A method for detecting all-sky cloud by multi-band brightness temperature difference collaborative inversion

Abstract

The application discloses a kind of multi-band brightness temperature difference collaborative inversion all-sky cloud detection methods, belong to atmospheric remote sensing and meteorological observation field, to solve the problem that cloud and clear separation, cloud phase state and cloud height inversion are susceptible to attitude deviation and light path pollution influence under complex environment, the present application constructs normalization multi-brightness temperature difference residual by double observation site multispectral thermal infrared metrology correction and unified sky grid mapping, utilizes sensor domain and sky domain motion inconsistency to identify pollution and implement dew point adaptive heating recovery, obtains cloud probability by fusing brightness temperature difference, texture and time sequence evidence, jointly inverts cloud microphysics, phase state and cloud base cloud top height, realizes the technical effect of improving complex environment effective observation rate and long-term unattended stability.

Description

Technical Field

[0001] This invention relates to the field of atmospheric remote sensing and meteorological observation, and in particular to a method for all-sky cloud detection by multi-band brightness-temperature difference collaborative inversion. Background Technology

[0002] All-sky cloud detection acquires cloud distribution information within a hemispherical area of ​​the sky through fisheye imaging or panoramic radiometric observation. It can be used to monitor cloud cover, cloud movement, cloud phase, and the positions of cloud bases and tops. Existing methods mostly segment clouds from clear skies based on visible light grayscale, single thermal infrared brightness temperature, or threshold values ​​for a few spectral bands, and estimate cloud parameters based on this.

[0003] When the observation scene includes conditions such as haze, thin clouds, near-ground temperature inversion, changes in radiative intensity at dawn and dusk, or low signal intensity at night, the differences in radiation and texture between clouds and the clear sky background are weakened. Fixed thresholds are difficult to adapt to changes in zenith angle, water vapor content, and equipment response simultaneously. Relying solely on single observational evidence is also insufficient to reliably distinguish between water clouds, ice clouds, and mixed clouds, and it is also difficult to provide a reliable description of the uncertainty in the location of cloud bases and cloud tops.

[0004] Furthermore, device attitude deviations can alter the correspondence between the fisheye image and the actual sky direction, while dust, condensation, and frost on the lens or mirror surface can cause fixed attenuation and blurring. If optical path contamination is mistaken for cloud formations, or heating is based solely on dew point risk, false contamination reports, ineffective heating, and thermal drift can easily occur, causing the clear-sky benchmark to be updated by contaminated samples over long-term operation.

[0005] Therefore, there is a need for a multi-band brightness-temperature difference collaborative inversion method for all-sky cloud detection that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a multi-band brightness-temperature difference collaborative inversion method for all-sky cloud detection. Addressing the problems of unstable cloud-clear segmentation under complex weather, dawn / dusk, and nighttime conditions, and the propagation of observation errors along the phase and cloud height inversion chain caused by attitude shifts, dust accumulation, dew, and frost, this invention proposes a method that simultaneously acquires multispectral thermal infrared observations from at least two spatially known observation positions. After reference source metrological correction, astronomical anchor point attitude correction, and unified sky grid mapping, normalized multi-brightness-temperature difference residuals are formed. Optical path obstruction is decomposed based on the motion inconsistency between the fixed component of sensor coordinates and the advection component of sky coordinates. Heating and recovery fault transfer are controlled using dew point risk and frost evidence. Finally, brightness-temperature difference, texture, and temporal evidence are weighted and fused using pixel reliability to jointly invert cloud microphysics, phase, and cloud base / top height. This invention achieves the technical effects of reducing false alarms and ineffective heating, improving the reliability of cloud parameter inversion in complex environments, and increasing the effective observation rate for unattended operation.

[0007] This invention provides a method for all-sky cloud detection based on multi-band brightness temperature difference co-inversion, including:

[0008] S1. Acquire all-sky radiation data of at least three thermal infrared spectral bands, as well as equipment attitude and environmental observation data. Perform radiometric calibration and attitude correction on the all-sky radiation data. Map the calibrated spectral data to a unified sky grid to obtain multi-spectral brightness temperature time series data and multi-brightness temperature difference residual vectors that have undergone differential and error normalization processing based on clear sky brightness temperature reference.

[0009] S2. Extract the observation components with fixed positions from the sensor coordinate time series, extract the observation components that move with the sky flow field from the sky coordinate time series, identify optical path contamination based on the inconsistency of the motion of the two types of observation components, and control the optical path recovery by combining the dew point state determined by the environmental observation data and the contamination identification result. Perform spectral compensation on observations that are not determined to be unusable to obtain the contamination mask, pixel reliability and reliable observation feature set.

[0010] S3. Based on the pixel reliability, fuse the multi-brightness temperature difference residual, cross-spectral texture and cloud advection time series features in the reliable observation feature set to obtain the cloud probability field and cloud mask;

[0011] S4. Based on the identification results of optical path contamination, pixel reliability and attitude correction results, control the spectral compensation, cloud detection admission and clear sky brightness temperature reference update of subsequent observation cycles, and use the obtained spectral compensation parameters, cloud detection admission status and updated clear sky brightness temperature reference for the next observation cycle.

