Radiometer multi-sensor data fusion preprocessing method for embedded platform

Abstract

The application discloses a radiometer multi-sensor data fusion preprocessing method for an embedded platform and relates to the technical field of embedded measurement. The method comprises the following steps: S1, controlling a scanning mechanism to perform a scanning motion; S2, in a back-scan stroke, calculating absolute gain estimation of a first frequency band receiving channel based on cold space brightness temperature radiation, and in an observation stroke, obtaining brightness temperature observation values of the first frequency band receiving channel and a second frequency band receiving channel based on scene brightness temperature radiation; S3, in the observation stroke, taking the absolute gain estimation of the first frequency band receiving channel as real-time gain of the first frequency band receiving channel, and setting real-time gain of the second frequency band receiving channel as a product of the real-time gain of the first frequency band receiving channel and a pre-calibrated inter-frequency transfer coefficient; and S4, periodically extending the back-scan stroke to calculate absolute gain estimation of the second frequency band receiving channel, and correcting the inter-frequency transfer coefficient based on a deviation of the absolute gain estimation of the second frequency band receiving channel and the product.

Description

Technical Field

[0001] This application relates to the field of embedded measurement technology, and more specifically, to a preprocessing method for multi-sensor data fusion of radiometers for embedded platforms. Background Technology

[0002] Radiometers, as passive microwave remote sensing devices, receive scene radiation energy to invert the physical characteristics of targets, and have significant application value in meteorological observation, resource exploration, and environmental monitoring. With the continuous improvement of observation requirements, single-band radiometers can no longer meet the accuracy requirements for multi-dimensional characterization of ground features; multi-sensor, multi-band collaborative observation has become the mainstream development trend. Integrating multi-band radiometer systems into embedded platforms can effectively improve the richness of data acquisition and the flexibility of system deployment, but it places higher demands on the real-time performance and consistency of front-end data. Multi-sensor data fusion preprocessing, as a key link connecting raw observations and high-precision inversion, directly determines the effectiveness of subsequent information fusion and the reliability of system output.

[0003] In existing technologies, data preprocessing methods for multi-band radiometers typically follow the approach of independent calibration for each channel. This involves establishing a periodic calibration process for each receiving channel and correcting channel gain fluctuations by inserting cold-space or calibration body observations. However, this approach faces certain limitations in embedded platforms. Frequent insertion of independent calibrations can disrupt the normal observation sequence, reducing the system's temporal resolution and observation efficiency.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a radiometer multi-sensor data fusion preprocessing method for embedded platforms to solve the above-mentioned technical problems.

[0006] This application provides a preprocessing method for multi-sensor data fusion of radiometers for embedded platforms. The radiometer includes a first frequency band receiving channel and a second frequency band receiving channel. The method comprises: S1, controlling a scanning mechanism to perform a non-uniform scanning motion, wherein the non-uniform scanning motion includes an observation stroke and a retrace stroke; S2, during the retrace stroke, calculating the absolute gain estimate of the first frequency band receiving channel based on cold air brightness temperature radiation, and during the observation stroke, acquiring the brightness temperature observation values ​​of the first frequency band receiving channel and the second frequency band receiving channel based on scene brightness temperature radiation; S3, during the observation stroke, using the absolute gain estimate of the first frequency band receiving channel as the real-time gain of the first frequency band receiving channel, and setting the real-time gain of the second frequency band receiving channel as the product of the real-time gain of the first frequency band receiving channel and a pre-calibrated inter-frequency transfer coefficient; S4, periodically extending the retrace stroke to calculate the absolute gain estimate of the second frequency band receiving channel, and correcting the inter-frequency transfer coefficient based on the deviation between the absolute gain estimate of the second frequency band receiving channel and the product.

[0007] Based on the embodiments provided in this application, by dividing the movement of the scanning mechanism into an observation stroke and a retrace stroke, and independently completing the absolute gain estimation of the first frequency band receiving channel during the retrace stroke, the gain calibration process and the scene observation process are seamlessly integrated in time. Compared to the prior art, which requires frequent insertion of independent calibration processes, this method effectively avoids the calibration action from encroaching on the continuity of observation under the limited processing resources and observation timing constraints of the embedded platform, thus ensuring the system's temporal resolution and observation efficiency.

[0008] Building upon this foundation, this application establishes a real-time gain correlation mechanism between the first and second frequency band receiving channels during the observation process. Specifically, the real-time gain of the second frequency band receiving channel is determined using the dynamic gain of the first frequency band receiving channel as a benchmark, combined with a pre-calibrated inter-frequency transfer coefficient. This ensures consistency in radiation values ​​between the two channels even without simultaneous independent absolute calibration. Furthermore, by periodically extending the retracement distance, an independent absolute gain estimate for the second frequency band receiving channel is obtained without affecting the conventional observation process. This estimate is then used to perform closed-loop correction on the aforementioned transfer coefficient. This mechanism effectively overcomes the multi-source data mismatch problem caused by channel characteristic differences during independent calibration of discrete channels. It ensures that multi-band radiation data maintains a unified radiation reference benchmark within the embedded platform, providing reliable front-end preprocessing support for subsequent high-precision data fusion. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of an optional radiometer multi-sensor data fusion preprocessing method for an embedded platform according to an embodiment of this application; Figure 2 This is a flowchart of another optional radiometer multi-sensor data fusion preprocessing method for embedded platforms according to an embodiment of this application; Figure 3 This is a flowchart of another optional radiometer multi-sensor data fusion preprocessing method for embedded platforms according to an embodiment of this application; Figure 4 This is a flowchart of another optional radiometer multi-sensor data fusion preprocessing method for embedded platforms according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application.

[0010] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0012] According to one aspect of the embodiments of this application, such as Figure 1 As shown, this application provides a preprocessing method for multi-sensor data fusion of a radiometer for embedded platforms. The radiometer includes a first frequency band receiving channel and a second frequency band receiving channel, comprising: S1 controls the scanning mechanism to perform non-uniform scanning motion, wherein the non-uniform scanning motion includes observation stroke and retrace stroke; S2, During the retrace stroke, the absolute gain estimate of the first frequency band receiving channel is calculated based on the cold air brightness temperature radiation. During the observation stroke, the brightness temperature observation values ​​of the first frequency band receiving channel and the second frequency band receiving channel are obtained based on the scene brightness temperature radiation. S3, During the observation process, the absolute gain estimate of the first frequency band receiving channel is used as the real-time gain of the first frequency band receiving channel, and the real-time gain of the second frequency band receiving channel is set as the product of the real-time gain of the first frequency band receiving channel and the pre-calibrated frequency transfer coefficient. S4, periodically extend the retrace stroke to calculate the absolute gain estimate of the second frequency band receiving channel, and correct the inter-frequency transfer coefficient based on the deviation of the product of the absolute gain estimate of the second frequency band receiving channel.

[0013] It should be noted that the physical meaning of the frequency transfer coefficient is clarified in this implementation plan.

[0014] The frequency transfer coefficient is not a simple empirical scalar, but a comprehensive representation of the differences in response between two frequency band receiving channels at the overall equipment level. Specifically, it includes the following physical differences: The differences in antenna feed efficiency response at different frequencies; the differences in gain frequency response of RF front-ends (including LNAs, mixers, filters, etc.); and the differences in receiver noise figure at different frequencies.

[0015] These differences together determine the inherent proportional relationship between the amplitudes of the output signals of the two channels under the same input brightness temperature.

[0016] In this invention, the inter-frequency transfer coefficient is defined as the ratio of power gains, rather than the ratio of complex gains. This is because the radiometer measures the power of the signal (i.e., brightness temperature), not phase information. Using the power gain ratio for calibration aligns with the scalar measurement nature of the radiometer, simplifies subsequent fixed-point arithmetic operations, and avoids the computational overhead and memory consumption associated with complex number operations. The pre-calibration process for this transfer coefficient is completed at the system-wide level. Specifically, before the radiometer system leaves the factory or is launched from the platform, it is placed in a controlled environmental chamber, and multiple known brightness temperature points covering the operating range are input, while the outputs of the two frequency band receiving channels are recorded. By fitting and analyzing these data, a stable power gain ratio relationship between the two channels, i.e., the inter-frequency transfer coefficient, can be determined throughout the entire operating temperature and dynamic range. This system-wide pre-calibration fully considers the comprehensive effects of all components, including the antenna, feeder, RF front-end, and intermediate frequency processing, ensuring the accuracy and systematic nature of the transfer coefficient. During system operation, the present invention makes only minor corrections to the pre-calibration coefficient through periodic calibration in step S4 to compensate for long-term slow device aging or environmental shifts, thereby ensuring the calibration accuracy of the system throughout its entire life cycle.

