A method, system, device and medium for retrieving land surface temperature on satellite

Abstract

The application discloses a method, system, device and medium for inverting ground surface temperature, which is applied to a satellite-ground cooperative system including a ground terminal and a satellite terminal, and comprises the following steps: the ground terminal acquires multi-source data about a ground surface temperature inversion task; the multi-source data at least includes remote sensing image data and meteorological data; an initial temperature inversion model which is constructed in advance is trained through the multi-source data; a target temperature inversion model obtained through training is subjected to light-weight processing, and the light-weight temperature inversion model is deployed to the satellite terminal after being obtained. The satellite terminal determines a target image area from real-time remote sensing images which are acquired in real time based on acquired task point information; the target image area is subjected to temperature inversion through the light-weight temperature inversion model, so that a ground surface temperature inversion result is obtained; and the ground surface temperature inversion result is transmitted to the ground terminal. Thus, the target image area can be directly subjected to temperature inversion at the satellite terminal, the load of a satellite-ground link is reduced, and the real-time performance of temperature monitoring is improved.

Description

Technical Field

[0001] This application relates to the field of remote sensing intelligent technology, and in particular to a method, system, device and medium for inverting on-board surface temperature. Background Technology

[0002] Land surface temperature is a crucial parameter in urban governance, agricultural production, disaster prevention and mitigation, ecological environment monitoring, and energy dispatch. Obtaining land surface temperature through satellite thermal infrared remote sensing imagery can provide data support for urban heat island monitoring, agricultural drought assessment, forest fire early warning, industrial heat source inspection, and emergency disaster relief.

[0003] Currently, surface temperature retrieval methods typically rely on a centralized ground-based processing model. This involves acquiring raw remote sensing images from satellites, transmitting the complete images to the ground, and then having the ground-based processing system perform data analysis, calibration, retrieval, and product generation.

[0004] However, this mode has the following shortcomings in actual satellite-ground collaborative missions. On the one hand, the original remote sensing images are large in size, with a single scene image typically reaching hundreds of megabytes (MB) or even higher, and a complete downlink would consume a large amount of satellite-ground link bandwidth. On the other hand, satellite transit windows are limited, and communication bandwidth and transmission time are restricted, making it difficult to meet the real-time requirements of disaster early warning, thermal anomaly monitoring, etc.

[0005] Therefore, how to save bandwidth resources while improving the real-time performance of surface temperature retrieval is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, one aspect of this application provides a method for retrieving on-board surface temperature, applied to a space-ground cooperative system including a ground end and a satellite end; the method includes: The ground-based terminal acquires multi-source data for the surface temperature retrieval task; the multi-source data includes at least remote sensing image data and meteorological data. The ground-based device trains the pre-constructed initial temperature inversion model using the multi-source data to obtain the target temperature inversion model. The ground terminal performs lightweight processing on the target temperature inversion model to obtain a lightweight temperature inversion model suitable for operation on the satellite terminal, and then deploys it to the satellite terminal. The satellite acquires real-time remote sensing images and mission location information related to the surface temperature inversion mission; and determines the target image area from the real-time remote sensing images based on the mission location information. The satellite terminal performs temperature inversion on the target image area using the lightweight temperature inversion model to obtain the surface temperature inversion result; and then transmits the surface temperature inversion result to the ground terminal.

[0007] Optionally, the remote sensing image data includes thermal infrared remote sensing images and historical remote sensing sequences corresponding to the thermal infrared remote sensing images; the meteorological data includes at least one of time data and weather data; the multi-source data also includes at least one of topographic elevation data, regional type data, and surface temperature annotation data.

[0008] Optionally, before training the pre-built initial temperature inversion model using the multi-source data at the ground end, the method further includes: The metadata and calibration data corresponding to the thermal infrared remote sensing image are analyzed, and the original digital quantization value of the thermal infrared remote sensing image is converted into a radiance value based on the calibration data. Based on the geographic coordinate information in the metadata, the latitude and longitude coordinates of the sample points are converted into pixel coordinates in the thermal infrared remote sensing image. Based on the pixel coordinates, local image samples are cropped from the remote sensing image data, and effective region masks for excluding invalid regions and effective label masks for limiting the training loss calculation region are generated for the local image samples.

[0009] Optionally, the ground-based device trains a pre-built initial temperature inversion model using the multi-source data, including: Obtain the task information for the surface temperature inversion task; The remote sensing image data and the task information are semantically enhanced using a remote sensing visual language model to obtain semantically enhanced information; the semantically enhanced information includes at least one of scene semantic information, terrain semantic information, region type semantic information, and anomalous region prior information. The semantic enhancement information is encoded to obtain remote sensing semantic conditions; Based on the multi-source data and the remote sensing semantic conditions, the initial temperature inversion model is iteratively trained.

[0010] Optionally, the initial temperature inversion model includes an encoder and a diffusion Transformer backbone network; the encoder includes an image encoder, a temporal encoder, a terrain encoder, and a conditional encoder. The image encoder is used to encode the thermal infrared remote sensing image; The time-series encoder is used to encode the historical remote sensing sequence; The terrain encoder is used to encode the terrain elevation data; The condition encoder is used to encode the meteorological data, the regional type data, and the remote sensing semantic conditions to obtain condition information; wherein, the condition information is injected into the diffusion Transformer backbone network through a classifier-free guidance mechanism. The diffusion Transformer backbone network employs a multi-path feature token fusion mechanism to fuse the features output by each encoder and output the land surface temperature prediction result.

[0011] Optionally, the lightweight temperature inversion model includes a first inference model and a second inference model; wherein, the first inference model includes at least one of a multilayer perceptron model, a linear regression model, and an XGBoost model; and the second inference model includes at least one of a U-Net model and a lightweight diffusion Transformer student model. The method further includes: If the satellite is in a low-power state or the computing load is greater than the preset load, the first inference model is invoked to perform temperature inversion on the target image area. If the available computing resources on the satellite meet the preset resource conditions, the second inference model is invoked to perform temperature inversion on the target image area.

