A star-ground collaborative carbon sink monitoring method based on multi-core heterogeneous acceleration

Abstract

The present application relates to a kind of star-ground collaborative carbon sink monitoring methods based on multi-core heterogeneous acceleration, including acquisition ground multi-source time series data and satellite multispectral image form star-ground data;Star-ground data is preprocessed, and C-Frame tensor with self-description frame header is generated;Collaborative flow self-adaptive bridge C-FAB framework is established, for C-Frame tensor is directly mapped to edge ai equipment, avoid traditional multiple memory copy;DS-LUE model is established in edge ai equipment, and the carbon sink result of the calculation area is generated by DS-LUE model, by fusing satellite multispectral image and ground humidity, soil and vegetation sensing data, in low power consumption, high throughput edge computing environment, fast, accurate carbon sink estimation and dynamic monitoring are realized, and support complex remote sensing processing task in different computing resources collaborative operation, significantly improve the performance and adaptability of remote sensing data processing.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous computing system architecture and scheduling methods thereof, and in particular to a satellite-ground collaborative carbon sink monitoring method based on multi-core heterogeneous acceleration. Background Art

[0002] Carbon sink monitoring is a key link in supporting the "dual carbon" and production-life carbon accounting; terrestrial ecosystems such as forests, farmlands, and grasslands are widely distributed and highly heterogeneous, and their carbon fixation rates change rapidly with seasons, weather, and management measures; to obtain real, temporally and spatially comparable carbon sink data, it is necessary to simultaneously utilize the wide-area coverage capabilities of satellite multispectral images and the near-ground fine observation capabilities of ground sensor networks, and complete the fusion operations of large-scale heterogeneous data within a semi-real-time scale.

[0003] However, existing technologies still have several key limitations: on the one hand, pure satellite inversion methods generally adopt an empirical formula with NDVI as the core and a fixed leaf light utilization efficiency (LUE), ignoring the inhibitory and promoting effects of soil moisture and air humidity on photosynthesis, and the regional average error is often higher than 5%; on the other hand, the traditional cloud-based heterogeneous architecture of "FPGA pre-processing → DDR → CPU copy → GPU inference" has to go through multiple memory copies and bus moves, and the bandwidth bottleneck and power consumption are significantly superimposed, resulting in a single-cycle processing delay of tens of minutes, making it difficult to achieve high-frequency observations in forest areas and agricultural areas with limited power and network conditions; at the same time, although ground flux towers or artificial sample plots can obtain high-precision data, they are limited by spatial coverage and cost and cannot meet the needs of large-scale applications. Summary of the Invention

[0004] The present invention aims to address key bottlenecks in the current field of satellite-ground collaborative carbon sink detection, such as the difficulty in spatiotemporal alignment of multi-source data, the high overhead of moving heterogeneous computing power, the insufficient accuracy of leaf light energy utilization models, and the high power consumption of edge deployment, and provide a satellite-ground collaborative carbon sink monitoring method based on multi-core heterogeneous acceleration.

[0005] To solve the above technical problems, the present invention adopts the following technical solution: a satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration, comprising the following steps:

[0006] S1. Collect ground multi-source time series data and satellite multispectral images to form satellite-ground data;

[0007] S2, preprocess the satellite-ground data to generate a C-Frame tensor with a self-describing frame header;

[0008] S3. Establish a collaborative flow adaptive bridge C-FAB framework with a built-in peer-to-peer direct DMA transmission submodule to directly map C-Frame tensors to edge AI devices;

[0009] S4. Establish a DS-LUE model in the edge AI device. The DS-LUE model is configured as follows:

[0010] Perform branch feature extraction on the data output by edge AI devices to obtain spatial-spectral domain features and physiological features;

[0011] Weighted splicing of spatial-spectral domain features and physiological features into a fusion tensor;

[0012] Perform two layers of 2D convolution on the fused tensor to obtain the light energy utilization rate map of the shade leaf and the sun leaf;

[0013] Calculate regional carbon sinks based on the light energy utilization rate maps of shade leaves and sun leaves.