[0017] It should also be explained that the core mechanism of the radiometer multi-sensor data fusion preprocessing method for embedded platforms provided by this invention lies in calculating the absolute gain estimate of the first frequency band receiving channel using the retrace stroke, and directly using it as the real-time gain of the observation stroke. The effectiveness of this mechanism relies on a key premise: within a single scan cycle (no more than 100ms), the gain drift of the radiometer system can be ignored. This solution will elaborate on the rationality of this premise from both the physical mechanism and engineering implementation perspectives.

[0018] The radio frequency (RF) front-end of a radiometer, particularly the low-noise amplifier (LNA), exhibits an inherent temperature coefficient of gain that varies with temperature. The typical LNA gain temperature coefficient is approximately 0.01 dB / °C. In practical engineering platforms (such as spaceborne or airborne systems), thermal control systems ensure that structural temperature changes are slow, with system thermal time constants typically on the order of tens to hundreds of seconds. This means that within a 100 ms scan period, the temperature change of the RF front-end components is less than 0.1°C. Based on this, the temperature-induced gain drift is less than 0.001 dB. This value is far smaller than the radiometric resolution NEDT of the first-band receiving channel (typically corresponding to a gain change on the order of 0.3 K), therefore, within a single scan period, the effect of temperature changes on the gain is completely masked by the system's intrinsic noise and can be safely ignored.

[0019] The phase noise of the local oscillator source can cause uncertainty in gain measurement during the integration time. The key design feature of this invention is that the first and second frequency band receiving channels share the same local oscillator source. This design offers two significant benefits: first, the phase noise of the two channels is highly correlated, and they can be canceled out by differential processing in subsequent data fusion; second, the energy accumulation of phase noise within the 100ms integration time is limited. Sharing the local oscillator source means that any instantaneous gain fluctuations caused by phase noise are presented in the same way in both channels. Therefore, when this method obtains the real-time gain of the second channel by multiplying the absolute gain of the first channel by the pre-calibrated inter-frequency transfer coefficient in step S3, the shared phase noise contribution of the two channels is effectively canceled out through the inverse operation of the multiplication, thus ensuring the stability of the transfer coefficient over a short period. Even if residual phase noise exists, its magnitude is much smaller than that of NEDT.

[0020] The power management modules of embedded platforms typically possess excellent transient response and filtering capabilities, maintaining stable supply voltage within milliseconds. Supply voltage fluctuations are strictly controlled to a negligible extent affecting gain (e.g., gain change less than 0.001dB). Therefore, power supply fluctuations do not substantially threaten the continuity assumption of gain within a single scan cycle.

[0021] This invention corrects the inter-frequency transfer coefficient by periodically extending the retrace stroke in step S4. The selection of this period needs to balance calibration accuracy and observation efficiency. This period is set to be much smaller than the coherence time of gain drift to ensure that the system always operates within the effective calibration region. Specifically, this period should be less than a fraction of the time required for temperature drift to cause gain error to accumulate to the NEDT level. At the same time, this period should not be too frequent to avoid reducing the proportion of effective observation time due to excessively extending the retrace stroke.

[0022] In a preferred embodiment of the invention, the period is set to 100 to 200 standard scan cycles (i.e., 10 to 20 seconds). This timescale is much smaller than the characteristic response time of the RF front-end component temperature change (typically tens of seconds), ensuring that the cumulative gain drift between two calibrations does not exceed the NEDT level; at the same time, the period is not too frequent, thus balancing calibration accuracy and observation efficiency.

[0023] Furthermore, the spatial resolution of the first frequency band receiving channel is lower than that of the second frequency band receiving channel, and the radiative resolution of the first frequency band receiving channel is higher than that of the second frequency band receiving channel. In some embodiments, taking a typical microwave radiometer configuration as an example, the first receiving channel is 6.9 GHz and the second receiving channel is 89 GHz. With the same antenna aperture, spatial resolution (i.e., half-power beamwidth) is proportional to wavelength; therefore, the wavelength ratio of 6.9 GHz to 89 GHz is approximately 12.9:1, and the corresponding beamwidth ratio is also approximately 12:1. For example, if the beamwidth of the 89 GHz channel is 1.5°, then the beamwidth of the 6.9 GHz channel is approximately 18°. This indicates that the spatial resolution of the first channel is significantly lower than that of the second channel.

[0024] Radiative resolution (NEDT) depends primarily on the system noise temperature, bandwidth, and integration time. Under similar system noise temperature and integration time conditions, NEDT is inversely proportional to the square root of the bandwidth. A typical bandwidth for an 89 GHz channel is approximately 500 MHz, while a typical bandwidth for a 6.9 GHz channel is approximately 50 MHz, a bandwidth ratio of approximately 10:1, and therefore a NEDT ratio of approximately 1:3.2. That is, the NEDT of the first channel (e.g., 0.3 K) is better than that of the second channel (e.g., 1.0 K).

[0025] During the observation journey, the scanning mechanism moves at a first angular velocity, causing the first and second frequency band receiving channels to receive scene brightness and temperature radiation. During the retrace journey, the scanning mechanism moves at a second angular velocity toward the cold air, causing the first and second frequency band receiving channels to receive cold air brightness and temperature radiation. The second angular velocity is less than the first angular velocity.

[0026] It should be noted that a typical scanning mechanism includes an antenna reflector, waveguide components, and counterweights, with an overall moment of inertia typically on the order of 0.01 kg·m² to 0.05 kg·m². In this embodiment, a permanent magnet synchronous motor driven by field-oriented control (FOC) is employed. This control method can achieve high-precision torque and speed closed-loop control. During speed switching, the motor driver outputs corresponding acceleration / deceleration currents according to a preset speed curve. Due to the limited system inertia, the angular acceleration generated by the motor at rated current is sufficient to complete a large speed change within tens of milliseconds without causing current overshoot or loss of synchronization.

[0027] In this embodiment, instead of a step-like speed switching, an S-curve or trapezoidal speed planning method is used to smoothly transition the speed change process. Specifically, before the end of the observation stroke, the control system gradually reduces the speed from the first angular velocity according to a preset deceleration curve; after completing the retrace pointing, it then increases the speed back to the first angular velocity according to the acceleration curve to enter the next observation stroke. This planning method ensures that the motor current is always within the rated range, while avoiding positioning overshoot or mechanism vibration caused by inertial impact. For a typical scenario with a first angular velocity of 60° / s and a second angular velocity of 5° / s, the transition process under the S-curve planning is smooth and controllable.

[0028] During the retrace stroke, the scanning mechanism needs to point towards the cold air at a relatively low second angular velocity to obtain a sufficiently long integration time for absolute gain estimation. However, low-speed movement places higher demands on the positioning accuracy of the servo system. Taking the first frequency band receiving channel (e.g., 6.9 GHz) as an example, its half-power beamwidth (HPBW) is approximately 30°. This means that as long as the scanning mechanism points the beam center within ±15° of the zenith direction, the cold air brightness temperature received by the antenna is approximately equal to the cosmic background radiation (2.7 K), and the deviation is negligible. Therefore, in this embodiment, when the second angular velocity is designed to be 5° / s, the beam center moves only 0.1° within the 20ms integration time, which is far less than 1 / 150 of the beamwidth, and its impact on the accuracy of cold air measurement is negligible. This design, which relaxes the pointing accuracy requirements by increasing the beamwidth, makes low-speed retrace a feasible engineering solution.

[0029] Furthermore, such as Figure 2 As shown, after S4, the method also includes: S5, perform the first inversion process based on the brightness temperature observation value of the first frequency band receiving channel and the real-time gain of the first frequency band receiving channel to obtain the first inversion result; based on the first inversion result and the gain estimation uncertainty of the first frequency band receiving channel, divide the observation scene into complex region and uniform region; S6. For complex regions, the second inversion process is performed based on the brightness temperature observation value of the second frequency band receiving channel and the real-time gain of the second frequency band receiving channel to obtain the second inversion result. In some embodiments, the first inversion process uses a lookup table method to invert the large-scale surface parameter distribution based on low-resolution brightness temperature observations and real-time gain from the first frequency band receiving channel. The second inversion process uses a lookup table method to obtain preliminary high-resolution inversion results based on high-resolution brightness temperature observations and real-time gain from the second frequency band receiving channel. Then, an iterative correction process is used to ensure consistency with the first inversion results, ultimately outputting the corrected high-resolution results. Both inversion processes rely on offline pre-computed inversion kernel basis functions and sparse projection matrices to ensure the real-time performance of the embedded platform.