[0012] Optionally, S1 includes: converting the response of the reference source in each thermal infrared band into a radiance correction amount, and combining it with the mirror temperature correction instrument self-emission.

[0013] The initial external orientation is optimized based on the initial orientation given by the inertial attitude of the equipment and the anchor point residual formed by the measured position of the sun, moon or star in the fisheye image and the corresponding astronomical calculation position. The equipment attitude reliability is determined based on the combined uncertainty of the inertial attitude residual and the anchor point residual.

[0014] Project each spectral pixel onto a unified sky grid divided by zenith angle and azimuth angle according to the corrected exterior orientation;

[0015] Conditional statistics of clear sky brightness temperature for each spectral band are determined from historical clear sky samples matched with zenith angle, water vapor characterization, and equipment operating parameters. The multi-brightness temperature difference residual vector is generated by subtracting the corresponding clear sky brightness temperature difference from the brightness temperature difference between spectral bands of the same grid cell and dividing by the uncertainty of the combination of calibration error and clear sky fluctuation.

[0016] Optionally, S2 includes: retaining the continuous multispectral brightness temperature frames in the sensor coordinate system and mapping them to the sky coordinate system according to the corrected exterior orientation;

[0017] In the sensor coordinate system, extract the continuously decaying component whose cross-frame position change is less than a fixed threshold in the sensor domain, and use the sky flow field determined by cloud texture displacement to predict the cloud advection position in the sky coordinate system.

[0018] The attenuation component that does not move with the sky flow field and remains fixed in the sensor coordinate system is identified as a candidate component for optical path occlusion.

[0019] When the absolute value of the formation time slope of the candidate component of the optical path obstruction is not greater than the dust accumulation slope threshold and the dew point margin is greater than the dew point risk threshold, it is determined to be dust accumulation. When the dew point margin is not greater than the dew point risk threshold, the absolute value of the formation time slope is greater than the liquid adhesion slope threshold and the mirror temperature is higher than the freezing point of water, it is determined to be condensation. When the dew point margin is not greater than the dew point risk threshold, the absolute value of the formation time slope is greater than the solid adhesion slope threshold and the mirror temperature is not higher than the freezing point of water, it is determined to be frost. The rest are determined to be unclassified pollution. The dust accumulation, condensation, frost or unclassified pollution are used as the pollution category.

[0020] The occlusion probability and pixel reliability are generated based on the common attenuation degree of each spectral band, the point diffusion blur increment, and the degree of satisfaction of classification conditions. Contamination masks and unusable pixel masks are generated based on the pollution category, occlusion probability, and pixel reliability. Each threshold is determined by the preset quantile of the corresponding statistics of the unoccluded historical samples.

[0021] Furthermore, the optical path restoration includes generating a heating control state in the following manner: when the dew point margin is not greater than the dew point risk threshold, the pollution category is condensation or frost, and the shading probability is not less than the shading activation threshold, a heating pulse is output;

[0022] During the heating process, the absolute value statistics of the multi-brightness temperature difference residual of the unobstructed sky grid pixels before and after heating, the transspectral texture sharpness, and the reference source response drift were used as recovery indicators.

[0023] Heating is stopped when the absolute value statistics of the multi-brightness temperature difference residual and the cross-spectral texture clarity are both restored to their respective historical unobstructed ranges, and the reference source response drift is not greater than the drift tolerance.

[0024] When the cumulative heating time reaches the safety time limit and the recovery index still does not meet the stop condition, or the reference source response drift is greater than the drift tolerance, heating is stopped, the corresponding pixel is written into the unusable pixel mask and a maintenance alarm is generated. The occlusion start threshold, historical unoccluded interval, drift tolerance and safety time limit are determined by the equipment calibration data and unoccluded operation samples.

[0025] Furthermore, spectral compensation includes determining spectral compensation coefficients, which are determined by estimating the instrument response drift, instrument self-emission correction, and optical path transmittance attenuation based on the reference source response, mirror temperature, and contamination type, respectively, and then performing error propagation on these three factors to obtain the compensation uncertainty.

[0026] The compensation uncertainty and the occlusion probability are jointly mapped to the attenuation of pixel reliability;

[0027] Only when the compensated reference source response is within the calibration response range and the compensation uncertainty is not greater than the compensation threshold, the compensated multispectral brightness temperature is written into the reliable observation feature set; otherwise, the corresponding spectral pixel is not written into the reliable observation feature set, but is written into the unusable pixel mask and is prohibited from being used as input for cloud detection and subsequent parameter inversion.

[0028] The calibration response range and compensation access threshold are determined by the equipment calibration data and unobstructed operation samples.

[0029] Optionally, S3 includes: calculating the consistency of amplitude and direction of multi-brightness temperature difference residual vector, the difference of local texture statistics of each spectral band, the degree of overlap of cross-spectral edge positions, and the degree of conformity between continuous frame texture displacement and the sky flow field in the same pixel neighborhood of the unified sky grid, and determining the corresponding evidence uncertainty based on the spatiotemporal fluctuations of each feature and the propagation results of preceding errors.