[0014] Preferably, in S1, the required ground multi-source time series data is acquired in real time through the ground terminal sensor, wherein the ground multi-source time series data includes soil volume moisture content, relative humidity, leaf area index and canopy temperature, and each sensor is calibrated with a sampling time stamp through a GNSS-synchronized 1 PPS signal;

[0015] The satellite multispectral image includes satellite image blocks of at least 8 bands, and the file header of the image carries the imaging start time, orbit attitude and imaging center coordinates.

[0016] Further preferably, the satellite multispectral image includes blue, green, red, near infrared, shortwave infrared 1 / 2, thermal infrared and panchromatic bands.

[0017] Preferably, the space-time alignment and synchronization of satellite-to-ground data on the FPGA board includes:

[0018] Taking the global positioning system time GPS-Time as the benchmark, the formula is:

[0019] ;

[0020] Among them, Tsat is the imaging start time, and Tsensor is the sensor data acquisition time.

[0021] when When the sensor data is missing, linear interpolation or spline interpolation is used to fill in the missing sampling points to ensure that the satellite and ground data correspond within the same observation period;

[0022] The WGS-8 UTM projection of the satellite image is converted to the same plane rectangular coordinate system as the ground sensor layout, and the pixel-point mapping relationship is established using the nearest neighbor or bilinear interpolation method between the sensor latitude and longitude and the image pixel center coordinates;

[0023] A 32-byte self-describing frame header is written into the BRAM on the FPGA board for each set of registered satellite-ground data. The frame header is written into the pipeline FIFO along with the data and is refreshed synchronously on the rising edge of 1 PPS to form a spatiotemporally consistent C-Frame tensor.

[0024] The frame header fields include the source identifier SrcID, the band mask BandMask, the unified timestamp TimeTag, the space block number TileID, the priority Prio and the CRC16 check.

[0025] Preferably, in S2, the preprocessing includes:

[0026] Radiometric calibration and atmospheric correction of satellite multispectral images to generate surface reflectance ;

[0027] Adjusting the surface reflectivity A 3×3 median filter is implemented to remove impulse noise, and then the edge details are enhanced by the first-order differential Laplacian operator to retain high-resolution texture for subsequent convolution processing;

[0028] Use parallel pipeline logic to calculate the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), and Soil Moisture Index (SWI) in real time;

[0029] Based on the pixel-point mapping relationship, the ground multi-source time series data and satellite pixel features are matched according to the unified time stamp TimeTag; when there is a time difference between satellite and ground sampling When , linear interpolation is used to fill in the missing values ​​to form a fusion feature vector F:

[0030] ;

[0031] in, is the soil volume moisture content, is the air humidity, is the leaf area index;

[0032] After normalizing each component of vector F to 0-1 according to the preset maximum-minimum range, it is converted to 16-bit fixed-point format and the normalization coefficients and offsets are written to the Lookup Table for back-end dequantization.

[0033] The normalized multispectral-ground fusion feature vector is rearranged into a C-Frame tensor according to row and column coordinates. It is then losslessly compressed in real time using the LZ4 hardware compression core embedded in the FPGA. The compressed tensor and its corresponding 32-byte self-describing frame header are written into a dual-ended FIFO.

[0034] Further preferably, radiation calibration and atmospheric correction are performed: in the on-chip multiplication and addition array of the FPGA, radiation calibration is completed pixel by pixel for the blue, green, red, near-infrared, short-wave infrared and thermal infrared bands according to the calibration coefficients provided by the satellite, and the atmospheric correction of the MODTRAN simplified model is performed by calling the table lookup method to generate the surface reflectivity.

[0035] Further preferably, the Normalized Difference Vegetation Index NDVI, Enhanced Vegetation Index EVI, Soil Adjusted Vegetation Index SAVI and Soil Moisture Index , the calculation formulas are as follows:

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] Among them, L=0.5, NIR is the near infrared band, RED is the red light band, BLUE is the blue light band, SWIR1 is shortwave infrared 1, and SWIR2 is shortwave infrared 2.