[0030] S7. Using the antenna pattern of the first frequency band receiving channel, the second inversion result is blurred to obtain a blurred result. The second inversion result is corrected based on the deviation between the blurred result and the first inversion result. S8, For a uniform region, correct the inter-frequency transfer coefficient based on the deviation between the observed and predicted brightness temperature values ​​of the second frequency band receiving channel; The predicted values ​​are not simple linear extrapolations, but rather forward calculations based on the radiative transfer model. Specifically, the first inversion results (such as surface temperature, soil moisture, and vegetation optical thickness) are input into the radiative transfer model to simulate and calculate the brightness temperature that the second-band receiving channel should observe under these surface parameters. The radiative transfer model is implemented on an embedded platform using a pre-computed lookup table, which discretizes the surface parameter space into a grid, pre-calculates the brightness temperature corresponding to each grid point, and stores it in Flash memory. During online runtime, the predicted values ​​are quickly obtained through table lookup and interpolation, avoiding complex online integration solutions.

[0031] S9 outputs the merged surface parameters.

[0032] It should be noted that regional delineation relies on the inversion sensitivity S(x,y), and the calculation of S(x,y) requires the first inversion result as a reference state. Specifically, the partial derivative of brightness temperature with respect to surface parameters in S(x,y) depends on the current state of the surface parameters. Only after obtaining a preliminary estimate of the surface parameters through the first inversion can the sensitivity be calculated at that state point, thereby determining whether the region is complex or homogeneous. If the first inversion is skipped, the sensitivity cannot be obtained, and regional delineation will lack a basis. Therefore, the first inversion must be performed before regional delineation can proceed.

[0033] In homogeneous regions, surface parameters change slowly in space, with a spatial scale larger than the beamwidth of the first frequency band receiving channel. In this case, the high spatial resolution information of the second frequency band receiving channel does not provide additional spatial detail (because there is no detail inherent in the region itself), and may even introduce unnecessary random errors due to its higher radiated noise. Therefore, in homogeneous regions, only the first inversion result needs to be used for cross-validation, without the need for computationally intensive high-resolution inversion. However, in complex regions, surface parameters change drastically, and the low resolution of the first channel cannot resolve details. Accurate inversion requires the high spatial resolution information of the second channel, making the second inversion indispensable.

[0034] In step S7, the second inversion result is blurred using the antenna pattern of the first channel. This is to downgrade the high-resolution result to the same spatial scale as the first inversion result for comparison. Without blurring, directly comparing a high-resolution pixel with a low-resolution pixel describes average physical quantities in different spatial ranges, resulting in a meaningless comparison. Only by smoothing the high-resolution result using the same spatial response function as the first channel can a low-resolution result with the same physical meaning as the first inversion result be obtained, thus enabling meaningful bias calculation and correction.

[0035] Furthermore, S5 includes: S51, calculate the gain estimation uncertainty based on the sample standard deviation of the absolute gain estimation of the first frequency band receiving channel. ; S52, Calculate the inversion sensitivity ; in, For the first frequency band receiving channel at the scanning sampling point Brightness temperature observations The surface parameters to be inverted are at the scanning sampling points The value, For real-time gain, This represents the radiometric resolution of the first frequency band receiving channel. Indicates the spatial location of the scan sampling point; S53, Regions exceeding a preset sensitivity threshold are marked as complex regions. Areas below a preset sensitivity threshold are marked as uniform regions; To avoid dimensional issues, this scheme actually uses a dimensionless normalized sensitivity S(x,y). During calculation, the partial derivative of brightness temperature with respect to surface parameters (reflecting the drastic change in brightness temperature with respect to parameters) is first obtained. This partial derivative is multiplied by the estimated value of the surface parameters and then divided by the observed brightness temperature value to obtain the sensitivity of the relative change in brightness temperature to the relative change in parameters. Then, this sensitivity is multiplied by the relative uncertainty of the gain estimate (i.e., the standard deviation of the gain estimate divided by the real-time gain) and then multiplied by the ratio of the observed brightness temperature value to the radiometric resolution of the first channel. After this processing, S(x,y) becomes a purely numerical index, characterizing how much of a relative error in the inversion parameters will result from the relative error of the gain at the current location.

[0036] For example, two typical thresholds are set based on the physical meaning of the inversion parameters: For soil moisture inversion, when S(x,y) is greater than 0.8, it is identified as a complex region; when S(x,y) is less than 0.3, it is identified as a homogeneous region. This is because in soil moisture inversion, the sensitivity of brightness temperature to moisture is usually between 0.2 and 1.5, and setting the threshold to 0.8 can separate sensitive regions (such as the wet-dry transition zone). For sea surface temperature inversion, when S(x,y) is greater than 0.5, it is identified as a complex region; when S(x,y) is less than 0.2, it is identified as a homogeneous region. The spatial variation of sea surface temperature is relatively gradual, so the threshold is slightly lower. These thresholds can be obtained through offline simulation and statistical analysis of measured data, and are loaded during system initialization.

[0037] S54: Allocate first computing resources to complex regions and perform second inversion processing; allocate second computing resources to uniform regions and perform cross-validation processing; the first computing resources are greater than the second computing resources.

[0038] In this embodiment, taking an embedded platform using an ARM Cortex-M7 processor (300MHz) as an example, the scan cycle is 10ms. Resource allocation is as follows: 80% of the CPU time (approximately 8ms) is allocated to the complex region. Within these 8ms, up to 5 iterative corrections are allowed (corresponding to the iterative process in claim 5), with each iteration involving sparse matrix multiplication, bias calculation, and weighted correction. For memory allocation, a sub-pixel buffer is allocated to the complex region, supporting a maximum of 100 sub-pixels. Each sub-pixel stores the inversion result, sensitivity, correction amount, etc., occupying approximately 400 bytes in total (calculated as 4-byte single-precision floating-point numbers).

[0039] Allocate 20% of the CPU time (approximately 2ms) to the uniform region. The inversion result for the uniform region only needs to store an average or representative value, thus occupying only 4 bytes of memory. During this time period, only one table lookup inversion and cross-validation calculation are performed, without iterative correction.

[0040] This differentiated allocation ensures the real-time performance and determinism of the embedded platform, while concentrating limited computing power in critical areas.

[0041] Furthermore, the second inversion process in S6 includes: S55 divides the complex region into sub-pixels that match the resolution of the second-band receiving channel; S56. Based on the brightness temperature observation value of the second frequency band receiving channel and the real-time gain of the second frequency band receiving channel, the initial second inversion result is obtained by performing a lookup table method. S57, using the antenna pattern of the first frequency band receiving channel to perform convolutional blurring on the initial second inversion result to obtain a blurred result; S58, calculate the deviation between the fuzzy result and the first inversion result; S59, based on the deviation and the local inversion sensitivity of the sub-pixel, the deviation is weighted and distributed to each sub-pixel to correct the initial second inversion result, and the corrected second inversion result is obtained. The local inversion sensitivity is calculated based on the brightness temperature observation value of the second frequency band receiving channel. For each sub-pixel, the calculation method for its local inversion sensitivity is similar to that of the global sensitivity in step S52, but it applies to the second frequency band receiving channel. Specifically, the partial derivative of the brightness temperature at that sub-pixel with respect to the surface parameters is first calculated, then multiplied by the relative uncertainty of the second channel gain estimate at that sub-pixel (i.e., the standard deviation of the gain estimate divided by the real-time gain), and finally divided by the radiometric resolution of the second channel. This local sensitivity reflects the sensitivity of the brightness temperature of the second channel to parameter changes at that sub-pixel, and also reflects the reliability of the observation at that location.