[0012] Optionally, the surface temperature inversion results include at least one of the following: target point temperature value, local temperature prediction map, temperature data confidence level, and thermal anomaly label; the method further includes: The ground terminal acquires target temperature data including either measured temperature data or preset reference temperature data. The ground terminal compares the surface temperature inversion results with the target temperature data to obtain temperature error analysis results; Based on the error analysis results, the ground terminal updates the model training samples or adjusts the model parameters to obtain the updated lightweight temperature inversion model. The ground terminal deploys the updated lightweight temperature inversion model to the satellite terminal.

[0013] Another aspect of this application provides a space-ground collaborative system, the system comprising a ground end and a satellite end; the ground end includes a multi-source data acquisition module, a model training module, and a lightweight module; the satellite end includes a target area determination module and a temperature inversion module; The multi-source data acquisition module is used to acquire multi-source data for the land surface temperature retrieval task; the multi-source data includes at least remote sensing image data and meteorological data. The model training module is used to train the pre-constructed initial temperature inversion model using the multi-source data to obtain the target temperature inversion model. The lightweight module is used to perform lightweight processing on the target temperature inversion model to obtain a lightweight temperature inversion model suitable for operation on the satellite, and then deploy it to the satellite. The target area determination module is used to acquire real-time remote sensing images and task location information related to the land surface temperature inversion task; and to determine the target image area from the real-time remote sensing images based on the task location information. The temperature inversion module is used to perform temperature inversion on the target image area using the lightweight temperature inversion model to obtain the surface temperature inversion result; and to transmit the surface temperature inversion result to the ground terminal.

[0014] Another aspect of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the steps of the method for retrieving the surface temperature of the satellite.

[0015] Another aspect of this application provides a computer-readable storage medium having a computer program stored thereon, the steps of the method for retrieving the surface temperature of a satellite being executed by a processor.

[0016] The method, system, equipment, and medium for inverting surface temperature on a satellite provided in this application have the following beneficial effects: By training and lightweighting the temperature inversion model using multi-source data at the ground end, and then deploying the lightweight temperature inversion model to the satellite, the satellite can directly complete the surface temperature inversion by selecting target image areas based on mission location information, without needing to download complete original remote sensing images to the ground. This reduces the amount of irrelevant image data processed and the amount of data downloaded, lowers the bandwidth usage of satellite-to-ground communication and the pressure on ground processing, shortens the latency of acquiring surface temperature inversion results, and improves the real-time response capability for tasks such as disaster early warning and thermal anomaly monitoring. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart illustrating a method for retrieving surface temperature on a satellite, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a satellite-ground cooperative system provided in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram illustrating the principle of an on-board surface temperature inversion method provided in this application embodiment.

[0020] Figure 4 This is a schematic diagram illustrating the principle of a temperature inversion model provided in an embodiment of this application.

[0021] Figure 5 This is a schematic diagram of the structure of a satellite-ground cooperative system provided in an embodiment of this application.

[0022] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0023] The attached diagram is labeled as follows: 1 represents the satellite-ground collaborative system, 10 represents the ground terminal, 11 represents the satellite terminal, 100 represents the multi-source data acquisition module, 101 represents the model training module, 102 represents the lightweight module, 110 represents the target area determination module, 111 represents the temperature inversion module, 60 represents the memory, 61 represents the processor, 62 represents the display screen, 63 represents the input / output interface, 64 represents the communication interface, 65 represents the power supply, 66 represents the communication bus, 601 represents the computer program, 602 represents the operating system, and 603 represents the data. Detailed Implementation

[0024] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0025] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0026] Figure 1 This is a schematic flowchart illustrating a method for retrieving surface temperature on a satellite, as provided in an embodiment of this application. Figure 2 This is a schematic diagram of a satellite-ground cooperative system provided in an embodiment of this application. The method provided in this application is applied to a satellite-ground cooperative system 1 including a ground terminal 10 and a satellite terminal 11. In a specific embodiment, the satellite terminal 11 and the ground terminal 10 establish a communication connection through a satellite-ground communication link.

[0027] In this embodiment, the ground terminal 10 is responsible for tasks such as multi-source data access, task planning, construction of model training samples, initial temperature inversion model training, lightweight model processing, and generation of model deployment packages. The satellite terminal 11 is responsible for real-time remote sensing image access, task location analysis, target image area determination, lightweight temperature inversion model operation, and surface temperature inversion result encapsulation and downlink.

[0028] The ground terminal 10 can be implemented by a ground server, ground workstation, cloud platform, local client or mission scheduling platform, and the satellite terminal 11 can be implemented by an onboard processor, onboard memory or onboard edge computing platform. This application does not limit the implementation of these features.

[0029] like Figure 1 As shown, the methods for retrieving surface temperature on a satellite include: S10: Acquire multi-source data for the land surface temperature retrieval task at the ground end; the multi-source data should include at least remote sensing image data and meteorological data; In a specific embodiment, Land Surface Temperature (LST) refers to the surface temperature at the interface between the Earth's surface and the atmosphere. After initiating an LST inversion task based on actual operational needs, the ground terminal 10 initializes the data interfaces, data points, model version, and runtime configuration required for the temperature inversion task. These operational needs may include, but are not limited to, rapid inspection of key urban heat island areas, drought assessment of agricultural irrigation areas, and screening for suspected thermal anomalies in forest fires or industrial heat sources. After initiating the LST inversion task, the ground terminal 10 acquires multi-source data related to the land surface temperature inversion task.

[0030] Multi-source data includes at least remote sensing imagery and meteorological data. Remote sensing imagery can provide the basis for observing surface thermal radiation, while meteorological data is used to supplement information on external environmental conditions such as weather and time, thus ensuring that subsequent model training does not rely solely on a single remote sensing image.

[0031] Specifically, after establishing the remote sensing image data and meteorological data corresponding to the LST inversion task on the ground end 10, data with the same spatial or temporal range are associated, and the task number, data time, coverage area and data version are recorded for subsequent model training.