[0041] Preferably, in S3, the PCIe Gen3×8 Root-Complex and AXI-MM bridge IP are instantiated in the FPGA board, and the peer-to-peer direct DMA transfer submodule's peer access base address register BAR is enabled. p2p ; By configuring DMA descriptors , directly transfer the C-Frame tensor from the FPGA on-chip DDR3 address Address transmitted to edge AI processor , its single frame delay is:

[0042] ;

[0043] Where S is the amount of compressed data and B is the measured bandwidth.

[0044] Preferably, the collaborative flow adaptive bridge C-FAB framework also includes a self-describing frame header parsing and disorder reordering submodule, which is configured to: when the C-Frame tensor arrives at the edge AI device, the frame header parsing logic is based on the field<TimeTag,TileID,Prio> Calculate hash key ,locate the buffer in the on-chip and off-chip joint hash table.

[0045] Preferably, the collaborative flow adaptive bridge C-FAB framework also includes a temperature-power-time adaptive scheduling submodule for real-time acquisition of edge AI device chip temperature. , peer-to-peer direct DMA transfer submodule bandwidth utilization NPU utilization of edge AI devices , and establish the cost function Minimize the time within 5ms:

[0046] ;

[0047] in, is the temperature threshold, , , are the weight coefficients of temperature, bandwidth, and NPU utilization on scheduling decisions; or When writing register Reg route Adjust the routing and switch subsequent frames to the FPGA local operator chain, and then switch back to the edge AI device when the temperature drops or the bandwidth is idle.

[0048] Preferably, the collaborative stream adaptive bridge C-FAB framework also includes a reconfigurable AXI-Stream cross router, which includes a 4×128-bit cross switch and a 16×32-bit AXI-Lite configuration register, and supports a dual mode of fixed routing at compile time and remapping at run time; when the temperature-power-time adaptive scheduling submodule decides to keep some tasks to be executed locally on the FPGA, the router is configured according to the register Reg route As shown, the C-Frame tensor is sent directly to the FPGA internal operator.

[0049] Preferably, in S4, the channel attention mechanism is used to calculate the weight vector to convert the spatial-spectral domain features and physiological characteristics Weighted concatenation into fused tensor :

[0050] ;

[0051] ;

[0052] in," " indicates channel-dimensional splicing, + =1, represents the weighted spatial-spectral domain features, Represents the weighted physiological features. This step realizes the dual-source attention fusion: when the satellite spectral domain features are more discriminative, On the contrary, when the ground moisture / leaf area index signal is more sensitive, it dominates, thereby dynamically adjusting the contribution of the two-source information to carbon sink inversion under different seasons or climatic conditions.

[0053] Further preferably, the output data transmitted to the edge AI processor via the C-FAB framework peer-to-peer DMA transmission submodule is the input of DS-LUE in the form of a compressed tensor C-Frame sorted by rows and columns with a 32B frame header; it is first decompressed by the on-chip LZ4 hardware on the Ascend side and restored to the shape of The fused feature tensor of , where C = 16 (8 spectral + 4 exponential + 4 ground parameters);

[0054] Further preferably, a satellite branch 3D-CNN is constructed: two-level 3D convolution-pooling is performed on 8 spectral channels to obtain spatial-spectral domain features ; Construct ground branch MLP: perform three-layer full connection on parameter meta-features such as soil volume moisture content, air humidity, leaf area index, etc. of ground multi-source time series data, and output physiological characteristics .

[0055] Preferably, in S4, the fused tensor Perform two layers of 2D convolution to obtain the light energy utilization rate map of shade leaves and sun leaves 、 :

[0056] ;

[0057] And upsampled to the original resolution, we get ;

[0058] Calculate regional carbon sinks on the FPGA board:

[0059] ;

[0060] ;

[0061] ;

[0062] in, is a regional carbon sink. GPP refers to the amount of organic carbon fixed per unit time by plant communities during photosynthesis before deducting their own respiratory consumption. f represents the absorption ratio of PAR by the vegetation canopy. PAR(x,y) refers to the irradiance at the pixel (x,y). NPP is the net carbon fixation after deducting the plant's own maintenance and growth respiration (Ra) from GPP. is the reflectance-photosynthetically active radiation conversion factor, The respiration coefficient is set according to the plant functional type.