[0042] In step S59, the total bias is distributed to each sub-pixel according to the proportion of local sensitivity. This weighting is not an arbitrary mathematical trick, but a physical weighting based on observation reliability. The physical basis is that in areas of high sensitivity, the brightness temperature observations of the second channel are more sensitive to changes in surface parameters; therefore, the inversion results in these areas have higher reliability. When systematic bias exists, the correction amount should be preferentially allocated to these areas. Conversely, in areas of low sensitivity (such as brightness temperature saturation regions), the observations of the second channel are not sensitive to parameter changes, and excessive allocation of correction amounts may lead to overcorrection. Through this weighting, iterative correction can more rationally utilize the effective information of high-frequency data and avoid introducing additional errors in invalid areas.

[0043] S60, calculate the residual between the corrected second inversion result and the first inversion result, and terminate the correction when the residual is less than the residual threshold.

[0044] The residual threshold in step S60 is used to terminate the iterative correction. This implementation scheme provides two specific examples: For soil moisture retrieval, the residual threshold is set to 0.005 m³ / m³. This is because the required accuracy for soil moisture application is typically between 0.02 and 0.05 m³ / m³, and setting the threshold to one-quarter to one-tenth of the required accuracy ensures sufficient correction while avoiding excessive iteration. For land surface temperature retrieval, the residual threshold is set to 0.5 K. The required accuracy for land surface temperature application is approximately 1 to 2 K, and a residual threshold of 0.5 K is sufficient to ensure good consistency between the corrected results and the initial retrieval results.

[0045] When the residual between the corrected second inversion result and the first inversion result, after further fuzzing, is less than the aforementioned threshold, the iteration terminates, and the final result is output. This adaptive termination mechanism ensures that no further computational resources are consumed after the expected accuracy is achieved.

[0046] It should be noted that the lookup table method pre-discretizes the surface parameter space into a grid, with each grid point storing the corresponding brightness temperature value. Taking soil moisture inversion as an example, the parameter space includes soil moisture (from 0 to 0.5, step size 0.01), surface temperature (from 270K to 320K, step size 1K), and vegetation optical thickness (from 0 to 1.0, step size 0.05), etc. For the three-dimensional parameter space, the total number of grid points is approximately 50,000, with each brightness temperature value stored in 2 bytes, resulting in a total storage size of approximately 100KB, which can be accommodated in embedded Flash. Trilinear interpolation is used during online table lookup, and the discretization error is controlled within 0.2K, which is less than the radiometric resolution of the second channel (approximately 1.0K). Therefore, the interpolation error does not affect the inversion accuracy.

[0047] The initial lookup result must fall within ±20% of the true value for subsequent iterative corrections to converge. This scheme ensures this condition through reasonable grid density and interpolation methods. In typical scenarios, the initial error of the lookup method is usually less than 10%, and iterative corrections can converge to below the residual threshold in two to three iterations. If the initial error is too large (e.g., due to system failure), iterative correction will not proceed; instead, the pixel will be marked as invalid.

[0048] The sparse matrix used in steps S57 and S59 derives its sparsity from the local support properties of the basis functions. In radiative transfer inversion, the surface parameters of neighboring pixels exhibit strong spatial correlation, while pixels farther apart show weak correlation. Therefore, when projecting the inversion problem onto a locally supported basis function space (such as wavelet or radial basis functions), non-zero elements in the projection matrix only appear in pixels within their corresponding local regions, and the proportion of non-zero elements can be controlled to within 10%. This sparsity is a mathematical manifestation of physical spatial locality, rather than an artificial approximation.

[0049] Furthermore, performing cross-validation includes: S61, based on the first inversion result and the pre-stored radiative transfer model, predict the brightness temperature of the second frequency band receiving channel; S62, calculate the cross-validation residual between the observed brightness temperature of the second frequency band receiving channel and the predicted brightness temperature of the second frequency band receiving channel; S63, determine whether the cross-validation residual exceeds the radiative resolution threshold of the second frequency band receiving channel; S64, if it exceeds, then calculate the correction amount of the inter-frequency transfer coefficient based on the absolute gain estimate of the second frequency band receiving channel; S65 feeds back the corrected inter-frequency transfer coefficient to S3 to update the real-time gain of the second frequency band receiving channel in the next scanning cycle. S66, based on the updated inter-frequency transfer coefficients, re-evaluates the observation weights of the second frequency band receiving channel in the next frame, enabling the temporal division of complex and uniform regions to be adaptively adjusted.

[0050] It needs to be explained that the division of complex and uniform regions is adaptively adjusted in time. This adjustment is based on the cross-validation results of the current frame, and it actually takes effect in the region division of the next frame, with a one-frame delay.

[0051] The scanning cycle of this scheme is no more than 100ms, meaning the system updates data at a frequency of no less than 10Hz. Spatial changes in surface parameters (such as soil moisture and surface temperature) are typically slow, with timescales much longer than 100ms (for example, the transition from bare soil to water takes several seconds or even longer). Therefore, a one-frame delay (100ms) is negligible for tracking real-world surface changes and will not cause substantial misalignment between the regional divisions and the actual scene.

[0052] In extreme cases, such as when the scanning beam suddenly moves from a bare soil area to a water body area, surface parameters undergo abrupt changes. In this situation, the first frame observation may result in a large cross-validation residual due to the lack of updated transfer coefficients, potentially leading to misclassification of the region. However, the state mechanism of this scheme (claim 8) detects the residual exceeding the limit and enters an enhanced calibration state. In the second frame, the system has corrected the transfer coefficients based on the cross-validation results of the first frame and updated the real-time gain and observation weights of the second channel accordingly, thus enabling a correct response to sudden scene changes. Therefore, although there is a one-frame delay in region division, through the coordinated operation of the state machine, the system can adapt to the sudden changes within two frames without affecting overall performance.

[0053] It should be noted that, under normal system operation and without drift, the residual between the brightness temperature observation value of the second frequency band receiving channel and the forward modeling prediction value based on the first inversion result theoretically follows a normal distribution with a mean of zero. The standard deviation of this distribution mainly consists of two parts: the radiated noise of the second channel itself (i.e., its radiative resolution NEDT_H) and the forward modeling error (including the uncertainty of the radiative transfer model and the residual error of the first inversion result). Therefore, the fluctuation amplitude of the residual is on the same order of magnitude as NEDT_H, which is the statistical basis for judging whether system drift has occurred.

[0054] This scheme sets the radiometric resolution threshold in step S63 to three times the radiometric resolution of the second channel. This value has clear statistical significance: for a normal distribution, the probability of a single residual falling outside three standard deviations is approximately 0.3%. This means that if a single exceedance is caused solely by random noise, the probability of misjudgment is extremely low. However, to further avoid overreacting to accidental noise, this scheme does not immediately perform coefficient correction upon a single exceedance, but instead uses the state machine mechanism in step S68 for continuous monitoring. Specifically, only when the cross-validation residuals of multiple consecutive scan cycles (e.g., three consecutive frames) exceed this threshold is it determined to be system drift and correction triggered. The probability of exceeding the limit for N consecutive frames decreases exponentially with N; for example, the random probability of exceeding the limit for three consecutive frames drops to approximately 0.3% cubed, which is less than one in a million, almost eliminating accidental noise factors and confirming it as a genuine drift.

[0055] As a specific example, for a configuration where the first frequency band receiving channel is 6.9 GHz and the second frequency band receiving channel is 89 GHz, the typical value of the radiative resolution NEDT_H for the second channel is 1.0 K, therefore the radiative resolution threshold is set to 3.0 K. For a configuration where the first frequency band is 18 GHz and the second frequency band is 150 GHz, the typical value of the NEDT_H for the second channel is 1.2 K, and the threshold is set accordingly to 3.6 K.

[0056] Even if system drift is detected, this embodiment does not directly assign the calculated correction amount to the transfer coefficient. Instead, it uses a first-order low-pass filter for smooth updates. Specifically, the new transfer coefficient is obtained by weighted averaging the current correction amount (calculated based on the deviation between the second channel absolute gain estimate and the current product) and the old transfer coefficient, where the weighting coefficient of the correction amount is a small value between 0.1 and 0.2. This filtering mechanism ensures that the transfer coefficient can slowly and smoothly track the actual system drift (such as device aging or slow temperature changes), while effectively suppressing measurement noise that may remain in a single correction and avoiding frequent coefficient jumps. For example, when the weight is 0.1, a single correction only changes 10% of the transfer coefficient, requiring multiple consecutive corrections to complete the full adjustment, which ensures the robustness of the calibration process.