[0032] S11: The ground end trains the pre-built initial temperature inversion model using multi-source data to obtain the target temperature inversion model; In a specific embodiment, the initial temperature inversion model is used to learn the mapping relationship between remote sensing image data, meteorological data, and other accessible conditions and land surface temperature. After reading multi-source data, the ground terminal 10 uses the land surface radiation state represented by the remote sensing image and the environmental conditions represented by the meteorological data as inputs to the initial temperature inversion model, and iteratively updates and trains the model parameters of the initial temperature inversion model. When the training reaches the preset iteration condition, the iteration stops and the target temperature inversion model is output. The preset iteration condition can be any one of the preset training rounds, convergence conditions, and accuracy requirements, which are not limited in this application. If the preset iteration condition is not reached, model training continues.

[0033] S12: The ground end performs lightweight processing on the target temperature inversion model to obtain a lightweight temperature inversion model suitable for satellite operation, and then deploys it to the satellite. Figure 3 A schematic diagram illustrating the principle of an on-board surface temperature inversion method provided in this application embodiment; further, as shown... Figure 3 As shown, the ground terminal 10 performs lightweight processing on the target temperature inversion model according to the model operation interface of the satellite terminal 11. The lightweight model weights, model structure configuration, and normalized parameters are combined into a model deployment package, which is then sent to the satellite terminal 11 via the satellite-to-ground communication link. After verifying the matching relationship between the model deployment package and the operating environment, the satellite terminal 11 loads the lightweight temperature inversion model into the onboard Artificial Intelligence (AI) acceleration unit.

[0034] It should be noted that lightweight processing can include at least one of knowledge distillation, model pruning, parameter quantization, and single-step inference optimization. Specifically, knowledge distillation uses a larger model trained on the ground-side platform as the teacher model to guide the lightweight student model in learning temperature inversion capabilities. Model pruning can remove low-contribution channels, attention heads, or redundant modules. Parameter quantization can convert model parameters from 32-bit floating-point numbers (FP32) to 16-bit half-precision floating-point numbers (FP16) or 8-bit integers (INT8), thereby reducing model storage and inference computation. Single-step inference optimization can compress the multi-step temperature generation process into a single or fewer-step inference step. In one optional embodiment, the compressed lightweight model weight file, model structure configuration, and normalized parameters together constitute the lightweight model deployment package.

[0035] In this application, the ground terminal 10 does not upload complete remote sensing images to the satellite terminal 11. Instead, it uploads only the mission locations, runtime configuration, model deployment packages, or model incremental update files via the satellite-ground link. The mission locations include at least the mission number, image identifier, latitude and longitude coordinates, batch size, inference mode, and downlink strategy.

[0036] S13: The satellite acquires real-time remote sensing images and mission location information for the land surface temperature inversion mission; and determines the target image area from the real-time remote sensing images based on the mission location information. After deploying a lightweight temperature inversion model, in a specific embodiment, satellite terminal 11 receives real-time remote sensing image streams from the thermal infrared remote sensing payload and receives mission location information from the satellite-to-ground communication link. Further, it determines the real-time remote sensing image to be processed based on image identifiers in the mission location, and then locates the target image region in the real-time remote sensing image based on the spatial location of the mission location. Therefore, satellite terminal 11 can perform temperature inversion without first sending the complete original image to ground terminal 10.

[0037] In one optional embodiment, when a single task includes one task location, the target image region is the local region corresponding to that task location. When a task includes multiple task locations, the satellite terminal 11 determines the local region corresponding to each task location and forms a processing queue according to a tile or scene caching method. It should be noted that the processing queue is used to constrain the processing order of each target image region and does not change the step of determining the target image region based on the task location information.

[0038] S14: The satellite uses a lightweight temperature inversion model to perform temperature inversion on the target image area to obtain the surface temperature inversion results; and then transmits the surface temperature inversion results to the ground.

[0039] Furthermore, satellite 11 reads the target image area and inputs it into a lightweight temperature inversion model to perform forward inference, obtaining the surface temperature inversion results corresponding to the mission location. Then, the surface temperature inversion results are associated with the mission location and image identifier and sent to ground 10, so that ground 10 can perform spatial mapping, result display, or business analysis based on the surface temperature inversion results.

[0040] It should be noted that the on-orbit processing of the satellite-end 11 includes, but is not limited to, real-time remote sensing image reading, conversion of digital quantization values ​​(DN) to radiance, masking of invalid areas, target point coordinate transformation, patch clipping, and patch batch processing organization.

[0041] It is understandable that the target temperature inversion model is the model that has been trained and is used for subsequent lightweight processing, which is different from the lightweight temperature inversion model that will eventually run on satellite-11. The target temperature inversion model retains the multi-source data representation capabilities formed during the ground training phase, while the lightweight temperature inversion model adapts to the computing power, power consumption, storage, and mission timeliness conditions of satellite-11 while maintaining the temperature inversion function.

[0042] In a specific embodiment, after the satellite-end 11 completes inference, it encapsulates the temperature inversion results, image identifiers, mission locations, pixel coordinates, model version, inference time, and quality identifier into a lightweight result.json file (which can be as small as 4KB) and downloads it to the ground-end 10, thereby reducing the bandwidth pressure on satellite-to-ground communication. The ground-end 10 receives and parses the result.json file, directly mapping it to the corresponding space unit for practical applications such as urban heat island monitoring, agricultural drought assessment, forest fire early warning, or industrial heat source inspection.

[0043] Therefore, the on-board surface temperature retrieval method provided in this application uses multi-source data to train and lightweight the temperature retrieval model at the ground terminal 10, and then deploys the lightweight temperature retrieval model to the satellite terminal 11. This allows the satellite terminal 11 to filter target image areas based on mission location information and directly complete the surface temperature retrieval without needing to download the complete original remote sensing image to the ground. This reduces the amount of irrelevant image data processed and the amount of data downloaded, lowers the bandwidth usage of satellite-to-ground communication and the pressure on ground processing, shortens the latency of obtaining surface temperature retrieval results, and improves the real-time response capability for tasks such as disaster early warning and thermal anomaly monitoring.