[0030] Only evidence whose corresponding spectral segment pixels are not marked by the unusable pixel mask and whose evidence uncertainty is finite is identified as valid evidence. Evidence weights that are negatively correlated with the uncertainty of valid evidence and sum to one after normalization are used for fusion, and the pixel reliability is multiplied into the fusion result to obtain the cloud probability field.

[0031] When no valid evidence is available, the corresponding pixel is written into the unavailable pixel mask and no definitive cloud judgment is output.

[0032] For pixels whose reliability is less than the reliability threshold, no cloud determination is output. For the remaining pixels, the cloud mask is generated according to the segmentation threshold determined by the cloud probability distribution of historical labeled samples, and the cloud motion field is estimated by the displacement of the cloud probability field of adjacent frames.

[0033] Furthermore, the cloud microphysical parameters and phase inversion based on the cloud probability field include: for pixels with cloud probability not less than the inversion probability threshold and pixel reliability not less than the reliability admission threshold, only the observed radiance corresponding to the multispectral brightness temperature after spectral band compensation and written into the reliable observation feature set is used; a cost function is constructed based on the difference between the observed radiance and the radiative transfer forward modeled radiance; and the parameter combination that minimizes the cost function is solved within the physical value range of cloud effective temperature, cloud optical thickness and particle effective scale.

[0034] The confidence level of the parameters is determined based on the residual covariance of the parameter combination.

[0035] Obtain the atmospheric temperature and humidity profile corresponding to the observation time, obtain the clear sky background temperature based on the clear sky brightness temperature reference, and input the difference between the cloud effective temperature corrected by the atmospheric temperature and humidity profile and the clear sky background temperature, as well as the cloud optical thickness, into the phase discrimination boundary established by water cloud samples, ice cloud samples and mixed cloud samples.

[0036] When the parameter falls into the non-overlapping region corresponding to water cloud, ice cloud or mixed cloud, it is determined to be the corresponding phase state; when the parameter falls into the boundary overlapping region or the parameter confidence is less than the inversion confidence threshold, it is determined to be an uncertain phase state.

[0037] Furthermore, the cloud base and cloud top height inversion includes: establishing a line-of-sight intersection relationship for cloud edge points obtained from at least two observation positions within the time synchronization error tolerance, based on the baseline between the observation positions and their respective corrected exterior orientations, and eliminating edge points with reprojection errors greater than the parallax error threshold;

[0038] Based on the clustering results of connected cloud bodies, phase continuity, and intersection vertical coordinates in the cloud mask, the retained points are paired to form cloud layers. The lower boundary points of each group along the vertical direction are determined as cloud base edge points, and the cloud base vertical coordinates and disparity confidence intervals of each cloud layer are determined by the median of the corresponding intersection vertical coordinates.

[0039] When the number of reserved point pairs for any cloud layer is less than the pairing number threshold or the overlap of clustering intervals of adjacent cloud layers is greater than the hierarchical tolerance, the cloud base of that cloud layer is marked as uncertain.

[0040] Based on the cloud optical thickness, phase state, and effective particle size, determine the vertical distribution parameters of cloud water content or cloud ice content. Combine the vertical coordinate intervals in the atmospheric temperature and humidity profile that meet the corresponding phase stability conditions to estimate the cloud geometric thickness. The sum of the cloud bottom vertical coordinates and the cloud geometric thickness is determined as the cloud top vertical coordinates. Then, propagate the parallax error, radiometric inversion residual, and profile error to generate the cloud top vertical coordinate confidence interval.

[0041] The time synchronization error tolerance is determined by the time synchronization calibration result of the observation position, the parallax error threshold is determined by the baseline measurement uncertainty between the exterior orientation and the observation position, and the pairing number threshold and the stratification tolerance are determined by the historical effective edge point pairing statistics.

[0042] Furthermore, when outputting the cloud phase confidence and the confidence intervals of the cloud base vertical coordinates and cloud top vertical coordinates, the optical path occlusion compensation uncertainty is determined based on the reference source response, mirror temperature, pollution type and occlusion probability. The inertial attitude residual and anchor point residual, the calibration error, the optical path occlusion compensation uncertainty, the cloud probability fusion uncertainty obtained from the propagation of evidence uncertainty and pixel reliability, the radiative transfer inversion residual, the parallax error of at least two observation positions and the atmospheric temperature and humidity profile error are propagated along each parameter processing chain.

[0043] When the propagation uncertainty corresponding to any pixel, spectral band, or time period is greater than the corresponding output tolerance, the corresponding output parameter is marked as uncertain and its confidence level or confidence interval is retained, and the observation samples of the pixel, spectral band, or time period are prohibited from being used for clear sky brightness temperature benchmark updates.

[0044] When the proportion of unavailable pixels or the proportion of uncertain output parameters exceeds the maintenance threshold within a consecutive preset observation period, a maintenance alarm is output.

[0045] The output tolerance, preset observation period, and maintenance threshold are determined by the parameter error distribution of historical valid observation samples and the lower limit of the valid observation rate.