[0063] Beneficial effect: The present invention cooperates with the C-FAB framework of the collaborative stream adaptive bridge + DS-LUE model. On the one hand, by integrating peer-to-peer direct DMA, frame header reordering, reconfigurable AXI-Stream routing and temperature-power-time scheduling in the FPGA, it realizes "zero-copy" bidirectional transmission of pre-processed tensors between the FPGA and the edge AI processor, reducing the single-frame latency by about 35% and the power consumption of the whole machine by about 22%; on the other hand, by constructing a dual branch of satellite 3D-CNN and ground MLP and introducing a channel attention mechanism, it adaptively fuses the spectrum with humidity, soil, and vegetation physiological parameters, and outputs a dynamic estimate of the LUE of shade leaves and sun leaves as the environment changes, reducing the annual average NPP error by about 11% in the actual measurement scenario; therefore, the present invention can achieve real-time carbon sink monitoring of no less than 15 frames per second in a 4096×4096×8 band imaging scene, taking into account accuracy, energy efficiency and deployment flexibility, and has significant engineering application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a satellite-ground collaborative carbon sink monitoring method based on multi-core heterogeneous acceleration according to the present invention;

[0065] Figure 2 This is the C-FAB framework collaborative flow adaptive bridge flow framework diagram of the present invention;

[0066] Figure 3 This is the flow chart of the DS-LUE model of the present invention. DETAILED DESCRIPTION

[0067] To make the objectives and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is merely intended to describe a satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration or several specific implementations of the present invention, and does not strictly limit the scope of protection specifically claimed in the present invention.

[0068] Example 1: The technical solution adopted by the present invention is as follows Figure 1 As shown, a satellite-ground collaborative carbon sink monitoring method based on multi-core heterogeneous acceleration includes the following steps:

[0069] S1. Collect ground multi-source time series data and satellite multispectral images to form satellite-ground data;

[0070] S11. Collect multi-source time series data on the ground and use soil moisture sensors deployed in the target area to obtain soil volume moisture content. , use the air humidity sensor to obtain relative humidity ; Use vegetation monitoring terminal to obtain leaf area index and canopy temperature Each sensor is sampled and time-stamped using a GNSS-synchronized 1 PPS signal.

[0071] S12. Receive satellite remote sensing images (Landsat 8 for example), including satellite image blocks of at least 8 bands, including blue, green, red, near infrared, shortwave infrared 1 / 2, thermal infrared, and panchromatic bands. The image file header carries the imaging start time Tsat, orbital attitude, and imaging center coordinates. ;

[0072] The spatiotemporal alignment and synchronization of satellite and ground data is achieved on an FPGA board (the Unisplendour Kosmo-2 K400 FPGA is used as an example in this example):

[0073] S13. Using the global positioning system time GPS-Time as a reference, use the formula:

[0074] ;

[0075] Among them, Tsat is the imaging start time, and Tsensor is the sensor data acquisition time.

[0076] when When the sensor data is missing, linear interpolation or spline interpolation is used to fill in the missing sampling points to ensure that the satellite and ground data correspond within the same observation period;

[0077] S14. Convert the WGS-8 coordinates (the most commonly used geocentric coordinate datum and ellipsoid parameters in the world) of the satellite image to the same plane rectangular coordinate system as the ground sensor layout, using the sensor latitude and longitude and the image pixel center coordinates ( )’s nearest neighbor or bilinear interpolation method to establish pixel-point mapping relationship;

[0078] S15. Write a 32-byte self-describing frame header for each set of aligned satellite-ground data in the BRAM on the FPGA board. The frame header fields include: source identifier SrcID (8 bits), band mask BandMask (5 bits), unified timestamp TimeTag (48 bits), spatial block sequence number TileID (12 bits), priority Prio (2 bits) and CRC16 checksum; the frame header is written into the pipeline FIFO together with the data and is synchronously refreshed at the rising edge of 1 PPS to form a temporally consistent C-Frame tensor, providing a unified input for subsequent preprocessing;

[0079] S2, preprocess the satellite-ground data to generate a C-Frame tensor with a self-describing frame header;

