculate the second similarity between the features of the task to be executed and the features of multiple historical tasks to obtain multiple similar historical tasks;

[0086] Historical tasks with similarity are those with a second similarity to the features of the task to be performed that is greater than the similarity threshold.

[0087] We summarize all the interpretation algorithms corresponding to the historical similarity tasks and construct a second set of interpretation algorithm recommendations.

[0088] In this implementation, intelligent algorithmic recommendations are performed based on task features and user behavior features. A fusion recommendation strategy based on user behavior collaborative filtering and task feature similarity is employed to ensure the objectivity and rationality of the recommendations while also considering user habits. The algorithmic recommendation results, which integrate objectivity, rationality, and user habits, are output and used as part of the interpretation of the algorithm's task parameters. The recommendation process is divided into:

[0089] Step 1: Collaborative filtering recommendation based on user behavior.

[0090] Based on the user ID and task target scenario label in the user behavior characteristics of the task to be executed, the number of times each interpretation algorithm was used by the user in the target scenario to be executed in history is obtained from the historical user behavior characteristics. Interpretation algorithms with a usage count greater than or equal to K1 (the usage count threshold) are selected to form the first interpretation algorithm recommendation set S1.

[0091] Step 2: Recommendation based on task feature similarity.

[0092] Obtain task features, calculate the cosine similarity between the features of the task to be executed and the features of historical tasks, select the top N (N≥1) historical tasks with similarity > M (similarity threshold) to form a set of similar tasks, select the interpretation algorithms selected in the set, sort them from highest to lowest according to the performance index of the interpretation algorithms, and form a second set of recommended interpretation algorithms S2.

[0093] Step 3: Merge the recommendation results. Solve for the intersection of S1 and S2.

[0094] If the intersection is empty, only the recommendation results based on task feature similarity are considered, and the TOP1 algorithm result of S2 is used as the final recommendation result.

[0095] If the intersection is not empty, sort the intersection according to S1, and use the TOP1 algorithm of the intersection as the final recommendation result.

[0096] The scheduling methods for remote sensing data interpretation tasks also include:

[0097] Based on the user behavior characteristics of the tasks to be executed, find the number of times the same user ID has executed a historical task in the same target scenario from multiple historical task user behavior characteristics;

[0098] If the number of times a historical task is executed is less than a preset execution threshold, then the set recommended by the first interpretation algorithm is set to be empty.

[0099] In this embodiment, in the collaborative filtering recommendation based on user behavior in the previous step 1, if the number of historical interpretation tasks executed for the current user ID is less than K2 (execution count threshold), no recommendation is made, that is, the recommendation set S1 of the first interpretation algorithm is empty.

[0100] In the scheduling method for remote sensing data interpretation tasks, the first set of recommended interpretation algorithms is sorted by the number of times they are used.

[0101] In the second set of recommended interpretation algorithms, the interpretation algorithms are ranked according to their performance metrics.

[0102] Algorithm performance metrics include accuracy and / or false alarm rate.

[0103] In the scheduling method for remote sensing data interpretation tasks, the algorithm selection criteria are as follows:

[0104] When the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms is empty, the interpretation algorithm ranked first in the second set of recommended interpretation algorithms is taken as the recommended interpretation algorithm for the corresponding task to be executed.

[0105] When the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms is not empty, the interpretation algorithm ranked first in the intersection is selected as the recommended interpretation algorithm for the corresponding task to be executed, based on the ranking of the recommended interpretation algorithms in the first set of recommended interpretation algorithms.

[0106] In the scheduling method for remote sensing data interpretation tasks, the process of calculating the optimal solution for resource scheduling is as follows:

[0107] The total estimated execution time of all tasks to be executed is calculated by estimating the execution time of each task on the corresponding execution machine; and the resource scheduling objective function is constructed by minimizing the total estimated execution time.