[0046] Optionally, S4 includes: selecting only grid cells with cloud probability less than clear sky update threshold, cell reliability not less than reliability update threshold, not marked by the unavailable cell mask, and device attitude reliability not less than attitude update threshold as clear sky brightness temperature reference update samples.

[0047] The updated samples were grouped according to zenith angle, water vapor characterization, and equipment operating parameters, and the clear sky brightness temperature condition statistics of each group were updated with time decay weight.

[0048] When the reference source response drift is greater than the drift tolerance, the device attitude reliability is less than the attitude update threshold, or the maintenance alarm is in a valid state, the baseline update of the corresponding spectral segment or the corresponding time period is frozen, wherein the clear sky update threshold, the reliability update threshold, and the attitude update threshold are determined by the misjudgment cost constraint of historical clear sky samples and calibration samples.

[0049] The beneficial effects of this invention are:

[0050] 1. By using reference source radiometric calibration, joint correction of inertial attitude and astronomical anchor points, and unified sky grid mapping, the brightness temperature of different observation positions and different thermal infrared bands has a consistent directional reference and metrological basis, reducing the impact of attitude shift and instrument response drift on brightness temperature difference and parallax inversion.

[0051] 2. By comparing the fixed components in the sensor coordinates with the cloud advection components in the sky coordinates, and combining dew point margin, transspectral attenuation, ambiguity and formation time characteristics to distinguish between ash accumulation, condensation and frost, heating is only performed when both dew point risk and evidence of dew and frost obstruction meet the conditions. Recovery is verified using brightness temperature difference, texture and reference source drift, which can reduce false alarms of pollution, invalid heating and thermal drift.

[0052] 3. By using pixel reliability weighted fusion of multi-brightness temperature difference residuals, cross-spectral texture and cloud advection time series features, and using contamination, attitude and compensation uncertainty for inversion access and clear sky benchmark update, it can provide phase state, cloud motion, cloud base and cloud top positions and confidence intervals while outputting cloud mask and cloud amount, thus enhancing the effective observation capability under complex environment and long-term operation conditions. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart of the all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to the present invention.

[0055] Figure 2 This is a flowchart of the optical path contamination detection and processing based on motion inconsistency recognition according to the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figures 1-2 A method for all-sky cloud detection based on multi-band brightness temperature difference co-inversion includes:

[0058] S1. Acquire all-sky radiation data of at least three thermal infrared spectral bands, as well as equipment attitude and environmental observation data. Perform radiometric calibration and attitude correction on the all-sky radiation data. Map the calibrated spectral data to a unified sky grid to obtain multi-spectral brightness temperature time series data and multi-brightness temperature difference residual vectors that have undergone differential and error normalization processing based on clear sky brightness temperature reference.

[0059] S2. Extract the observation components with fixed positions from the sensor coordinate time series, extract the observation components that move with the sky flow field from the sky coordinate time series, identify optical path contamination based on the inconsistency of the motion of the two types of observation components, and control the optical path recovery by combining the dew point state determined by the environmental observation data and the contamination identification result. Perform spectral compensation on observations that are not determined to be unusable to obtain the contamination mask, pixel reliability and reliable observation feature set.

[0060] S3. Based on the pixel reliability, fuse the multi-brightness temperature difference residual, cross-spectral texture and cloud advection time series features in the reliable observation feature set to obtain the cloud probability field and cloud mask;

[0061] S4. Based on the identification results of optical path contamination, pixel reliability and attitude correction results, control the spectral compensation, cloud detection admission and clear sky brightness temperature reference update of subsequent observation cycles, and use the obtained spectral compensation parameters, cloud detection admission status and updated clear sky brightness temperature reference for the next observation cycle.

[0062] In this specific embodiment, S1 includes:

[0063] This specific implementation involves simultaneous execution from observation positions A and B, located 2.4 km apart and whose coordinates were determined by geodetic surveying. Each observation position is equipped with a 1024×1024 pixel fisheye radiation imaging component with three thermal infrared spectral bands of 8.7 μm, 10.8 μm, and 12.0 μm, a cavity reference source, a mirror temperature sensor, an inertial measurement unit, temperature, humidity, and air pressure sensors, and a satellite timing module with a timing error of no more than 10 ms. The electronic equipment reads the original counts, reference source counts, mirror temperature, three-axis attitude quaternions, ambient temperature, relative humidity, air pressure, equipment operating temperature, and power supply status of each spectral band at a 30-second observation cycle. Radiometric calibration and attitude correction are performed on the all-sky radiation data, and the original observation record is composed of the observation position identifier, observation cycle number, spectral band number, pixel row and column number, and unified timestamp.

[0064] The radiation processing module first removes counts that exceed the detector's linear response range. For isolated bad spots, it fills them with the median of effective pixels in a 3×3 neighborhood of the same spectral band. For areas with continuous missing areas exceeding the neighborhood, it only writes a missing measurement mark without performing spatial filling. Then, it corrects the detector gain drift based on dark field counts and reference source counts. The bad spot table, linear response range, reference source emissivity, and mirror self-emission coefficient for the three spectral bands are stored in the calibration database along with the calibration batch number. The calibration database is regenerated after the detector, reference source, or mirror is replaced.