[0080] S21, radiometric calibration and atmospheric correction: In the on-chip multiplication and addition array of the FPGA board, radiometric calibration is completed pixel by pixel for the blue, green, red, near infrared, short-wave infrared and thermal infrared bands according to the calibration coefficients provided by the satellite, and the atmospheric correction of the MODTRAN simplified model is performed by calling the table lookup method to generate the surface reflectance. ;

[0081] S22, adjust the surface reflectivity A 3×3 median filter is implemented to remove impulse noise, and then the edge details are enhanced by the first-order differential Laplacian operator to retain high-resolution texture for subsequent convolution processing;

[0082] S23, using parallel pipeline logic to calculate the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Soil Adjusted Vegetation Index (SAVI) in real time And soil moisture index SWI, the calculation formulas are as follows:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] Among them, L=0.5; NIR is the near infrared band, RED is the red light band, BLUE is the blue light band, SWIR1 is shortwave infrared 1, and SWIR2 is shortwave infrared 2.

[0088] S24, based on the pixel-point mapping relationship, the soil volume moisture content , air humidity , leaf area index The ground parameters and satellite pixel features are matched according to the unified time stamp TimeTag; when there is a time difference between satellite and ground sampling When , linear interpolation is used to fill in the missing values ​​to form a fusion feature vector F:

[0089] ;

[0090] S25. After normalizing each component in vector F to 0-1 according to a preset maximum-minimum range, convert it to a 16-bit fixed-point format, and write the normalization coefficient and offset into the Lookup Table for back-end dequantization.

[0091] S26, rearrange the normalized multispectral-ground fusion feature vector into a C-Frame tensor according to row and column coordinates, perform real-time lossless compression using the LZ4 hardware compression core embedded in the FPGA, and write the compressed tensor and its corresponding 32-byte self-describing frame header into a dual-ended FIFO;

[0092] S3. Establish the collaborative flow adaptive bridge C-FAB framework, refer to Figure 2 As shown, a built-in peer-to-peer direct DMA transfer submodule is used to map the C-Frame tensor directly to the LPDDR4 of the edge AI device (in this embodiment, Huawei Ascend 310 AI acceleration board is used as an example);

[0093] S31, peer-to-peer direct DMA transfer submodule (PCIe Gen3×8 DMA controller configured as peer-to-peer mode):

[0094] Instantiate the PCIe Gen3×8 Root-Complex and AXI-MM bridge IP on the FPGA board, enabling peer-to-peer (P2P) access to the base address register (BAR). p2p ; By configuring DMA descriptors , directly transfer the C-Frame tensor from the FPGA on-chip DDR3 address Address transmitted to edge AI processor , its single frame delay is:

[0095] ;

[0096] Where S is the amount of compressed data and B is the measured bandwidth;

[0097] S32, self-description frame header parsing and out-of-order reordering submodule:

[0098] When the C-Frame tensor arrives at Ascend 310 via P2P-DMA transmission, the frame header parsing logic is based on the field<TimeTag,TileID,Prio> Calculate hash key , locate the buffer in the on-chip and off-chip joint hash table, TimeID is the spatial block number (12 bits), indicating the spatial block to which the C-Frame tensor belongs, and Prio indicates the importance ranking level of the spatial block in subsequent processing;

[0099] S33, Reconfigurable AXI-Stream Crossbar Router:

[0100] The router includes 4×128-bit crossbar switches and 16×32-bit AXI-Lite configuration registers, supporting "compile-time fixed routing + run-time remapping" dual modes; when the Tiny-RISC scheduling core decides to keep some tasks on the FPGA for local execution, the router registers Reg route As shown, the C-Frame is sent directly to the internal operator of FPGA;

[0101] S34, temperature-power-time adaptive scheduling submodule (Tiny-RISC scheduling core):

[0102] Collect Ascend chip temperature during scheduling verification , P2P bandwidth utilization and NPU utilization , and minimized within a 5 ms period using the following function as the cost function:

[0103] ;

[0104] in, is the temperature threshold, usually 85°C, , , are the weight coefficients that measure the importance of temperature, bandwidth, and NPU utilization to the scheduling decision; or When writing register Reg route Adjust the routing and switch subsequent frames to the FPGA local operator chain. Switch back to Ascend when the temperature drops or the bandwidth is idle.