[0044] It is understood that remote sensing imagery, meteorological data, and various types of data related to the spatial state of the Earth's surface can all affect LST inversion. Therefore, in one optional embodiment, the remote sensing imagery includes thermal infrared remote sensing images and corresponding historical remote sensing sequences; the meteorological data includes at least one of temporal data and weather data; and the multi-source data also includes at least one of topographic elevation data, regional type data, and surface temperature labeling data.

[0045] In specific embodiments, thermal infrared remote sensing images can be used to reflect the current surface radiation state, and historical remote sensing sequences can be used to reflect temperature changes in the same area at different times. In an optional embodiment, the accompanying files for thermal infrared remote sensing images include metadata XML and calibration XML files. Time data and weather data can be used to reflect the imaging time and environmental conditions. Time data may include, but is not limited to, imaging date, imaging time, season, and day / night indicator. Weather data may include, but is not limited to, temperature, humidity, wind speed, cloud cover, or precipitation status.

[0046] Topographic elevation data can be used to reflect differences in terrain. Regional type data can be used to distinguish surface categories such as built-up areas, road areas, forests, and irrigation areas. Surface temperature labeling data can be used to provide the target temperature required for model training.

[0047] Based on the above embodiments, to ensure that thermal infrared remote sensing images can form training samples with spatial location and effective annotation range, as an optional embodiment, such as... Figure 3 As shown, the acquired multi-source data is preprocessed before training the initial temperature inversion model on the ground. This preprocessing includes, but is not limited to, radiometric calibration of remote sensing images, metadata parsing, latitude and longitude to pixel coordinate transformation, generation of invalid region masks, and construction of local patch samples.

[0048] Specifically, before training the pre-built initial temperature inversion model using multi-source data at the ground end 10, the on-board surface temperature inversion method also includes: The metadata and calibration data corresponding to the thermal infrared remote sensing image are analyzed, and the original digital quantization values ​​of the thermal infrared remote sensing image are converted into radiance values ​​based on the calibration data. Based on the geographic coordinate information in the metadata, the latitude and longitude coordinates of the sample points are converted into pixel coordinates in the thermal infrared remote sensing image. Based on pixel coordinates, local image samples are cropped from remote sensing image data, and effective region masks are generated for the local image samples to exclude invalid regions and effective label masks to limit the training loss calculation area.

[0049] In a specific embodiment, the ground-based unit 10 analyzes the thermal infrared remote sensing images to obtain corresponding metadata and calibration data. Based on the calibration data, the raw DN of the thermal infrared remote sensing images is converted into radiance values. Radiance values ​​refer to a physical quantity characterizing the intensity of radiant energy received by the thermal infrared channel, used to ensure that the raw DN of different image batches is entered into a unified training scale. Metadata may include, but is not limited to, information such as image identifiers, imaging time, four-corner latitude and longitude coordinates, and projected coordinates. Calibration data may include, but is not limited to, the gain parameters and bias parameters of the thermal infrared channel.

[0050] Furthermore, the ground terminal 10 converts the latitude and longitude coordinates of the sample points into pixel coordinates in the thermal infrared remote sensing image based on the geographic coordinate information in the metadata. In a specific conversion embodiment, the latitude and longitude information or projected coordinate information of the four corners of the image is read, and the row and column numbers corresponding to the sample points are determined according to the image coordinate reference and pixel resolution.

[0051] Then, local image samples are cropped from the remote sensing image data based on pixel coordinates, and effective region masks are generated for black borders, invalid fill values, cloud occlusion, or other invalid areas. Effective label masks are generated for locations with surface temperature annotations, ensuring that the training loss is calculated only at effective labeled locations or within effective regions.

[0052] For training samples, a pre-defined fixed-size local patch is cropped centered on the labeled point or task point, and the patch's center position, internal labeled point positions, and effective label mask are recorded. The label mask is used to limit the training loss calculation to only the effective labeled position or effective region, avoiding interference from unlabeled areas with model learning.

[0053] For example, in one optional embodiment, a 128×128 pixel local image sample is cropped centered on the pixel coordinates of the sample points, and the center position of the patch, the position of the internal annotation points, and the two types of masks are recorded. Under FP32 storage conditions, the data size of a single 128×128 pixel patch is approximately 192KB.

[0054] In another optional embodiment, when the same local image sample contains multiple valid annotation points, the valid label mask sets the corresponding positions of each annotation point as valid. When only the center point is labeled, only the center point or the valid region surrounding the center point participates in the training loss calculation.

[0055] It should be noted that preprocessing of the thermal infrared remote sensing image needs to be performed by both the ground terminal 10 and the satellite terminal 11. In one optional embodiment, after the ground terminal 10 or the satellite terminal 11 reads the TIFF image, the metadata XML file, and the calibration XML file, it parses the gain and bias parameters in the calibration file and converts the original DN values ​​into radiance values ​​according to the channels. Subsequently, it parses the four-corner latitude and longitude information or projected coordinate information in the metadata and calculates the row and column pixel coordinates of the target point in the image based on the latitude and longitude of the task point. For black borders, invalid fill values, cloud occlusion, or other invalid areas in the image, the system generates an effective mask and excludes these areas during training and inference.

[0056] For onboard temperature inference on satellite 11, in one optional embodiment, satellite 11 receives the thermal infrared remote sensing payload image stream (the data size of a single scene is typically about 0.8GB to 0.9GB). Satellite 11 reads the TIFF images in real-time on orbit and parses the XML files, completing DN to radiance conversion, target point pixel coordinate calculation, invalid area mask generation, and 128×128 pixel patch cropping, avoiding the need to download the complete original image. Satellite 11 organizes the cropped patches into a batch processing queue using tiles or scene caching. During the model inference phase, the onboard AI acceleration unit is responsible for the parallel forward inference of the lightweight temperature inference model.

[0057] It is understandable that the same radiance may correspond to different surface conditions in different scenes, terrains, and region types. To provide constraints corresponding to the task scenario during model training, as an optional implementation, such as... Figure 3 As shown, semantic augmentation is performed on the pre-trained data.