[0065] The formula is used to calculate the radiance of the spectral band. Perform measurement calibration. Indicates spectral segment The corrected radiance, in units of W·m⁻²·sr⁻¹·μm⁻¹, This indicates the thermal infrared spectral band index, and the values ​​are 8.7 μm, 10.8 μm, or 12.0 μm. This represents the response gain obtained from two-point calibration of the reference source, expressed in W·m⁻²·sr⁻¹·μm⁻¹ per count. This represents the current raw cell count, in units of count. This represents the dark field count in the same spectral band, with the unit being counts. This represents the radiance bias correction amount calculated from the reference source response, with units of W·m⁻²·sr⁻¹·μm⁻¹. Indicates the mirror surface in the spectral band The effective emissivity within the range is a dimensionless calibrated quantity ranging from 0 to 1. This represents the blackbody radiance as an integral of the spectral response function, with units of W·m⁻²·sr⁻¹·μm⁻¹. This indicates the mirror temperature, expressed in K, and is derived from the mirror temperature sensor. When the radiance correction exceeds the effective radiance range recorded in the calibration database, the corresponding pixel will be marked as calibration unavailable.

[0066] The brightness temperature conversion module pre-generates a radiance-brightness temperature lookup table according to the response function, center wavelength, and Planck constant of each spectral band. The table records the spectral band number, radiance node, corresponding brightness temperature, interpolation slope, valid range, and version number. During operation, piecewise linear interpolation is performed using the corrected radiance as the key, and the result is written to the brightness temperature field. If the lookup key is below the lower bound or above the upper bound, no extrapolation is performed, and an invalid brightness temperature flag is written. In this specific implementation, the lookup table covers 180 K to 340 K and the node interval is 0.1 K. Its version is bound to the aforementioned calibration batch number.

[0067] The attitude correction module uses the exterior orientation quaternion output by the inertial measurement unit as the initial value. Within the observation period in which the sun, moon, or stars can be identified, the azimuth and zenith angles at the corresponding times are generated by the astronomical ephemeris table. The centroid of the measured bright target in the fisheye image is back-projected as the unit line of sight. The weighted Gauss-Newton iteration is used to minimize the sum of squares of the angles between the measured line of sight and the astronomically calculated line of sight at each anchor point, as well as the square of the attitude increment relative to the initial inertial value. The anchor point weight is the reciprocal of the centroid positioning variance. The module ends when the angle increment of two consecutive iterations is less than 0.001° or when the iteration reaches 20 times. If the astronomical anchor point cannot be obtained, the initial inertial value is retained and the attitude state is marked as inertial only.

[0068] The attitude covariance is obtained by multiplying the inverse matrix of the optimization equation with the anchor point residual variance. The attitude correction module compares the root mean square of the standard deviations of the three rotation angles with the 0.20° attitude admission upper limit and generates the device attitude reliability according to the rule of linear decay from 1 to 0. The reliability is recorded as 1 when the root mean square is not greater than 0.02° and as 0 when it is not less than 0.20°. The intermediate interval is linearly interpolated according to the endpoints. The upper and lower limits are determined by the distribution of repeated positioning error of anchor points at different zenith angles in the calibration field. The output record simultaneously retains the inertial attitude residual, anchor point residual, attitude covariance and the number of anchor points used.

[0069] The attitude quality module calculates the combined uncertainty based on the sum of squares of the anchor point residuals and the sum of squares of the inertial attitude residuals, and corrects the residual degrees of freedom according to the number of anchor points. When the combined uncertainty is greater than 0.20°, the device attitude reliability is set to 0 and participation in clear sky reference updates is prohibited in this cycle. When the number of anchor points is less than 2, the attitude source code is written as inertial only and the upper limit of reliability is limited to 0.70.

[0070] The grid mapping module uses the corrected exterior orientation to rotate the calibration line of sight of each fisheye pixel to the local northeast celestial coordinate system, calculates the azimuth and zenith angle of the line of sight, and maps it to a unified sky grid with azimuth intervals of 0.25°, zenith angle intervals of 0.25°, and zenith angles not greater than 85°. When multiple pixels are received in the same grid cell, the brightness temperature average is calculated according to the inverse distance weight of the angle between the pixel's line of sight and the grid center. The weights are normalized to sum to 1 within the same spectral band. When fewer than 2 pixels are received or all pixels have the calibration unavailable flag, the grid cell is recorded as missing measurement. The mapping table stores the observation position identifier, spectral band number, source pixel, target grid index, normalized weights, and exterior orientation version number.

[0071] The clear sky benchmark module establishes conditional statistical records from historical samples of the most recent 90 days that have been manually verified or automatically determined to be clear sky, divided into bins based on zenith angle of 2°, precipitable water of 2 mm, and equipment operating temperature of 5 K. Each record saves the median of brightness temperature of three spectral bands, the median brightness temperature difference between spectral bands, the interquartile range, the number of samples, the effective time period, and the benchmark version. The water vapor characterization is obtained by integrating the precipitable water from the atmospheric temperature and humidity profile at the same moment. When there is no accurate record for the current combination, only weighted interpolation based on the number of samples is performed within the adjacent zenith angle and precipitable water range. If there are still no records with no less than 200 historical samples, the corresponding grid is marked as no clear sky benchmark.