[0057] Furthermore, the method is executed on an embedded platform with memory resources less than 100KB, and the method satisfies the following conditions: The scanning period of non-uniform scanning motion is no more than 100ms, and the single-frame processing delay is a deterministic delay; wherein, the variance of the deterministic delay is less than five percent of the scanning period, and the single-frame processing delay is less than the scanning period. The calculation of absolute gain estimation, real-time gain, frequency transfer coefficient, and S52 inversion sensitivity only performs fixed-point lookup table method and fixed-point arithmetic operations, and does not perform floating-point matrix inversion or floating-point iterative optimization.

[0058] The allocation of the first and second computing resources is achieved through a preset lookup table method. The input of the lookup table method is the spatial proportion of complex regions and uniform regions in the current frame, and the output is the number of CPU cycles allocated to complex regions and uniform regions.

[0059] This embodiment designs a unified fixed-point representation format for all physical quantities involved in the calculation, ensuring that the numerical range and accuracy meet the requirements. Gain quantities (including real-time gain and absolute gain estimation) adopt the Q15 format, i.e., 16-bit integer representation with an implicit 15 decimal places, representing a range of 0 to 2, with an accuracy better than one part in thirty thousand, far exceeding the actual requirements for radiometer gain calibration. Brightness temperature values ​​also adopt the Q15 format, but with a range extended to 0 to 655K, and an accuracy of approximately 0.02K, better than the radiometer's own radiometric resolution. The frequency transfer coefficient adopts the Q15 format, ranging from 0 to 4, with an accuracy better than 0.0001, sufficient to ensure the accuracy of the second channel gain calculation. Sensitivity indicators, based on their typical value range (0 to 2), adopt a custom Q14 format, balancing range and accuracy.

[0060] For division, this embodiment employs Newton's iteration method to implement fixed-point division. Newton's iteration method can obtain a quotient with 16-bit precision in three to four iterations. Each iteration involves only multiplication and addition / subtraction, allowing for efficient execution on embedded platforms. Initial values ​​are quickly obtained through table lookup, ensuring fast convergence. For square root operations, this scheme uses a combination of table lookup and linear interpolation. A pre-calculated table of 256 uniformly distributed square roots covers the entire input range of the square root function. During online execution, the system finds the two nearest entries based on the input value and obtains a high-precision result through linear interpolation. This method has minimal computational cost, and accuracy can be further improved by appropriately increasing the number of entries (e.g., 512 entries), fully meeting the requirements of radiometer inversion.

[0061] To verify whether the accuracy of the fixed-point implementation meets the requirements, a comparative test of fixed-point and floating-point implementations was conducted during the system development phase. The test covered all typical working scenarios (including soil moisture inversion, sea surface temperature inversion, etc.) and extreme boundary conditions. The results show that the difference between the fixed-point and floating-point inversion results is less than 0.1% of the dynamic range of the inversion parameters. Taking soil moisture as an example, its dynamic range is 0 to 0.5 m³ / m³, and a 0.1% error is equivalent to 0.0005 m³ / m³, far less than the required accuracy of 0.02 m³ / m³. Therefore, the accuracy loss of fixed-point calculations has a negligible impact on the final inversion results.

[0062] This example illustrates how to ensure deterministic latency from three aspects: memory management, interrupt priority design, and worst-case execution time analysis.

[0063] All data structures in this scheme (including brightness temperature observation buffers, gain estimation storage areas, sub-pixel inversion result buffers, sparse matrix storage areas, etc.) are statically allocated at compile time, without using any dynamic memory allocation functions (such as malloc, free). Static allocation ensures the determinism of memory access addresses, avoiding the unpredictable latency and memory fragmentation problems caused by dynamic allocation. All buffer sizes are determined at compile time based on system configuration and do not require adjustment at runtime, ensuring that the execution time of memory operations is fixed.

[0064] System interrupt priorities are strictly divided according to the real-time requirements of the tasks. Timer interrupts (used for synchronizing scan cycles) are set to the highest priority to ensure accurate scanning cycle timing and prevent delays from other interrupts. DMA transfer completion interrupts (used for data acquisition) are set to the second highest priority. Other interrupts (such as serial communication and watchdog timers) are set to lower priorities. This embodiment prohibits interrupt nesting; that is, when an interrupt is being executed, subsequent interrupts are suspended until the current interrupt is completed, eliminating the uncertainty introduced by interrupt nesting. All interrupt service routines are designed for extremely simple operation, only performing necessary data transfers and flag settings, with the main computational tasks executed in the main loop, further reducing interrupt latency fluctuations.

[0065] During the system design phase, this solution conducted a worst-case execution time analysis on the critical path (i.e., the entire process from brightness temperature observation input to output fusion result). Through a combination of static code analysis and actual measurements, the most computationally intensive steps were identified as sparse matrix-vector multiplication and iterative correction loops (maximum 5 times). Tests show that on an ARM Cortex-M7 platform with a maximum clock frequency of 300MHz, the worst-case execution time is approximately 7.2ms. For example, setting the scan cycle to 10ms and reserving a 2.8ms margin for interrupt responses and other system overhead ensures that the single-frame processing latency is less than the scan cycle under all circumstances.

[0066] In one specific implementation, such as Figure 3 As shown, Figure 3 This paper demonstrates a complete implementation of the invention in both 6.9 GHz (first band) and 89 GHz (second band) configurations. In this embodiment, the 6.9 GHz channel has a lower spatial resolution (approximately 18°) but a higher radiative resolution (approximately 0.3K), while the 89 GHz channel has a higher spatial resolution (approximately 1.5°) but a lower radiative resolution (approximately 1.0K). Both channels share the same scanning mechanism, which moves at a non-uniform speed.

[0067] like Figure 3 As shown, during the retracement, the scanning mechanism points towards the cold sky (within ±15° of the zenith direction) at a relatively low second angular velocity, completing the cold sky brightness temperature acquisition within a 10ms retracement time, and calculating the absolute gain estimate for the 6.9GHz channel based on the 2.7K cosmic background radiation. During the observation journey, the scanning mechanism points towards the scene at a higher first angular velocity, simultaneously acquiring brightness temperature observations for both the 6.9GHz and 89GHz channels within 90ms.

[0068] Subsequently, the 6.9 GHz absolute gain estimate calculated from the retrace travel is directly used as the real-time gain of this channel during the observation travel. The pre-calibrated inter-frequency transfer coefficient (which reflects the frequency response differences in antenna feed efficiency, RF front-end gain, and noise figure) is multiplied by the 6.9 GHz real-time gain to obtain the real-time gain of the 89 GHz channel. To ensure long-term stability, the retrace travel is extended by 50 ms every 150 scan cycles, the absolute gain estimate of the 89 GHz channel is directly calculated, and the inter-frequency transfer coefficient is smoothly corrected based on the deviation from the current product value.

[0069] After gain calibration, the first inversion was performed using the brightness temperature observations from the 6.9 GHz channel and its high-precision real-time gain to obtain the preliminary distribution of surface parameters (such as soil moisture). Based on the first inversion results, the inversion sensitivity was calculated. Regions with a sensitivity higher than 0.8 were classified as complex regions (such as urban edges and mountains), and regions with a sensitivity lower than 0.3 were classified as uniform regions (such as water bodies and large areas of vegetation). Computational resources were allocated accordingly: 80% of the CPU time was allocated to complex regions, and 20% of the CPU time was allocated to uniform regions.

[0070] For complex regions, the region is divided into sub-pixels matching the resolution of the 89 GHz channel, and an initial high-resolution inversion result is obtained using a lookup table method. Subsequently, the high-resolution result is blurred using the antenna pattern of the 6.9 GHz channel to ensure it matches the spatial scale of the first inversion result, and the deviation between the two is calculated. Based on the local inversion sensitivity of each sub-pixel (reflecting the sensitivity of the 89 GHz brightness temperature to parameter changes at that location), the total deviation is weighted and distributed to each sub-pixel, iteratively correcting the initial inversion result until the residual is less than 0.005 m³ / m³ (for soil moisture inversion). For homogeneous regions, the brightness temperature of the 89 GHz channel is predicted using the first inversion result through a forward model, and cross-validation residuals are calculated with actual observations. If the residual exceeds 3.0K (i.e., 3 times the radiometric resolution of the 89 GHz channel) for three consecutive frames, it is considered system drift, and the inter-frequency transfer coefficient is corrected using a low-pass filter (weighting coefficient is 0.1). Finally, the corrected high-resolution inversion result for complex regions is fused with the inversion result for homogeneous regions to output a surface parameter distribution with quality indicators.