[0105] The output data transmitted to the edge AI processor through the C-FAB framework peer-to-peer DMA transmission submodule is the input of DS-LUE in the form of a compressed tensor C-Frame sorted by rows and columns with a 32B frame header; it is first decompressed by the on-chip LZ4 hardware on the Ascend side and restored to the shape of The fused feature tensor of , where C = 16 (8 spectral + 4 index + 4 ground parameters);

[0106] S4. Build the DS-LUE model in the edge AI device, refer to Figure 3 As shown, the DS-LUE model is configured as follows:

[0107] S41. Perform branch feature extraction on the data output by the edge AI device to obtain spatial-spectral domain features and physiological features;

[0108] Constructing a satellite branch 3D-CNN: Performing two-stage 3D convolution-pooling on 8 spectral channels to obtain spatial-spectral features ;

[0109] Constructing the ground branch MLP: The equal parameter meta-features perform three-layer full connection and output physiological features ;

[0110] S42, weighting and concatenating the spatial-spectral domain features and physiological features into a fusion tensor;

[0111] Use the channel attention mechanism to calculate the weight vector W, and integrate the spatial-spectral features and physiological characteristics Weighted concatenation into fused tensor :

[0112] ;

[0113] ;

[0114] in," " indicates channel-dimensional splicing, + =1, represents the weighted spatial-spectral domain features, Represents the weighted physiological features. This step realizes the dual-source attention fusion: when the satellite spectral domain features are more discriminative, On the contrary, when the ground moisture / leaf area index signal is more sensitive, it dominates, thereby dynamically adjusting the contribution of the two-source information to carbon sink inversion under different seasons or climatic conditions.

[0115] S43. Fusion Tensor Perform two layers of 2D convolution to obtain the light energy utilization rate map of shade leaves and sun leaves :

[0116] ;

[0117] And upsampled to the original resolution, we get ;

[0118] S44. Calculate regional carbon sinks on the FPGA board:

[0119] ;

[0120] ;

[0121] ;

[0122] in, is a regional carbon sink. GPP refers to the amount of organic carbon fixed per unit time by plant communities during photosynthesis before deducting their own respiratory consumption. f represents the absorption ratio of PAR by the vegetation canopy. PAR(x,y) refers to the irradiance at the pixel (x,y). NPP is the net carbon fixation after deducting the plant's own maintenance and growth respiration (Ra) from GPP. is the reflectance-photosynthetically active radiation conversion factor, The respiration coefficient is set according to the plant functional type.

[0123] In a specific case, the method of the present invention was experimentally verified in an agro-forestry mixed ecological experimental area, with the experimental period being from March 2024 to November 2024;

[0124] 1. Experimental equipment and conditions:

[0125] Satellite data: Landsat-8 satellite remote sensing images are used, with a spatial resolution of 30m and an imaging period of 16 days.

[0126] Ground data: Soil moisture sensors (SM150T, 30 locations), air humidity sensors (HMP155, 10 locations), and vegetation monitoring terminals (LI-COR LAI-2200, 10 locations) were used, with data sampling frequencies of once per hour and once per day, respectively.

[0127] Computing platform: Inventive platform (Zhiguang Tongchuang Kosmo-2 K400 FPGA + Huawei Ascend 310 AI accelerator); traditional comparison platform (Xilinx Virtex UltraScale FPGA + Intel Xeon E5 CPU + NVIDIA Tesla T4 GPU).

[0128] 2. Experimental methods:

[0129] The following methods were used for comparative experiments:

[0130] Method 1 (method of the present invention): FPGA (Kosmo-2 K400) completes preprocessing and directly sends data to the Ascend 310 chip through PCIe Gen3×8 DMA to run the dual-source leaf light energy utilization efficiency (DS-LUE) model.

[0131] Method 2 (traditional CPU + GPU heterogeneous method): FPGA (Xilinx Virtex UltraScale) preprocesses the data, which is copied multiple times to DDR memory. The CPU (Intel Xeon E5) performs further preprocessing before sending it to the GPU (NVIDIA Tesla T4) to run the traditional fixed light energy utilization model.