[0108] By using the particle swarm optimization method, and with the constraint that each task to be executed is assigned to only one executor, the optimal solution for resource scheduling is obtained by solving the resource scheduling objective function.

[0109] In this implementation, the task scheduling is based on operations research and optimization, using the particle swarm optimization algorithm to provide a task scheduling scheme for n tasks to be executed. It includes three sub-steps: problem definition, duration prediction model, and particle swarm optimization solution, as follows:

[0110] Step 1, Problem Definition:

[0111] Assuming there are n tasks to be executed and m available execution machines (each with varying performance and remaining resources), this problem is solved using operations research optimization methods, as follows:

[0112] Decision variables: Define a binary decision matrix X, where X... ij This indicates whether task i is assigned to execution machine j. If task i is assigned to execution machine j, then X... ij =1, otherwise 0.

[0113] Objective function: Minimize execution time. Where n is the number of tasks and m is the number of execution machines.

[0114]

[0115] Constraint: Each task can only be assigned one execution machine.

[0116] Step 2: Construct a task execution time prediction model using linear regression (the execution time prediction model is the same for historical tasks and tasks to be executed). Calculate the execution time prediction value for each task to be executed using historical tasks with known exact times and the execution time prediction value of historical tasks obtained based on the execution time prediction model.

[0117] Step 3: Solve for the variable decision matrix X using the particle swarm optimization method:

[0118] 1) Particle initialization:

[0119] Given the constraints, k particles are randomly generated, each representing a scheduling scheme. Each particle is represented by an n*m two-dimensional decision vector matrix X, where the matrix represents the particle's position information, and each element represents the assignment relationship between a task and an executor. If task i is assigned to executor j, then X... ij =1.

[0120] 2) Initialization of particle velocity matrix:

[0121] Initialize the velocity matrix V for each particle, with a size of N*M.

[0122] 3) Calculate the initial fitness value (estimated overall execution time) for each particle; a lower value indicates better performance. Based on the time prediction model, estimate the execution time of each task i on its assigned execution machine j, and sum these values ​​to obtain the fitness value.

[0123]

[0124] 4) Update the particle velocity matrix:

[0125] Update the position (decision) vector of each particle. For each particle, update its velocity and position (decision) vectors based on its pending position, velocity, optimal position for individual fitness, and optimal position for global fitness. The velocity update formula is as follows:

[0126]

[0127] in Let ω represent the velocity of particle i in the d-th dimension at the t-th iteration, where ω represents the particle's inertia; c1 and c2 represent the learning rate, and r1 and r2 represent random factors, ranging from [0,1]. Let d represent the value of the d-th dimension of particle i when its fitness is at its best within iteration t. The value of the d-th dimension of a particle when its fitness is globally optimal within t iterations. Let represent the value of the d-th dimension of particle i in the t-th iteration.

[0128] 5) Update particle position (decision) information based on the velocity matrix:

[0129] The formula for updating the position (decision) vector of particle i is as follows, updating the value of each dimension in turn.

[0130]

[0131] 6) After updating the position vector (decision vector), recalculate the fitness of each particle. Update the optimal position (decision) of each particle and the global optimal position (decision) based on the fitness. When a certain number of iterations or the global optimal particle fitness function value converges, output the global optimal position (decision) vector as the scheduling matching result. Otherwise, repeat (4) and (5) until the output condition is met.

[0132] Step 4, Interpretation Task Execution: Extract the solution results and execute the interpretation task according to the scheduling plan.

[0133] Scheduling schemes are extracted from the globally optimal particles (an n*m two-dimensional decision matrix, where n represents the number of interpretation tasks to be executed and m represents the number of executors). Each row of the matrix represents the decision variables for each interpretation task. The maximum value in each row is selected as the scheduling result for each interpretation task. For example, if the maximum value in the j-th column of the i-th row is the maximum value in the i-th row, then the i-th interpretation task to be executed is assigned to executor j.