[0071] All calculations in the entire process are implemented using fixed-point arithmetic and lookup table methods. Key physical quantities are represented using Q15 format fixed-point representation. Division and square root operations are implemented using Newton's iteration method or lookup table interpolation.

[0072] Furthermore, the method is executed via a three-state state machine, where the three states include a standard observation state, an enhanced calibration state, and a degraded operation state. The three-state state machine executes as follows: S67, under standard observation conditions, performs a scan cycle of a preset standard duration, and based on the inversion sensitivity... Regularization of the standard strength; S68 monitors the deviation in S58 or the cross-validation residual in S62. When the deviation exceeds the first preset threshold or the cross-validation residual exceeds the second preset threshold, the system switches from the standard observation state to the enhanced calibration state. S69, in enhanced calibration mode, extends the integration time of the retrace stroke to more than twice the preset standard duration to enhance the accuracy of absolute gain estimation; S70, in enhanced calibration state, enhance regularization intensity to reduce the observation weight of the second frequency band receiving channel and increase the observation weight of the first frequency band receiving channel; S71, continue monitoring in enhanced calibration state. If the deviation drops below the first preset threshold or the cross-validation residual drops below the second preset threshold, return to standard observation state. If the deviation still exceeds the first preset threshold and the cross-validation residual still exceeds the second preset threshold after a preset number of scan cycles, switch to degraded operation state. S72, in degraded operation mode, pauses the output of the second frequency band receiving channel, outputs only the first inversion result and the degraded flag, and extends the retrace stroke to more than three times the preset standard duration to recalibrate the first frequency band receiving channel, while freezing the update of the inter-frequency transfer coefficient until calibration is completed.

[0073] In some embodiments, the preset standard duration of the scan travel under standard observation conditions is set to 10 ms. The corresponding second angular velocity is 5° / s (scan angle range 0.05°), which is sufficient to complete cold air pointing and basic integration. For a configuration with a first frequency band of 6.9 GHz and a second frequency band of 89 GHz, this duration meets the signal-to-noise ratio requirements for gain estimation.

[0074] This example uses a method that preserves the integrity of the observation path by extending the scan period. The specific implementation is as follows: In enhanced calibration mode, the scan cycle is extended from the standard 100ms to 140ms (i.e., the retrace distance is extended from 10ms to 50ms, while the observation distance remains unchanged at 90ms). In degraded operation mode, the scan cycle is further extended to 130ms (the retrace distance is extended to 40ms, while the observation distance remains unchanged at 90ms), or adjusted according to actual needs.

[0075] Maintaining the integrity of the observation journey ensures the continuity of scene observations and data consistency, avoiding insufficient spatial sampling due to compressed observation journeys. While extending the scan cycle may temporarily reduce the frame rate, this is acceptable given that the system is already handling abnormal situations during enhanced calibration and degraded operation. Furthermore, this approach avoids the servo control complexity caused by speed changes during retracement and prevents frame synchronization disruptions caused by inserting extra retrace cycles.

[0076] When using this scheme, the servo control needs to adapt to changes in the scanning cycle. Specifically, the control system dynamically adjusts the timer interrupt cycle based on the current state machine state: 100ms in standard state, 140ms in enhanced calibration state, and 130ms in degraded operation state. The motion trajectory planning (S-curve) of the scanning mechanism is recalculated according to the new cycle to ensure stable scanning at the first angular velocity during the observation stroke and cold-space staring at the second angular velocity (5° / s) during the extended retrace stroke. The calculation of the speed switching point is recalibrated based on the new cycle to ensure a smooth transition of motion.

[0077] For the combination of 6.9 GHz and 89 GHz, the second channel radiometric resolution is 1.0K. Therefore, the first preset threshold (the bias threshold triggering enhancement calibration) is set to 3.0K; the second preset threshold (the cross-validation residual threshold triggering degradation) is set to 4.0K (appropriately relaxed after considering model error). For the combination of 18 GHz and 150 GHz, the second channel radiometric resolution is 1.2K. Therefore, the first preset threshold is set to 3.6K; the second preset threshold is set to 4.8K.

[0078] In addition, the stability guarantee of this scheme is explained from three aspects: hysteresis design, counter mechanism, and recovery path control.

[0079] This embodiment employs a hysteresis design for the state transition thresholds. Specifically, the trigger threshold for transitioning from the standard state to the enhanced calibration state is higher than the recovery threshold for returning from the enhanced calibration state to the standard state. Taking deviation monitoring as an example, the deviation threshold (first preset threshold) for entering the enhanced calibration state from the standard state is set to three times the radiometric resolution of the second channel; while the recovery threshold for returning from the enhanced calibration state to the standard state is set to twice the radiometric resolution of the second channel. This hysteresis design means that after entering the enhanced calibration state, the system needs to improve its performance to a level better than the trigger condition before returning, avoiding oscillations around the threshold boundaries. For example, when the deviation fluctuates between 2.8 and 3.2 times the radiometric resolution, the system will not frequently switch states but will remain stable in the already entered state.

[0080] All state transitions in this embodiment employ a multi-frame exceedance judgment mechanism. Specifically, a state transition is triggered only if the deviation or cross-validation residual exceeds the trigger threshold for M consecutive frames (M being 3). A single-frame exceedance is only recorded as a warning and does not immediately change the state. For example, if a frame experiences a deviation that briefly exceeds 3 times the radiometric resolution due to transient noise, but returns to normal in the next frame, the system remains in the standard observation state and will not mistakenly enter the enhancement calibration state. The probability of exceeding the limit for 3 consecutive frames is extremely low (approximately one in a million, as mentioned in the aforementioned statistics), thus almost ensuring that the trigger for the transition is a real, continuous system drift rather than random noise.

[0081] To prevent system recovery with defects, this embodiment sets strict recovery conditions for returning to standard observation status from degraded operation. First, the system must complete the recalibration of the first frequency band receiving channel. Recalibration requires three consecutive extended retrace cycles, with the integration time of each extended retrace cycle extended to more than three times the standard duration, ensuring high signal-to-noise ratio cold-air measurements. Second, after recalibration, the system must continuously monitor the deviation and cross-validation residual for N frames (N=5) until both are below the recovery threshold before returning to standard status from degraded state. This dual verification mechanism ensures that the system's calibration status is fully restored before returning to normal mode, avoiding repeated degrades due to incomplete calibration.

[0082] Furthermore, the table lookup method in S56 includes: S73 pre-computes and stores the sparse approximation of the inversion kernel basis function and projection matrix. The non-zero elements of the sparse approximation account for less than 10%, and are stored in the Flash memory of the embedded platform. S74, when executed online, projects the brightness temperature observations onto the basis function space through sparse matrix-vector multiplication to obtain the coefficient vector. The computational complexity of sparse matrix-vector multiplication is less than 15% of that of dense matrix multiplication. S75, preliminary inversion results are obtained based on the linear combination of coefficient vector basis functions; S76, apply hard truncation physical constraints to the preliminary inversion result, truncate surface parameter values ​​less than zero to zero, truncate surface parameter values ​​greater than saturation values ​​to saturation values, and mark the truncated sub-pixels with saturation flags. The saturation flags are used to reduce the weight of the sub-pixel in subsequent fusion.

[0083] It's important to note that the physical basis of sparsity lies in the spatial locality of microwave radiation transmission: the sparsity of the projection matrix stems from the inherent spatial locality during radiation transmission. In passive microwave remote sensing, the brightness temperature observation at a given pixel location is primarily influenced by the surface parameters of that pixel and its neighboring pixels. Specifically, the main lobe and near sidelobes of the antenna pattern contribute the vast majority of the received energy, while the far sidelobes and back lobes contribute very little. This means that when projecting brightness temperature observations onto the basis functions of the surface parameter space, each observation is strongly correlated only with a few spatially adjacent basis functions, and its correlation with distant basis functions is negligible. Therefore, the non-zero elements in the projection matrix are mainly concentrated in a range of approximately 3×3 to 5×5 pixels near the main diagonal, naturally exhibiting high sparsity. This sparsity is an inherent reflection of physical phenomena, not an artificially imposed approximation.