[0058] Specifically, as an optional embodiment, the ground terminal 10 trains a pre-built initial temperature inversion model using multi-source data, including: Obtain mission information for the land surface temperature inversion task; Using a remote sensing visual language model, semantic enhancement is performed on remote sensing image data and task information to obtain semantic enhancement information. The semantic enhancement information includes at least one of scene semantic information, landform semantic information, region type semantic information, and anomalous region prior information. Encode the semantic enhancement information to obtain remote sensing semantic conditions; Based on multi-source data and remote sensing semantic conditions, the initial temperature inversion model is iteratively trained.

[0059] In specific embodiments, task information may include, but is not limited to, task number, task location, target area, service type, inference mode, download strategy, and model version. The remote sensing visual language model is used to perform scene understanding on remote sensing image data and task information, outputting semantic descriptions of densely built-up urban areas, roads, water bodies, green spaces, farmland, irrigation areas, forests, industrial heat sources, or suspected fire points. In an optional embodiment, the remote sensing visual language model may be a Falcon remote sensing large model; this application does not limit the type of remote sensing visual language model.

[0060] Semantic enhancement information is obtained by semantically enhancing remote sensing image data and task information through a remote sensing visual language model. Among them, scene semantic information is used to describe the overall scene such as city, farmland or forest, landform semantic information is used to describe the surface morphology, area type semantic information is used to describe local categories such as building area, road area, water body, green space or irrigation area, and anomalous area prior information is used to describe suspected fire or industrial heat source areas.

[0061] Furthermore, the semantic enhancement information is encoded to obtain remote sensing semantic conditions, i.e., conditional tokens are formed. The multi-source data and remote sensing semantic conditions are then used as the output of the initial temperature inversion model, which is then iteratively trained. Figure 3 As shown, after each iteration of training, it is determined whether the preset iteration conditions have been met. If so, the target temperature inversion model is output; otherwise, the iteration loop continues.

[0062] Therefore, the remote sensing visual language model transforms the meaning of tasks and scenes into trainable remote sensing semantic conditions, enabling the initial temperature inversion model to obtain scene, topography and anomalous area constraints when fusing imagery and meteorological information, thereby improving the temperature inversion accuracy of the target temperature inversion model.

[0063] Figure 4 This is a schematic diagram of the principle of a temperature inversion model provided in an embodiment of this application. In an optional embodiment, the initial temperature inversion model includes an encoder and a diffusion Transformer backbone network; the encoder includes an image encoder, a temporal encoder, a terrain encoder, and a conditional encoder. Image encoders are used to encode thermal infrared remote sensing images; A time-series encoder is used to encode historical remote sensing sequences; A terrain encoder is used to encode terrain elevation data; The condition encoder is used to encode meteorological data, regional type data, and remote sensing semantic conditions to obtain condition information; the condition information is injected into the diffusion Transformer backbone network through a classifier-free guidance mechanism. The Diffusion Transformer backbone network employs a multi-path feature token fusion mechanism to fuse the features output by each encoder and output the land surface temperature prediction result.

[0064] In a specific embodiment, such as Figure 4 As shown, the image encoder encodes thermal infrared remote sensing images to extract spatial texture and thermal radiation features of the target image area. The temporal encoder encodes historical remote sensing sequences to extract temperature change trends of the same surface area over historical time periods. The topographic encoder encodes topographic elevation data to supplement elevation and landform relief information. The conditional encoder encodes meteorological data, regional type data, and remote sensing semantic conditions to obtain conditional information. This conditional information is injected into the diffusion Transformer backbone network through a classifier-free guidance mechanism (CFG) to constrain the temperature generation or temperature inversion process.

[0065] The Diffusion Transformer backbone network employs a multi-path feature token fusion mechanism to fuse features from each encoder and output land surface temperature prediction results. This mechanism allows image tokens, temporal tokens, terrain tokens, and conditional tokens to be input into the backbone network in a preset order, enabling the model to simultaneously reference current imagery, historical states, terrain elevation, and task semantics during the same inference process.

[0066] In this embodiment, an image encoder, a temporal encoder, a terrain encoder, and a condition encoder are used to encode spatial observation, historical changes, terrain environment, and conditional constraints, respectively. The diffused Transformer backbone network then performs multi-path feature token fusion, so that temperature prediction no longer depends on a single thermal infrared image, thereby improving the stability of temperature inversion in complex scenarios.

[0067] Based on the above embodiments, as an optional embodiment, the lightweight temperature inversion model includes a first inference model and a second inference model; wherein, the first inference model includes at least one of a multilayer perceptron (MLP) model, a linear regression model, and an XGBoost model; and the second inference model includes at least one of a U-Net model and a lightweight diffusion Transformer student model. The method for retrieving surface temperature on a satellite also includes: If the satellite is in a low-power state or the computing load is greater than the preset load, the first inference model is called to perform temperature inversion on the target image area. If the available computing resources on the satellite meet the preset resource conditions, the second inference model is invoked to perform temperature inversion on the target image area.

[0068] It is understandable that the power consumption and computational load of satellite terminal 11 are different in different mission phases. In order to keep the mission executable under low resource conditions and output continuous space temperature results when resources are sufficient, the corresponding model can be selected for temperature inversion based on the actual power consumption and computational load of satellite terminal 11.

[0069] In a specific embodiment, the first inference model is used to quickly estimate the temperature of the target point under low power consumption or high computational load conditions. Specifically, if the satellite terminal 11 is in a low power consumption state or the computational load is greater than a preset load, the satellite terminal 11 calls the first inference model to perform temperature inversion on the target image area. Among them, the MLP model can output temperature values ​​based on the center point and neighborhood features, the linear regression model can be used for rapid engineering estimation, and the XGBoost model can use multiple structured features to predict the center point temperature.