[0072] For the same grid cell Spectral pairs Using formula Generate normalized multi-brightness temperature difference residuals. Represents grid cells Spectral pairs dimensionless residuals Indicates a unified sky grid index. and This indicates two different thermal infrared spectral band indices. Represents grid cells In the spectrum The brightness temperature is expressed in Kelvin. Indicates that the same grid cell is in the same spectral band The brightness temperature is expressed in Kelvin. Indicates the current condition within the spectral band. With spectral segment The median temperature difference under clear skies, in Kelvin. Indicates spectral segment The standard uncertainty of brightness temperature calibration, composed of the reference source response, mirror temperature error, and interpolation error, is expressed in K. Indicates spectral segment The standard uncertainty of the same type, in units of K, This represents the robust standard deviation obtained by dividing the interquartile range of the clear-sky brightness temperature difference by 1.349, with units in K. The three spectral band pairs 8.7 / 10.8, 10.8 / 12.0 and 8.7 / 12.0 were selected in sequence. When the denominator was less than 0.05 K, 0.05 K was used as the lower limit of the same unit to avoid division by zero.

[0073] The electronic equipment writes the brightness temperature of the three spectral bands, the three-dimensional multi-brightness temperature difference residual vector, the standard uncertainty of each component, the equipment attitude reliability, the grid missing measurement mark, the calibration batch number, the exterior orientation version number, and the clear sky reference version number of each observation period into the multi-spectral brightness temperature time series dataset. The same record is simultaneously stored in the sensor pixel coordinate buffer and the unified sky grid buffer for S2 to perform dual-coordinate motion decomposition. Grid cells that are missing any spectral band, have no clear sky reference, or have an equipment attitude reliability of 0 do not generate valid residual vectors but retain the missing measurement reason code.

[0074] In this specific embodiment, S2 includes:

[0075] S2 reads the sensor pixel coordinate buffer and unified sky grid buffer written by S1 for the most recent 20 observation periods, aligns consecutive frames with observation bit identifier, observation period number and spectral band number, and establishes sensor domain attenuation sequence and sky domain texture sequence respectively. Frames with timestamp intervals deviating from 30 s by more than 10 ms are not included in displacement estimation but are retained in the quality log. When there are fewer than 12 valid frames in the window, the pollution identification status is recorded as insufficient evidence.

[0076] When there are fewer than 12 valid frames in the window, if there is a quality record from the previous valid period with a generation time of no more than 3 observation periods, the pollution mask and unusable cell mask in it are used, the cell reliability is multiplied by 0.70, and only the compensation brightness temperature, occlusion probability and insufficient evidence reason code are retained. The cloud detection admission is reduced to output probability only, and microphysical inversion, cloud height inversion and clear sky benchmark update are prohibited. If there is no previous valid record that meets the time requirements, the affected grid is written into the temporarily unusable cell mask, the cell reliability is set to 0 and the reliable observation feature set is not written into the grid. After the window regains no less than 12 frames, the pollution evidence is re-estimated for the 20-frame window and the fallback state is overwritten with the new period quality record. The clear sky benchmark update is still prohibited in the first recovery period.

[0077] The sensor domain processing module calculates the relative transmittance of the current frame using the corrected radiance of the unobstructed reference frame as the denominator. It takes the median of the relative transmittance of the three spectral bands to form a common attenuation map. Then, it calculates the Theil-Sen temporal slope and position centroid for each pixel of the 20-frame common attenuation map. When the centroid movement of the attenuation connected region is no more than 1.5 sensor pixels within 6 consecutive frames, the common attenuation is no less than the 99th percentile of the unobstructed historical sample, and the connected area is no less than 16 pixels, it writes a fixed position candidate record. The record fields include candidate number, centroid, area, formation slope, attenuation of the three spectral bands, and number of consecutive frames.

[0078] The sky domain processing module calculates the local contrast of a 5×5 neighborhood on a 10.8 μm grid brightness temperature map and uses the multi-scale pyramid Lucas-Kanade method to track texture points with a contrast no lower than the 75th percentile of historical unoccluded samples. After the tracking points are filtered based on the condition that the forward and backward errors are no more than 0.5 grid cells, the two-dimensional advection vector of each 15°×15° sky partition is fitted with Huber loss. When there are fewer than 20 effective tracking points in a partition, the result of the distance weighting between adjacent partitions is used. When there are fewer than 20 effective tracking points in all partitions, the sky flow field is recorded as unpredictable and fixed candidates are prohibited from being directly judged as contamination.

[0079] For the region where the sensor domain position is fixed and projected onto a unified sky grid, the motion inconsistency is expressed by the formula. calculate, Represents grid cells The dimensionless inconsistency of motion, Indicates a unified sky grid index. This represents the sensor domain displacement vector obtained by mapping a fixed candidate continuous centroid, with units in grid cells. This represents the cloud texture displacement vector predicted by the sky flow field at the same time interval, with units in grid cells. The normalized scale representing the displacement of 0.25 grid cells is determined by the 95th percentile of the displacement noise of the static clear-sky sequence. The two-dimensional Euclidean norm is defined as a candidate component for optical path occlusion when the motion inconsistency is not less than 2.0 and the number of fixed candidate frames is not less than 6. The threshold corresponds to a displacement difference that is not less than twice the 95th percentile of the static clear sky displacement noise.