[0134] The interpretation task begins and ends based on the scheduling results.

[0135] The scheduling method for remote sensing data interpretation tasks obtains the estimated execution time of each task on the corresponding execution machine. The process is as follows:

[0136] Based on the task parameters of the task to be executed and the hardware parameters of the executor, a first linear equation including the task parameters and hardware parameters is constructed as a model for predicting the execution time of the task to be executed.

[0137] Based on the task parameters of historical tasks and the hardware parameters of the executor, a second linear equation including the task parameters and hardware parameters is constructed as a model for predicting the execution time of historical tasks.

[0138] The task parameters include the sample size corresponding to the task to be executed, the parameter size of the corresponding execution time prediction model, the spatial resolution of the sample image, and the spectral resolution of the sample image; the hardware parameters include the memory of the execution machine, the CPU clock frequency, the number of CPU cores, the GPU clock frequency, and the number of GPU cards used.

[0139] The loss function of the historical task is calculated using the actual execution time of the historical task and the execution time prediction model of the historical task.

[0140] Based on multiple loss functions, the partial derivatives of different coefficients of the loss functions are calculated to obtain all coefficients in the historical task execution time prediction model.

[0141] Substitute all coefficients into the execution time prediction model for the tasks to be executed, and then calculate the estimated execution time of each task on the corresponding execution machine.

[0142] In this embodiment, the linear regression method is used to construct the task execution time prediction model in step 2 above. The specific process is as follows:

[0143] 1) Model definition: Let the task execution time be T. The factors affecting the execution time are: task sample size x1, CPU clock frequency x2, number of CPU cores used x3, required memory x4, GPU clock frequency used x5, number of GPU cards used x6, algorithm model complexity x7, spatial resolution of sample image x8, and spectral resolution of sample image x9. The execution time prediction model can be expressed as T=w0+w1x1+w2x2+w3x3+w4x4+w5x5+w6x6+w7x7+w8x8+w9x9.

[0144] 2) Collect historical task execution data to create a historical task execution data set:

[0145] Collect the following features for each task: execution time T, factors affecting execution time, task sample size x1, CPU clock frequency x2, number of CPU cores used x3, required memory x4, GPU clock frequency used x5, number of GPU boards used x6, algorithm model complexity x7, spatial resolution of sample images x8, and spectral resolution of sample images x9.

[0146] 3) Set the linear fitting loss function for the duration prediction model:

[0147] Let the number of historical tasks be n, and the loss function be the sum of squared errors. Here, T(i) represents the actual execution time of the i-th task, and T^(i) represents the estimated execution time.

[0148]

[0149] 4) Parameter solution.

[0150] Based on the loss function J(w0,w1,w2,w3,w4,w5,w6,w7,w8,w9), we calculate the partial derivatives of the loss function with respect to w0,w1,w2,w3,w4,w5,w6,w7,w8,w9 respectively. Setting the partial derivatives to zero, we obtain a system of equations. Solving this system yields the values ​​of w0,w1,w2,w3,w4,w5,w6,w7,w8,w9.

[0151]

[0152] Based on the above set of equations, a model for predicting the execution time of the task to be executed with determined coefficients was further obtained.

[0153] Where n represents the number of historical task samples, y (i) Let x represent the actual execution time of the i-th historical task. (i) This represents the feature information of historical task i.

[0154] According to one aspect of the present invention, an electronic device is provided, comprising: one or more CPUs, one or more memories, and one or more computer programs; wherein the CPU is connected to the memory, and the one or more computer programs are stored in the memory; when the electronic device is running, the CPU executes the one or more computer programs stored in the memory to cause the electronic device to perform the remote sensing data interpretation task scheduling method as described above.

[0155] According to one aspect of the present invention, a computer-readable storage medium is provided for storing computer instructions, which, when executed by a CPU, implement a scheduling method for remote sensing data interpretation tasks as described above.