[0084] In this embodiment, the projection matrix is ​​sparsified during the offline pre-computation stage. Specifically, a strategy is adopted to retain the K elements with the largest absolute values ​​in each row, where K is between 10 and 15. This method ensures that the element with the largest energy contribution in each row is fully retained, while elements with small contributions (below a certain threshold) are set to zero. By adjusting the value of K, the energy retention rate can be controlled. In this scheme, the value of K is chosen such that the energy retention rate of the sparse matrix is ​​greater than 99%, meaning that the total energy of the discarded elements accounts for less than 1%, and its impact on the inversion accuracy is negligible.

[0085] It should be explained that in the blurring process of step S7, this scheme dynamically adjusts the weights of truncated sub-pixels using a saturation flag. Specifically, for sub-pixels marked with a saturation flag, their weights when participating in spatial convolution blurring are reduced to one-tenth of their normal value (e.g., multiplied by 0.1). The purpose of this design is that when the inversion result of a sub-pixel is truncated, it indicates that the brightness temperature observation at that location exceeds the effective range of the model (e.g., brightness temperature saturation), and the reliability of the inversion result of that sub-pixel is low. Reducing its weight can prevent this outlier from contaminating the inversion results of neighboring pixels through blurring, thereby ensuring the stability of the overall inversion result.

[0086] In some embodiments, for sub-regions divided into complex areas, this scheme counts the number of sub-pixels marked with a saturation flag. If the proportion of saturated sub-pixels in a certain region exceeds 50% (i.e., more than half of the sub-pixels are truncated), the entire region is marked as "unreliable." This quality flag is included in the final output for reference by upper-layer applications (such as data assimilation systems or users). This quality flag mechanism ensures the transparency of the output results, enabling users to identify and appropriately use inversion data with different levels of reliability.

[0087] Furthermore, the first frequency band receiving channel is a 6.9 GHz channel, and the second frequency band receiving channel is an 89 GHz channel; or the first frequency band receiving channel is an 18 GHz channel, and the second frequency band receiving channel is a 150 GHz channel; The first frequency band receiving channel and the second frequency band receiving channel share the same local oscillator source and perform non-uniform scanning motion through the same scanning mechanism; the cold air brightness temperature radiation is the cosmic background radiation, and its brightness temperature value is 2.7K.

[0088] The above only lists two preferred frequency band combinations: 6.9GHz and 89GHz, and 18GHz and 150GHz.

[0089] The frequency band combinations applicable to this invention must meet two basic conditions. First, the two frequency bands must share the same local oscillator source or have coherent local oscillator sources to ensure the correlation of phase noise. This is the physical basis for the stability of the inter-frequency transfer coefficient in claim 1; that is, sharing a local oscillator makes the phase noise of the two channels highly correlated, which can be mutually canceled through gain product calibration. Second, the antenna feeds of the two frequency bands must be physically co-located or installed close together to ensure field-of-view registration. If there is a large deviation in the pointing of the antennas of the two frequency bands, they will observe different spatial regions, and subsequent pixel-level fusion will lose its physical meaning.

[0090] This approach is not suitable for all low-frequency and high-frequency combinations. For example, the combination of 6.9 GHz and 150 GHz is not recommended. This is because the 150 GHz band experiences strong atmospheric attenuation (especially in the presence of water vapor), and its observed brightness temperature is significantly affected by the atmosphere, while the 6.9 GHz band is essentially transparent to the atmosphere. This results in significant differences in observation conditions between the two bands, with the frequency transfer coefficient varying drastically under different atmospheric conditions, making stable tracking through periodic calibration difficult. Furthermore, the brightness temperature of the 150 GHz band at the top of the atmosphere (outside the zenith) deviates significantly from the 2.7 K cosmic background, rendering the cold-space calibration reference inaccurate and affecting the accuracy of the absolute gain estimation for the first channel.

[0091] For the recommended 89 GHz and 150 GHz combination, this scheme confirms the effectiveness of its cold-air calibration. In the millimeter-wave band, atmospheric absorption towards the zenith under clear conditions is mainly concentrated in the troposphere, with an optical thickness of less than 0.1. Therefore, when observing from the zenith, the atmospheric top brightness temperature in the 89 GHz and 150 GHz bands remains close to the cosmic background radiation (2.7 K), and the cold-air calibration benchmark is reliable. This is also an important reason for selecting these frequency band combinations in this scheme.

[0092] Low-frequency antennas have large apertures (e.g., approximately 0.6m for 6.9GHz), while high-frequency antennas have small apertures (e.g., only approximately 0.04m for 89GHz). This solution integrates the two using an offset parabolic design: the high-frequency feed is placed in the sidelobe or edge region of the low-frequency antenna, allowing both bands to share the same reflector. This design avoids the space and weight overhead of configuring a separate antenna for the high-frequency band. The main beam of the low-frequency antenna is formed by parabolic reflection, while the high-frequency feed utilizes the edge portion of the parabolic surface to form a relatively independent beam, with the two beams approximately coinciding in the far field. By optimizing the offset position and feed design, the beam center pointing deviation between the two bands can be controlled within an acceptable range.

[0093] In one specific implementation, Figure 4Another embodiment of the present invention is shown in 18GHz (first frequency band) and 150GHz (second frequency band) configurations, highlighting the operation mechanism of the three-state machine, servo adaptation with extended retrace travel, fixed-point arithmetic guarantee, and hardware integration scheme.

[0094] like Figure 4 As shown, the system operates under a control framework based on a three-state machine, which includes a standard observation state, an enhanced calibration state, and a degraded operation state. The transitions between states are achieved through a hysteresis threshold and a continuous counting mechanism.

[0095] Under standard observation conditions, a 100ms scan cycle (10ms retrace travel, 90ms observation travel) is used, and data fusion is performed according to a preset standard regularization intensity. When the cross-validation residuals of three consecutive frames exceed 3.6K (i.e., 3 times the radiometric resolution of the 150GHz channel) or the deviation exceeds 4.8K, the system switches to enhanced calibration mode. In enhanced calibration mode, the scan cycle is extended to 140ms (keeping the observation travel unchanged at 90ms, and the retrace travel extended to 50ms), while the regularization intensity is enhanced, and the weight of the observation values ​​in the second channel is reduced. When the residuals of three consecutive frames drop below 2.4K and the deviation drops below 3.6K, the system returns to standard observation mode; if the indicators continue to deteriorate for multiple consecutive frames under enhanced calibration mode (residuals exceeding 4.8K and deviation exceeding 6.0K), it switches to degraded operation mode. In degraded operation mode, the scan cycle is adjusted to 130ms (40ms retrace travel), the output of the second channel is paused, only the first inversion result and the degraded flag are output, and the update of the inter-frequency transfer coefficients is frozen. Once the first channel has completed recalibration (three consecutive extended retrace cycles, with the integration time reaching more than three times the standard duration) and the residual for five consecutive frames is below 2.4K, the system returns to the standard observation state.

[0096] In terms of servo control, the extended retrace stroke is achieved by maintaining the integrity of the observation stroke and dynamically adjusting the scan cycle. The control system switches the timer interrupt cycle (100ms, 140ms, or 130ms) according to the current state machine state and replans the S-curve speed trajectory to ensure that the motor current does not exceed the limit and that there is no overshoot in positioning.

[0097] All calculations were performed using fixed-point arithmetic, with key physical quantities using Q15 or custom formats. Division was implemented using Newton's iteration method (3 to 4 iterations), and square roots were achieved using a lookup table with 256 entries plus linear interpolation. Comparison between fixed-point and floating-point calculations showed that the difference in inversion results was less than 0.1% of the parameter dynamic range. To ensure deterministic latency, the system employed static memory allocation (disabling malloc), fixed interrupt priorities, and prohibited nesting. Worst-case execution time analysis ensured that the critical path (sparse matrix-vector multiplication and iterative correction) was completed within 7.2ms, with a 2.8ms margin. The variance of single-frame processing latency was less than 5% of the scan cycle.

[0098] In the implementation of the inversion algorithm, the projection matrix adopts a sparse approximation, retaining the 10 elements with the largest absolute values ​​in each row, with non-zero elements accounting for less than 8%. For the inversion results obtained by the lookup table method, a hard truncation physical constraint is applied (setting values ​​less than zero to zero and values ​​greater than the saturation value to the saturation value), and a saturation flag is marked. In subsequent blurring processing, the weight of saturated sub-pixels is reduced to one-tenth of the normal value; if the proportion of saturated sub-pixels in a certain region exceeds 50%, the entire region is marked with an "unreliable" quality flag.