[0080] The environmental status module calculates the dew point temperature based on the ambient temperature and relative humidity using the WMO-recommended Magnus coefficient table embedded in the device, and then uses the formula... Obtain dew point margin, This indicates the dew point margin, and the unit is K. The value represents the mirror surface temperature in Kelvin (K) and is derived from a mirror temperature sensor. The dew point temperature is expressed in K and is calculated from the ambient temperature and relative humidity of the same period. The Magnus coefficient table saves the effective temperature range, coefficient and version number according to the water surface and ice surface conditions respectively. When the dew point margin is not greater than 1.5 K, it enters the dew point risk state. This threshold is determined by the point with the minimum misjudgment cost of samples with no condensation and samples with condensation in the equipment dew frost test.

[0081] The pollution classification module applies deterministic rules to each candidate component of optical path obstruction: when the absolute value of the formation time slope is no greater than 0.002 per cycle and the dew point margin is greater than 1.5 K, it is marked as dust accumulation; when the dew point margin is no greater than 1.5 K, the absolute value of the formation time slope is greater than 0.008 per cycle and the mirror temperature is higher than 273.15 K, it is marked as condensation; when the dew point margin is no greater than 1.5 K, the absolute value of the formation time slope is greater than 0.005 per cycle and the mirror temperature is no higher than 273.15 K, it is marked as frost; and the rest are marked as unclassified pollution. The three slope thresholds are taken as the minimum quantile of the misjudgment cost of the corresponding statistics of the unobstructed, standard dust, condensation and frost calibration sequences and saved with the calibration batch.

[0082] The occlusion probability is expressed by the formula calculate, Represents grid cells The occlusion probability is between 0 and 1. Indicates a unified sky grid index. This represents the weight of the evidence to be jointly attenuated, and in this specific implementation, it is set to 0.45. This represents a dimensionless value representing the common attenuation of the three spectral bands, linearly normalized to the 99th percentile of the unobstructed and fully obstructed calibration values ​​and truncated to 0 to 1. This represents the incremental weight of the point diffusion fuzziness and is set to 0.30. This represents the increment of the edge spread width relative to the unobstructed median, normalized to a dimensionless value between 0 and 1 after being truncated by a fully blurred calibration value. This indicates the degree to which the classification criteria are met, and is set to a weight of 0.25. It represents the minimum normalization margin of each judgment value of the current category to the corresponding threshold and is truncated to a dimensionless value between 0 and 1. The sum of the three weights is 1 and is obtained by minimizing the Brier loss on the independent contamination label set. The higher the occlusion probability, the stronger the evidence of optical path occlusion.

[0083] The quality control module first subtracts the occlusion probability from 1 to obtain the occlusion reliability, and then multiplies it by the S1 device attitude reliability and the effective calibration flag of each spectral band to form the initial pixel reliability. When the occlusion probability is not less than 0.60, the corresponding area and one grid cell outside it are written into the contamination mask. When the occlusion probability is not less than 0.85, the calibration of any spectral band is unavailable, or the sky flow field is unpredictable and the fixed attenuation lasts for more than 30 minutes, the unavailable pixel mask is written. The boundary expansion is used to cover the adjacent contamination caused by the point diffusion function, and its width is obtained by standard occlusion patch calibration.

[0084] The heating controller adopts a four-state machine of standby, heating, recovery verification and fault lockout. When three conditions are met simultaneously, such as dew point margin not greater than 1.5 K, pollution category of condensation or frost and shading probability not less than 0.75, it switches from standby state to heating state and outputs a heating pulse with a duty cycle of 40% and a period of 5 seconds. The dew point margin and shading probability are reassessed every 30 seconds. Heating is not triggered by dust accumulation or unclassified pollution. Before the maintenance personnel reset, the fault lockout state prohibits heating again.

[0085] Before heating, the recovery verification module selects grid cells that have not entered the contamination mask and have a zenith angle of less than 70° to establish a background reference set. Simultaneously, it establishes a heated candidate region using the contamination connectivity region that triggers heating and one grid cell extending outward. The background reference set is used to calculate the median absolute value of the multi-brightness temperature difference residuals, the cross-spectral edge gradient correlation coefficient, and the relative drift of the reference source response to exclude changes in the sky scene and thermal drift. The heated candidate region records the median absolute value of the residuals, gradient correlation coefficient, occlusion probability, and light path transmittance before and after heating at the same grid location. Heating stops and enters the recovery verification state only when the background reference set remains within the unoccluded historical range for three consecutive observation periods, the median absolute value of the residuals in the heated candidate region returns to the unoccluded range, the gradient correlation coefficient is not lower than the 5th percentile of the unoccluded range, the occlusion probability is less than 0.30, the light path transmittance is not less than 0.80, and the relative drift of the reference source response is not greater than 0.5%. The cumulative heating reaches 300... If s still does not meet all the above conditions or the reference source response drift exceeds 0.5%, immediately stop heating, write the unusable cell mask and generate a maintenance alarm with observation position, heated candidate area, contamination type and failure reason;

[0086] During the heating process, the controller calculates the cumulative heating time according to each 5-second pulse cycle, calculates the temperature rise margin based on the difference between the current mirror temperature and 273.15 K, and compares the current residual of the reference set with the upper limit of the historical interval to restore the state. When the temperature rise margin is greater than 8 K, the current pulse is stopped immediately. Heating is stopped when the recovery state is passed for three consecutive observation cycles and the drift of the reference source response is not greater than the drift tolerance. In this specific embodiment, the drift tolerance is 0.5%.