[0156] Computer-readable storage media can include any medium capable of storing or transmitting information. Examples of computer-readable storage media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, and so on. Code segments can be downloaded via computer networks such as the Internet and intranets.

[0157] The present invention provides a method for scheduling remote sensing data interpretation tasks, comprising: obtaining task data to be executed and user behavior data of the tasks to be executed; obtaining a first set of recommended interpretation algorithms and a second set of recommended interpretation algorithms based on pre-stored historical task features and historical task user behavior features, respectively; solving for the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms, and selecting an algorithm from either the first set of recommended interpretation algorithms or the second set of recommended interpretation algorithms as the recommended interpretation algorithm for the task to be executed; assuming that at least one executor is invoked to execute multiple tasks to be executed, and considering the recommended interpretation algorithm as a factor affecting the estimated execution time of the corresponding task to be executed, calculating the optimal solution for resource scheduling with the minimum estimated total execution time of the multiple tasks to be executed as the optimization objective; and completing the scheduling of the interpretation tasks based on the optimal solution for resource scheduling.

[0158] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0159] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the CPU of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the CPU of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0160] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0161] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises 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 terminal device. 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 terminal device that includes said element.

[0162] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A scheduling method for remote sensing data interpretation tasks, characterized in that, The steps are as follows: Obtain data on tasks to be executed and user behavior data for those tasks; The task data to be executed and the user behavior data of the task to be executed are respectively based on the pre-stored historical task features and historical task user behavior features to obtain the first interpretation algorithm recommendation set and the second interpretation algorithm recommendation set; Find the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms, and select an algorithm from the first set of recommended interpretation algorithms or the second set of recommended interpretation algorithms as the recommended interpretation algorithm for the task to be executed, based on whether the intersection is empty and the preset algorithm selection conditions. Assuming that at least one execution machine is invoked to execute multiple tasks to be executed, and the recommended interpretation algorithm is used as a factor affecting the estimated execution time of the corresponding tasks to be executed, the optimal solution for resource scheduling is calculated with the goal of minimizing the estimated total execution time of the multiple tasks to be executed; The number of executors is less than or equal to the number of tasks to be executed; Based on the optimal resource scheduling solution, multiple tasks to be executed are assigned to their respective execution machines to complete the scheduling of the interpretation tasks.

2. The scheduling method for remote sensing data interpretation tasks according to claim 1, characterized in that, The process of obtaining the characteristics of historical tasks is as follows: Obtain the task target scene label, task size, image parameters, algorithm operation indicators, user ID, interpretation algorithm used by the user, and the number of times each interpretation algorithm was used from historical tasks; Through feature engineering, the task target scene label, task size and image parameters are vectorized to obtain historical task features; Furthermore, by using task target scenario labels and user IDs as dimension labels, and statistically analyzing the interpretation algorithms used by users and the number of times each interpretation algorithm was used, a statistical feature vector is formed as the user behavior features of historical tasks.

3. The scheduling method for remote sensing data interpretation tasks according to claim 2, characterized in that, The process of obtaining the recommendation set of the first interpretation algorithm and the recommendation set of the second interpretation algorithm is as follows: Obtain the task target scene label, task size, image parameters, algorithm operation indicators, user ID, the interpretation algorithm used by the user, and the number of times each interpretation algorithm is used in the task to be executed; Through feature engineering, the task target scene label, task size and image parameters are vectorized to obtain the features of the task to be executed. In addition, the task target scene label and user ID are used as dimension labels, and the interpretation algorithms used by users and the number of times each interpretation algorithm is used are counted to form a statistical feature vector as the user behavior features of the task to be executed; Based on the user behavior characteristics of the task to be executed, the number of times each interpretation algorithm is used under the same user ID in the same task target scenario is found from multiple historical task user behavior characteristics; Select interpretation algorithms whose usage frequency is greater than or equal to a preset usage frequency threshold, and construct a first set of recommended interpretation algorithms; Calculate the second similarity between the features of the task to be executed and the features of multiple historical tasks to obtain multiple similar historical tasks; The similarity history task is a historical task whose second similarity with the features of the task to be executed is greater than a similarity threshold; We summarize all the interpretation algorithms corresponding to the historical similarity tasks and construct a second set of interpretation algorithm recommendations.