[0132] Method 3 (pure GPU cloud method): Satellite raw data and ground data are uniformly transmitted to the cloud server (CPU+GPU architecture) via the network. After being pre-processed by the CPU (Intel Xeon E5), they are sent to the GPU (Tesla T4) to run the traditional fixed light energy utilization model.

[0133] 3. Experimental results and analysis, as shown in 1:

[0134] Table 1 Performance index comparison

[0135]

[0136] Compared with Method 2, the method of the present invention reduces the single-frame data processing latency by 36.27%, increases the data transmission bandwidth by 51.60%, and reduces overall power consumption by 22.22%. Compared with the cloud-based method, the latency advantage of the present invention is even more significant, with a significant reduction in power consumption of 66.4%.

[0137] Table 2 Comparison of carbon sink estimation accuracy (NPP annual average)

[0138]

[0139] As can be seen from Tables 1 and 2 above, the method of the present invention is significantly superior to the traditional CPU+GPU heterogeneous method and the pure GPU cloud method in terms of processing performance, power consumption, data transmission efficiency and carbon sink monitoring accuracy, and has obvious practical application value and promotion prospects.

[0140] The above describes the implementation mode of the present invention in detail with reference to the embodiments. However, the present invention is not limited to the above implementation mode. After knowing the contents described in the present invention, ordinary technicians in this technical field can make several equivalent transformations and substitutions without departing from the principles of the present invention. These equivalent transformations and substitutions should also be regarded as falling within the scope of protection of the present invention.

Claims

1. A satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration, characterized by: The steps include: S1. Collect ground multi-source time series data and satellite multispectral images to form satellite-ground data; In S1, the required ground multi-source time series data are acquired in real time through ground terminal sensors. The ground multi-source time series data includes soil volume moisture content, relative humidity, leaf area index, and canopy temperature. Each sensor is calibrated with a sampling time stamp through a GNSS-synchronized 1 PPS signal. The satellite multispectral image includes at least 8-band satellite image blocks, and the image file header carries the imaging start time, orbit attitude and imaging center coordinates; S2, preprocess the satellite-ground data to generate a C-Frame tensor with a self-describing frame header; S3. Establish a collaborative flow adaptive bridge C-FAB framework with a built-in peer-to-peer direct DMA transmission submodule to directly map C-Frame tensors to edge AI devices; In S3, the PCIe Gen3×8 Root-Complex and AXI-MM Bridge IP are instantiated in the FPGA board, enabling peer-to-peer access to the base address register (BAR). p2p ; By configuring DMA descriptors , directly transfer the C-Frame tensor from the FPGA on-chip DDR3 address Address transmitted to edge AI processor , its single frame delay is: ； Where S is the amount of compressed data and B is the measured bandwidth; The collaborative flow adaptive bridge C-FAB framework also includes a self-describing frame header parsing and out-of-order reordering submodule, which is configured to: when the C-Frame tensor arrives at the edge AI device, the frame header parsing logic is based on the field<TimeTag,TileID,Prio> Calculate hash key ,locate the buffer in the on-chip and off-chip joint hash table; The collaborative flow adaptive bridge C-FAB framework also includes a temperature-power-time adaptive scheduling submodule for real-time acquisition of edge AI device chip temperature. , peer-to-peer direct DMA transfer submodule bandwidth utilization NPU utilization of edge AI devices , and establish the cost function : ； in, is the temperature threshold, , , are the weight coefficients of temperature, bandwidth, and NPU utilization on scheduling decisions; or When writing register Reg route Adjust the routing and switch subsequent frames to the FPGA local operator chain, and then switch back to the edge AI device when the temperature drops or the bandwidth is idle; S4. Establish a DS-LUE model in the edge AI device. The DS-LUE model is configured as follows: Perform branch feature extraction on the data output by edge AI devices to obtain spatial-spectral domain features and physiological features; Weighted splicing of spatial-spectral domain features and physiological features into a fusion tensor; Perform two layers of 2D convolution on the fused tensor to obtain the light energy utilization rate map of the shade leaf and the sun leaf; Calculate regional carbon sinks based on the light energy utilization rate maps of shade leaves and sun leaves; In S4, a satellite branch 3D-CNN is constructed and two-level 3D convolution-pooling is performed to obtain spatial-spectral domain features. ;Build ground branch MLP, perform three-layer full connection, and output physiological features ; Use channel attention mechanism to calculate weight vector to integrate spatial-spectral features and physiological characteristics Weighted concatenation into fused tensor : ； ； Among them, "∥" represents channel dimension splicing, + =1, represents the weighted spatial-spectral domain features, Represents the weighted physiological characteristics.