[0070] In another optional embodiment, if the available computing resources of satellite 11 meet the preset resource conditions, the second inference model can be invoked to perform temperature inversion on the target image area. The second inference model is used to output a local temperature prediction map or thermal anomaly area result with stronger spatial continuity. The U-Net model is suitable for encoder-decoder spatial temperature map inversion, and the lightweight diffusion Transformer student model can be obtained from the target temperature inversion model of ground 10 through knowledge distillation and compression.

[0071] It should be noted that the initial temperature inversion model trained on the ground-based 10, the target temperature inversion model obtained after training, and the lightweight temperature inversion model after lightweight processing can all be either the first inference model or the second inference model. It is important to note that the three models should be of the same type throughout the entire model training and lightweight processing chain.

[0072] Therefore, the satellite terminal 11 can dynamically select the temperature inversion model based on factors such as power consumption, computing power, storage, and task timeliness, avoiding forcing the running of complex models when resources are scarce, and also avoiding outputting only low-information results when resources are abundant.

[0073] Based on the above embodiments, as an optional embodiment, the surface temperature inversion result includes at least one of the following: target point temperature value, local temperature prediction map, temperature data confidence level, and thermal anomaly label. The on-board surface temperature inversion method further includes: The ground-based system acquires target temperature data, including either measured temperature data or preset reference temperature data. The ground end compares the surface temperature inversion results with the target temperature data to obtain temperature error analysis results; Based on the error analysis results, the ground-based system updates the model training samples or adjusts the model parameters to obtain the updated lightweight temperature inversion model. The updated lightweight temperature inversion model will be deployed from the ground to the satellite.

[0074] In a specific embodiment, after the satellite terminal 11 completes the inference, it can encapsulate the surface temperature inversion results along with the task location, image identifier, pixel coordinates, model version, inference time, and quality identifier into a lightweight result file and download it to the ground terminal 10.

[0075] The ground terminal 10 acquires target temperature data, which includes either measured temperature data or preset reference temperature data. In a specific embodiment, the target temperature data may come from a ground temperature station, ground sample collection, or historical temperature data, and this application does not limit this.

[0076] Furthermore, after obtaining the target temperature data, the surface temperature inversion results are compared with the target temperature data to obtain temperature error analysis results. In specific implementation, error analysis can be performed by calculating the difference, comparing the mean of the inverted temperature with the target temperature data, or calculating the difference and then averaging the results. This application does not limit the scope of error analysis.

[0077] Furthermore, based on the error analysis results, the model training samples are updated or the model parameters are adjusted to obtain an updated lightweight temperature inversion model, and the updated lightweight temperature inversion model is redeployed to the satellite.

[0078] In one optional embodiment, for areas with large errors, the ground terminal 10 can add the corresponding image patch, multi-source condition information and reference temperature data to the new training sample library, retrain or fine-tune the target temperature inversion model, and then generate a new deployment package after lightweight processing and uplink it to the satellite terminal 11.

[0079] Thus, the process forms a closed-loop mechanism of "mission point uplink—on-board image access—on-board preprocessing—Patch caching—batch inference—lightweight downlink of result.json—ground error analysis—model update—compressed uplink—on-board redeployment".

[0080] Under this mechanism, large-volume images are processed on the satellite, while the ground end 10 mainly uploads task locations, operational configurations, or model increments, and the satellite end 11 only downloads high-value temperature results. This enables efficient, stable, and sustainable optimization and inversion of surface temperature under conditions of limited computing power, limited power consumption, limited communication windows, and significant migration across regions, seasons, and geomorphological environments.

[0081] To help those skilled in the art better understand the technical solution of this application, a specific embodiment will be described below.

[0082] Currently, in a feasible temperature inversion implementation method, ground-based task scheduling can be initiated through a designated task scheduling platform. The size of a single thermal infrared TIFF image is approximately 850MB to 906MB, the lightweight model file `best_model` is approximately 30MB, and the data size of a single 128×128 pixel patch is approximately 192KB. The satellite-to-ground result file `result.json` is approximately 4KB. The CPU handles reading, XML parsing, calibration, and cropping, while a 5090D GPU or equivalent onboard AI accelerator handles batch inference. Under normal conditions, latency is in the range of seconds to tens of seconds, and in batch patching or multi-scene concurrent scenarios, it can be in the range of tens of seconds. Current validation results show that this process can complete end-to-end operational validation from remote sensing scene understanding to intelligent inversion of surface parameters, and can continuously improve cross-environment generalization ability and inversion accuracy based on subsequent new data and feedback samples.

[0083] In one optional embodiment, when implementing temperature inversion using the on-board surface temperature inversion method provided in this application, the training dataset constructed on the ground terminal 10 has a fixed-size local patch size set to 128×128 pixels. Under three-channel FP32 storage conditions, the data size of a single patch is approximately 192KB. For a single scene of thermal infrared TIFF imagery, the original data size is approximately 850MB to 906MB in the specific validation data. After knowledge distillation, model pruning, parameter quantization, and one-step optimization on the ground terminal 10, the compressed lightweight model weight file is named best_model and is approximately 30MB in size. The single-task result.json file, which is packaged and downloaded after inference on the satellite terminal 11, is approximately 4KB in size.

[0084] After training and on-orbit deployment (average of 10 tests) using the inversion method provided in this application, the distributed on-orbit inference strategy at the uplink mission points on the satellite-to-ground link reduces the amount of satellite-to-ground communication data to about 0.0005% compared to the traditional scheme of directly downloading about 850MB to 906MB of original panoramic images, thus alleviating the bandwidth pressure on the satellite-to-ground link.

[0085] In terms of accuracy and performance, the lightweight temperature inversion model achieves an average macro accuracy of 94.18% for overall surface temperature inversion in complex multi-source scenarios. Under normal single-scene tasks, the end-to-end response time (including on-orbit image parsing, calibration, patch cropping to model inference) is completed within seconds to ten seconds. For suspected thermal anomaly areas in sudden forest fires or industrial heat source scenarios, the on-orbit intelligent screening macro recognition rate reaches 91.35%, thereby improving the real-time performance of satellite remote sensing applications and the operational performance of the space-ground collaborative network.