4. The scheduling method for remote sensing data interpretation tasks according to claim 3, characterized in that, Also includes: Based on the user behavior characteristics of the task to be executed, find the number of times the same user ID has executed the historical task in the same target scenario from multiple historical task user behavior characteristics; If the number of times the historical task is executed is less than the preset execution number threshold, then the first interpretation algorithm recommendation set is set to be empty.

5. The scheduling method for remote sensing data interpretation tasks according to any one of claims 2 to 4, characterized in that, In the first set of recommended interpretation algorithms, the interpretation algorithms are sorted by the number of times they are used. In the second set of recommended interpretation algorithms, the interpretation algorithms are ranked according to their performance metrics. The algorithm's performance metrics include accuracy and / or false alarm rate.

6. The scheduling method for remote sensing data interpretation tasks according to claim 5, characterized in that, The algorithm selection criteria are as follows: When the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms is empty, the interpretation algorithm ranked first in the second set of recommended interpretation algorithms is taken as the recommended interpretation algorithm for the corresponding task to be executed. When the intersection of the first set of recommended interpretation algorithms and the second set of recommended interpretation algorithms is not empty, the interpretation algorithm ranked first in the intersection is selected as the recommended interpretation algorithm for the corresponding task to be executed, based on the ranking of the recommended interpretation algorithms in the first set of recommended interpretation algorithms.

7. The scheduling method for remote sensing data interpretation tasks according to any one of claims 1 to 4 and 6, characterized in that, The process of calculating the optimal solution for resource scheduling is as follows: The total estimated execution time of all tasks to be executed is calculated by estimating the execution time of each task on the corresponding execution machine; and the resource scheduling objective function is constructed by minimizing the total estimated execution time. By using the particle swarm optimization method, and with the constraint that each task to be executed is assigned to only one executor, the optimal solution for resource scheduling is obtained by solving the resource scheduling objective function.

8. The scheduling method for remote sensing data interpretation tasks according to claim 7, characterized in that, The estimated execution time for each task to be executed on the corresponding execution machine is obtained as follows: Based on the task parameters of the task to be executed and the hardware parameters of the executor, a first linear equation including the task parameters and hardware parameters is constructed as a model for predicting the execution time of the task to be executed. Based on the task parameters of historical tasks and the hardware parameters of the executor, a second linear equation including the task parameters and hardware parameters is constructed as a model for predicting the execution time of historical tasks. The task parameters include the sample size corresponding to the task to be executed, the parameter size of the corresponding execution time prediction model, the spatial resolution of the sample image, and the spectral resolution of the sample image; the hardware parameters include the memory, CPU clock frequency, CPU core count, GPU clock frequency, and number of GPU cards used in the execution machine. The loss function of the historical task is calculated using the actual execution time of the historical task and the execution time prediction model of the historical task. Based on multiple loss functions, the partial derivatives of different coefficients of the loss functions are calculated to obtain all coefficients in the historical task execution time prediction model; Substitute all coefficients into the execution time prediction model of the task to be executed, and then calculate the estimated execution time of each task on the corresponding execution machine.

9. An electronic device, characterized in that, include: One or more CPUs, one or more memories, and one or more computer programs; wherein the CPU is connected to the memory, and the one or more computer programs are stored in the memory, and when the electronic device is running, the CPU executes the one or more computer programs stored in the memory to cause the electronic device to perform the scheduling method for remote sensing data interpretation tasks as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by the CPU, implement the scheduling method for remote sensing data interpretation tasks as described in any one of claims 1 to 8.

Images