[0099] In terms of hardware integration, the two frequency band channels share the same local oscillator source to ensure phase noise correlation. The antenna adopts an offset parabolic design, with the 150GHz feed placed in the sidelobe region of the 18GHz antenna. Through a combination of mechanical adjustment and electrical calibration, the pointing deviation of the beam centers of the two frequency bands is controlled within 0.15° (less than one-tenth of the half-power beamwidth of the 150GHz channel), ensuring field-of-view registration accuracy.

[0100] Through the above design, this embodiment achieves efficient and robust inversion of surface parameters (such as atmospheric water vapor, cloud parameters, etc.) under 18GHz and 150GHz configurations, fully verifying the versatility and engineering feasibility of the method of the present invention.

[0101] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described radiometer multi-sensor data fusion preprocessing method for embedded platforms is also provided. This electronic device may be... Figure 5 The terminal device or server shown. This embodiment uses this electronic device as an example of a server. Figure 5 As shown, the electronic device includes a memory 402, a processor 404, and a transmission device 406. The memory 402 stores a computer program, and the processor 404 is configured to execute the steps of any of the above method embodiments through the computer program.

[0102] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0103] Optionally, the transmission device 406 is used to receive or send data via a network. Specific examples of the network described above may include wired and wireless networks. In one example, the transmission device 406 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 406 is a Radio Frequency (RF) module used to communicate with the Internet wirelessly. Furthermore, the electronic device also includes a display 408 and a connection bus 410, which connects the various module components within the electronic device.

[0104] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A preprocessing method for multi-sensor data fusion of a radiometer for embedded platforms, the radiometer including a first frequency band receiving channel and a second frequency band receiving channel, characterized in that, include: S1, control the scanning mechanism to perform non-uniform scanning motion, wherein the non-uniform scanning motion includes observation stroke and retrace stroke; S2, during the retrace stroke, the absolute gain estimate of the first frequency band receiving channel is calculated based on the cold air brightness temperature radiation; during the observation stroke, the brightness temperature observation values ​​of the first frequency band receiving channel and the second frequency band receiving channel are obtained based on the scene brightness temperature radiation. S3, during the observation process, the absolute gain estimate of the first frequency band receiving channel is used as the real-time gain of the first frequency band receiving channel, and the real-time gain of the second frequency band receiving channel is set as the product of the real-time gain of the first frequency band receiving channel and the pre-calibrated frequency transfer coefficient. S4, periodically extend the retrace stroke to calculate the absolute gain estimate of the second frequency band receiving channel, and correct the inter-frequency transfer coefficient based on the deviation between the absolute gain estimate of the second frequency band receiving channel and the product.

2. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 1, characterized in that, The spatial resolution of the first frequency band receiving channel is lower than that of the second frequency band receiving channel, and the radiative resolution of the first frequency band receiving channel is higher than that of the second frequency band receiving channel. During the observation journey, the scanning mechanism moves at a first angular velocity to enable the first frequency band receiving channel and the second frequency band receiving channel to receive scene brightness and temperature radiation. During the retrace journey, the scanning mechanism moves at a second angular velocity toward the cold air direction to enable the first frequency band receiving channel and the second frequency band receiving channel to receive cold air brightness and temperature radiation. The second angular velocity is less than the first angular velocity.

3. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 1, characterized in that, Following S4, the method further includes: S5, perform a first inversion process based on the brightness temperature observation value of the first frequency band receiving channel and the real-time gain of the first frequency band receiving channel to obtain a first inversion result; based on the first inversion result and the gain estimation uncertainty of the first frequency band receiving channel, divide the observation scene into complex regions and uniform regions. S6, For the complex region, a second inversion process is performed based on the brightness temperature observation value of the second frequency band receiving channel and the real-time gain of the second frequency band receiving channel to obtain a second inversion result; S7. Using the antenna pattern of the first frequency band receiving channel, the second inversion result is blurred to obtain a blurred result. The second inversion result is corrected based on the deviation between the blurred result and the first inversion result. S8, For the uniform region, the inter-frequency transfer coefficient is corrected based on the deviation between the observed and predicted brightness temperature values ​​of the second frequency band receiving channel; S9 outputs the merged surface parameters.

4. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 3, characterized in that, S5 include: S51, calculate the gain estimation uncertainty based on the sample standard deviation of the absolute gain estimation of the first frequency band receiving channel. With inversion sensitivity; S53, regions with inversion sensitivity higher than a preset sensitivity threshold are marked as complex regions, and regions with inversion sensitivity lower than the preset sensitivity threshold are marked as uniform regions; S54, allocate a first computing resource to the complex region and perform the second inversion process, allocate a second computing resource to the uniform region and perform cross-validation process, wherein the first computing resource is greater than the second computing resource.

5. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 4, characterized in that, The second inversion process in S6 includes: S55, the complex region is divided into sub-pixels that match the resolution of the second frequency band receiving channel; S56, Based on the brightness temperature observation value of the second frequency band receiving channel and the real-time gain of the second frequency band receiving channel, the initial second inversion result is obtained by performing a lookup table method; S57, using the antenna pattern of the first frequency band receiving channel to perform convolutional blurring on the initial second inversion result to obtain a blurred result; S58, Calculate the deviation between the fuzzy result and the first inversion result; S59, Based on the deviation and the local inversion sensitivity of the sub-pixel, the deviation is weighted and distributed to each sub-pixel to correct the initial second inversion result, and the corrected second inversion result is obtained; S60, calculate the residual between the corrected second inversion result and the first inversion result, and terminate the correction when the residual is less than the residual threshold.

6. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 5, characterized in that, The cross-validation process includes: S61, based on the first inversion result, predict the brightness temperature of the second frequency band receiving channel; S62, calculate the cross-validation residual between the observed brightness temperature of the second frequency band receiving channel and the predicted brightness temperature of the second frequency band receiving channel; S63, determine whether the cross-validation residual exceeds the radiative resolution threshold of the second frequency band receiving channel; S64, if it exceeds, then calculate the correction amount of the inter-frequency transmission coefficient based on the absolute gain estimate of the second frequency band receiving channel; S65, the corrected inter-frequency transfer coefficient is fed back to S3 to update the real-time gain of the second frequency band receiving channel in the next scanning cycle; S66, based on the updated inter-frequency transfer coefficients, re-evaluate the observation weights of the second frequency band receiving channel in the next frame, so that the division of the complex region and the uniform region is adaptively adjusted in time.

7. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 1, characterized in that, The method is executed on an embedded platform, and the memory resources of the embedded platform are less than 100KB. The method satisfies the following: The scanning period of the non-uniform scanning motion is no more than 100ms, and the single-frame processing delay is a deterministic delay; wherein, the variance of the deterministic delay is less than five percent of the scanning period, and the single-frame processing delay is less than the scanning period.

8. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 4, characterized in that, The method is executed via a three-state state machine, wherein the three states include a standard observation state, an enhanced calibration state, and a degraded operation state. The three-state state machine executes as follows: Under the standard observation conditions, a scanback process of a preset standard duration is performed, and regularization is performed based on the inversion sensitivity; In the enhanced calibration state, the integration time of the retrace stroke is extended to more than twice the preset standard duration; In the enhanced calibration state, the enhanced regularization intensity is used to reduce the observation weight of the second frequency band receiving channel and increase the observation weight of the first frequency band receiving channel. In the degraded operation state, the output of the second frequency band receiving channel is paused, only the first inversion result is output, and the retrace journey is extended to more than three times the preset standard duration to recalibrate the first frequency band receiving channel. At the same time, the update of the frequency transfer coefficient is frozen until the calibration is completed.

9. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 5, characterized in that, The table lookup method in S56 includes: The sparse approximation of the inversion kernel basis function and projection matrix is ​​pre-computed and stored. The non-zero elements of the sparse approximation account for less than 10%, and are stored in the Flash memory of the embedded platform.

10. The radiometer multi-sensor data fusion preprocessing method for embedded platforms according to claim 1, characterized in that, The first frequency band receiving channel is a 6.9 GHz channel, and the second frequency band receiving channel is an 89 GHz channel; or the first frequency band receiving channel is an 18 GHz channel, and the second frequency band receiving channel is a 150 GHz channel; The first frequency band receiving channel and the second frequency band receiving channel share the same local oscillator source and perform the non-uniform scanning motion through the same scanning mechanism.

Images