[0086] In the above embodiments, the method for inverting the surface temperature on the satellite has been described in detail. This application also provides an embodiment corresponding to a satellite-ground cooperative system.

[0087] Figure 5 This is a schematic diagram of the structure of a satellite-ground cooperative system provided in an embodiment of this application, as shown below. Figure 5 As shown, the system includes a ground terminal 10 and a satellite terminal 11; the ground terminal 10 includes a multi-source data acquisition module 100, a model training module 101, and a lightweight module 102; the satellite terminal 11 includes a target area determination module 110 and a temperature inversion module 111, wherein: The multi-source data acquisition module 100 is used to acquire multi-source data for the land surface temperature inversion task; the multi-source data includes at least remote sensing image data and meteorological data. The model training module 101 is used to train the pre-built initial temperature inversion model using multi-source data to obtain the target temperature inversion model. The lightweight module 102 is used to perform lightweight processing on the target temperature inversion model to obtain a lightweight temperature inversion model suitable for operation on the satellite terminal 11, and to deploy it to the satellite terminal 11. The target area determination module 110 is used to acquire real-time remote sensing images and task location information related to the land surface temperature inversion task; and to determine the target image area from the real-time remote sensing images based on the task location information. The temperature inversion module 111 is used to perform temperature inversion on the target image area using a lightweight temperature inversion model to obtain the surface temperature inversion result; and to transmit the surface temperature inversion result to the ground terminal 10.

[0088] The remote sensing image data includes thermal infrared remote sensing images and corresponding historical remote sensing sequences; the meteorological data includes at least one of time data and weather data; the multi-source data also includes at least one of topographic elevation data, regional type data and surface temperature annotation data.

[0089] Furthermore, the ground terminal 10 in the satellite-ground collaborative system provided in this application embodiment also includes a data parsing module, a coordinate transformation module, and a sample clipping module. Wherein: The data parsing module is used to parse the metadata and calibration data corresponding to the thermal infrared remote sensing image, and convert the raw digital quantization value of the thermal infrared remote sensing image into a radiance value based on the calibration data. The coordinate transformation module is used to convert the latitude and longitude coordinates of sample points into pixel coordinates in thermal infrared remote sensing images based on the geographic coordinate information in the metadata. The sample cropping module is used to crop local image samples from remote sensing image data based on pixel coordinates, and generate effective region masks for excluding invalid regions and effective label masks for limiting the training loss calculation area for the local image samples.

[0090] The model training module is specifically used for: Obtain mission information for the land surface temperature inversion task; Using a remote sensing visual language model, semantic enhancement is performed on remote sensing image data and task information to obtain semantic enhancement information. The semantic enhancement information includes at least one of scene semantic information, landform semantic information, region type semantic information, and anomalous region prior information. Encode the semantic enhancement information to obtain remote sensing semantic conditions; Based on multi-source data and remote sensing semantic conditions, the initial temperature inversion model is iteratively trained.

[0091] The initial temperature inversion model includes an encoder and a diffusion Transformer backbone network; the encoder includes an image encoder, a temporal encoder, a terrain encoder, and a conditional encoder. Image encoders are used to encode thermal infrared remote sensing images; A time-series encoder is used to encode historical remote sensing sequences; A terrain encoder is used to encode terrain elevation data; The condition encoder is used to encode meteorological data, regional type data, and remote sensing semantic conditions to obtain condition information; the condition information is injected into the diffusion Transformer backbone network through a classifier-free guidance mechanism. The Diffusion Transformer backbone network employs a multi-path feature token fusion mechanism to fuse the features output by each encoder and output the land surface temperature prediction result.

[0092] The lightweight temperature inversion model includes a first inference model and a second inference model; wherein, the first inference model includes at least one of a multilayer perceptron model, a linear regression model, and an XGBoost model; and the second inference model includes at least one of a U-Net model and a lightweight diffusion Transformer student model. The satellite terminal 11 in the satellite-ground collaborative system 1 provided in this application embodiment further includes a first invocation module and a second invocation module. Wherein: The first calling module is used to call the first inference model to perform temperature inversion on the target image area if the satellite terminal 11 is in a low power consumption state or the computing load is greater than the preset load. The second invocation module is used to invoke the second inference model to perform temperature inversion on the target image area if the available computing resources of the satellite terminal 11 meet the preset resource conditions.

[0093] The ground terminal 10 of the satellite-ground collaborative system 1 provided in this application embodiment further includes a target temperature data acquisition module, a comparison module, an error analysis module, and an update module. Wherein: The target temperature data acquisition module is used to acquire target temperature data, including either measured temperature data or preset reference temperature data; wherein, the surface temperature inversion result includes at least one of the following: target point temperature value, local temperature prediction map, temperature data confidence level, and thermal anomaly marker. The comparison module is used to compare the surface temperature inversion results with the target temperature data to obtain temperature error analysis results. The error analysis module is used by the ground terminal 10 to update the model training samples or adjust the model parameters based on the error analysis results, so as to obtain the updated lightweight temperature inversion model. The update module is used by the ground terminal 10 to deploy the updated lightweight temperature inversion model to the satellite terminal 11.

[0094] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 6 As shown, the electronic device includes: a memory 60 for storing computer programs; The processor 61 is used to execute a computer program to implement the steps of the method for inverting the surface temperature on the satellite as described in the above embodiments.

[0095] The electronic devices provided in this embodiment may include, but are not limited to, laptops or desktop computers.

[0096] The processor 61 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 61 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 61 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 61 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 61 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0097] The memory 60 may include one or more computer-readable storage media, which may be non-transitory. The memory 60 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 60 is used to store at least the following computer program 601, which, after being loaded and executed by the processor 61, is capable of implementing the relevant steps of the on-board surface temperature inversion method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 60 may also include an operating system 602 and data 603, etc., and the storage method may be temporary storage or permanent storage. The operating system 602 may include Windows, Unix, Linux, etc. The data 603 may include, but is not limited to, the relevant data involved in the on-board surface temperature inversion method.