2. The satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration according to claim 1 is characterized by: Perform spatial and temporal alignment and synchronization of satellite and ground data on the FPGA board, including: Taking the global positioning system time GPS-Time as the benchmark, the formula is: ； Among them, Tsat is the imaging start time, and Tsensor is the sensor data acquisition time; when When the sensor data is missing, linear interpolation or spline interpolation is used to fill in the missing sampling points to ensure that the satellite and ground data correspond within the same observation period; The WGS-8 UTM projection of the satellite image is converted to the same plane rectangular coordinate system as the ground sensor layout, and the pixel-point mapping relationship is established using the nearest neighbor or bilinear interpolation method between the sensor latitude and longitude and the image pixel center coordinates; A 32-byte self-describing frame header is written into the BRAM on the FPGA board for each set of registered satellite-ground data. The frame header is written into the pipeline FIFO along with the data and is refreshed synchronously on the rising edge of 1 PPS to form a spatiotemporally consistent C-Frame tensor. The frame header fields include the source identifier SrcID, the band mask BandMask, the unified timestamp TimeTag, the space block number TileID, the priority Prio and the CRC16 check.

3. The satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration according to claim 2 is characterized by: In S2, the preprocessing includes: Radiometric calibration and atmospheric correction of satellite multispectral images to generate surface reflectance ; Adjusting the surface reflectivity A 3×3 median filter is implemented to remove impulse noise, and then the first-order differential Laplacian operator is used to enhance edge details and retain high-resolution textures; Use parallel pipeline logic to calculate the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), and Soil Moisture Index (SWI) in real time; Based on the pixel-point mapping relationship, the ground multi-source time series data and satellite pixel features are matched according to the unified time stamp TimeTag; when there is a time difference between satellite and ground sampling When , linear interpolation is used to fill in the missing values ​​to form a fusion feature vector F: ； in, is the soil volume moisture content, is the air humidity, is the leaf area index; After normalizing each component of vector F to 0-1 according to the preset maximum-minimum range, it is converted to 16-bit fixed-point format and the normalization coefficients and offsets are written to the Lookup Table for back-end dequantization. The normalized multispectral-ground fusion feature vector is rearranged into a C-Frame tensor according to row and column coordinates. It is then losslessly compressed in real time using the LZ4 hardware compression core embedded in the FPGA. The compressed tensor and its corresponding 32-byte self-describing frame header are written into a dual-ended FIFO.

4. The satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration according to claim 1 is characterized by: The collaborative stream adaptive bridge C-FAB framework also includes a reconfigurable AXI-Stream cross router, which includes a 4×128-bit cross switch and a 16×32-bit AXI-Lite configuration register, supporting a dual mode of fixed routing at compile time and remapping at run time; when the temperature-power-time adaptive scheduling submodule decides to keep some tasks on the FPGA for local execution, the router is configured according to the register Reg route As shown, the C-Frame tensor is sent directly to the FPGA internal operator.

5. The satellite-ground coordinated carbon sink monitoring method based on multi-core heterogeneous acceleration according to claim 4 is characterized by: In S4, for fused tensors Perform two layers of 2D convolution to obtain the light energy utilization rate map of shade leaves and sun leaves : ； And upsampled to the original resolution, we get ; Calculate regional carbon sinks on the FPGA board: ； ； ； in, is a regional carbon sink. GPP refers to the amount of organic carbon fixed per unit time by plant communities during photosynthesis before deducting their own respiratory consumption. f represents the absorption ratio of PAR by the vegetation canopy. PAR(x,y) refers to the irradiance at the pixel (x,y). NPP is the net carbon fixed after deducting the plant's own maintenance and growth respiration from GPP. is the reflectance-photosynthetically active radiation conversion factor, is the breathing coefficient.

Images