[0098] In some embodiments, the electronic device may further include a display screen 62, an input / output interface 63, a communication interface 64, a power supply 65, and a communication bus 66.

[0099] Those skilled in the art will understand that Figure 6 The structures shown do not constitute a limitation on electronic devices and may include more or fewer components than those shown.

[0100] The electronic device provided in this application includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the method for inverting the surface temperature on the satellite in the above embodiments.

[0101] It should be noted that although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Claims

1. A method for inverting surface temperature on a satellite, characterized in that, Applied to a space-ground collaborative system including ground and satellite ends; the method includes: The ground-based terminal acquires multi-source data for the surface temperature retrieval task; the multi-source data includes at least remote sensing image data and meteorological data. The ground-based device trains the pre-constructed initial temperature inversion model using the multi-source data to obtain the target temperature inversion model. The ground terminal performs lightweight processing on the target temperature inversion model to obtain a lightweight temperature inversion model suitable for operation on the satellite terminal, and then deploys it to the satellite terminal. The satellite acquires real-time remote sensing images and mission location information related to the surface temperature inversion mission; and determines the target image area from the real-time remote sensing images based on the mission location information. The satellite terminal performs temperature inversion on the target image area using the lightweight temperature inversion model to obtain the surface temperature inversion result; and then transmits the surface temperature inversion result to the ground terminal.

2. The method for inverting on-board surface temperature as described in claim 1, characterized in that, The remote sensing image data includes thermal infrared remote sensing images and the corresponding historical remote sensing sequences; the meteorological data includes at least one of time data and weather data; the multi-source data also includes at least one of topographic elevation data, regional type data, and surface temperature annotation data.

3. The method for inverting on-board surface temperature as described in claim 2, characterized in that, Before training the pre-built initial temperature inversion model using the multi-source data at the ground end, the method further includes: The metadata and calibration data corresponding to the thermal infrared remote sensing image are analyzed, and the original digital quantization value of the thermal infrared remote sensing image is converted into a radiance value based on the calibration data. Based on the geographic coordinate information in the metadata, the latitude and longitude coordinates of the sample points are converted into pixel coordinates in the thermal infrared remote sensing image. Based on the pixel coordinates, local image samples are cropped from the remote sensing image data, and effective region masks for excluding invalid regions and effective label masks for limiting the training loss calculation region are generated for the local image samples.

4. The method for inverting on-board surface temperature as described in claim 2, characterized in that, The ground-based system trains a pre-built initial temperature retrieval model using the multi-source data, including: Obtain the task information for the surface temperature inversion task; The remote sensing image data and the task information are semantically enhanced using a remote sensing visual language model to obtain semantically enhanced information; the semantically enhanced information includes at least one of scene semantic information, terrain semantic information, region type semantic information, and anomalous region prior information. The semantic enhancement information is encoded to obtain remote sensing semantic conditions; Based on the multi-source data and the remote sensing semantic conditions, the initial temperature inversion model is iteratively trained.

5. The method for inverting on-board surface temperature as described in claim 4, characterized in that, The initial temperature inversion model includes an encoder and a diffusion Transformer backbone network; the encoder includes an image encoder, a temporal encoder, a terrain encoder, and a conditional encoder; The image encoder is used to encode the thermal infrared remote sensing image; The time-series encoder is used to encode the historical remote sensing sequence; The terrain encoder is used to encode the terrain elevation data; The condition encoder is used to encode the meteorological data, the regional type data, and the remote sensing semantic conditions to obtain condition information; wherein, the condition information is injected into the diffusion Transformer backbone network through a classifier-free guidance mechanism. The diffusion Transformer backbone network employs a multi-path feature token fusion mechanism to fuse the features output by each encoder and output the land surface temperature prediction result.

6. The method for inverting on-board surface temperature as described in claim 1, characterized in that, The lightweight temperature inversion model includes a first inference model and a second inference model; wherein, the first inference model includes at least one of a multilayer perceptron model, a linear regression model, and an XGBoost model; and the second inference model includes at least one of a U-Net model and a lightweight diffusion Transformer student model. The method further includes: If the satellite is in a low-power state or the computing load is greater than the preset load, the first inference model is invoked to perform temperature inversion on the target image area. If the available computing resources on the satellite meet the preset resource conditions, the second inference model is invoked to perform temperature inversion on the target image area.

7. The method for inverting on-board surface temperature as described in any one of claims 1 to 6, characterized in that, The surface temperature inversion results include at least one of the following: target point temperature value, local temperature prediction map, temperature data confidence level, and thermal anomaly labeling; the method further includes: The ground terminal acquires target temperature data including either measured temperature data or preset reference temperature data. The ground terminal compares the surface temperature inversion results with the target temperature data to obtain temperature error analysis results; Based on the error analysis results, the ground terminal updates the model training samples or adjusts the model parameters to obtain the updated lightweight temperature inversion model. The ground terminal deploys the updated lightweight temperature inversion model to the satellite terminal.

8. A space-ground cooperative system, characterized in that, The system includes a ground terminal and a satellite terminal; the ground terminal includes a multi-source data acquisition module, a model training module, and a lightweight module; the satellite terminal includes a target area determination module and a temperature inversion module. The multi-source data acquisition module is used to acquire multi-source data for the land surface temperature retrieval task; the multi-source data includes at least remote sensing image data and meteorological data. The model training module is used to train the pre-constructed initial temperature inversion model using the multi-source data to obtain the target temperature inversion model. The lightweight module is used to perform lightweight processing on the target temperature inversion model to obtain a lightweight temperature inversion model suitable for operation on the satellite, and then deploy it to the satellite. The target area determination module is used to acquire real-time remote sensing images and task location information related to the land surface temperature inversion task; and to determine the target image area from the real-time remote sensing images based on the task location information. The temperature inversion module is used to perform temperature inversion on the target image area using the lightweight temperature inversion model to obtain the surface temperature inversion result; and to transmit the surface temperature inversion result to the ground terminal.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for inverting the surface temperature on the satellite according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for inverting the surface temperature on the satellite according to any one of claims 1 to 7.

Images