[0087] The pollution evidence update module calculates the decay increment based on the current common decay amount and the common decay amount of the previous period, calculates the fuzzy increment based on the current point diffusion width and the unobstructed median, and calculates the category satisfaction based on the classification condition normalization margin. When any input exceeds the calibrated valid range, the evidence is set to invalid. When all three pieces of evidence are invalid, the pollution category is written as unclassified pollution.

[0088] For observations where the unusable pixel mask was not written, the formula is used. Implement spectral compensation, Represents grid cells In the spectrum The compensated radiance is expressed in W·m⁻²·sr⁻¹·μm⁻¹. Indicates a unified sky grid index. Indicates the thermal infrared spectral index. This indicates that S1 has been adjusted for optical path attenuation after deducting the self-emission of the mirror from the nominal emissivity, and the units are the same. This represents only the residual correction amount relative to the nominal self-emission correction of S1, and is in the same unit. This residual correction amount is obtained by multiplying the difference between the estimated actual emissivity at the current mirror temperature and the nominal emissivity of the S1 calibration batch by the radiance of a blackbody at the same temperature, and does not include the nominal self-emission component that has been subtracted by S1. This indicates the mirror temperature, and the unit is K. This represents the instrument response drift relative to the center value of the calibration batch, expressed in units of the same value. This represents the optical path transmittance estimated from the current pollution category and common attenuation, and is a dimensionless value between 0 and 1. If the transmittance is not greater than 0.35, division is not performed and the pixel is set to unusable.

[0089] The compensation uncertainty is propagated from the standard uncertainty of observed radiance, mirror temperature, self-emission coefficient, response drift, and optical path transmittance according to the first-order sensitivity coefficient of the aforementioned compensation formula. All radiance terms are unified as W·m⁻²·sr⁻¹·μm⁻¹, and transmittance remains dimensionless. The quality control module divides the compensation uncertainty by the width of the calibration response interval of the same spectral band and truncates it to 0 to 1. It forms the reliability attenuation amount together with the maximum value of the occlusion probability. Only when the reference source response after compensation is in the calibration response interval, the compensation uncertainty is not greater than 2.5 times the standard uncertainty of the unoccluded calibration of the spectral band, and the final pixel reliability is not less than 0.30, is the compensation radiance, compensation brightness temperature, and compensation uncertainty written into the reliable observation feature set.

[0090] The compensation admission module calculates the compensation brightness temperature by looking up the brightness temperature conversion table according to the compensation radiance, calculates the normalization error according to the compensation uncertainty and the width of the calibration response interval, and calculates the final pixel reliability by subtracting the reliability attenuation from 1. When the reference source response after compensation exceeds the calibration interval, the candidate for that spectral segment is deleted. When the final pixel reliability is less than 0.30, an unusable branch is selected. When no less than two valid spectral segments are still retained in the same grid, the corresponding spectral segment pair is matched and the multi-brightness temperature difference residual is updated.

[0091] The reliable observation feature set output by S2 stores compensated brightness temperature, multi-brightness temperature difference residuals, local texture, sky flow field, compensation uncertainty, contamination category, occlusion probability, and pixel reliability according to observation position, observation period, grid cell, and spectral band. Spectral bands that fail the compensation admission only retain the reason code and are prohibited from entering S3. Contamination mask, unusable pixel mask, heating status, maintenance alarm, and compensation parameter version are synchronously written into the periodic quality record for S3 to perform evidence fusion and for S4 to control the next observation period.

[0092] In this specific embodiment, S3 includes:

[0093] S3 uses the reliable observation feature set of S2 as the only observation input, and establishes evidence records for each observation position, observation period and unified sky grid cell. The evidence records include the multi-brightness temperature difference residuals of three spectral band pairs, compensated brightness temperature, 7×7 neighborhood texture statistics, cross-spectral edge map, sky flow field, compensated uncertainty, pixel reliability, unusable pixel flags and cloud probability of the previous period. Any spectral band marked by an unusable pixel mask does not generate evidence items.

[0094] The brightness temperature difference module uses the Euclidean mode of the three-dimensional multi-brightness temperature difference residual vector as amplitude evidence, and maps the cosine similarity between the current residual vector and the principal direction of the historical cloud sample as directional consistency evidence. The historical principal direction is determined by the principal components of the unit residual vectors of labeled water cloud, ice cloud and mixed cloud samples according to the zenith angle and precipitable water content. The amplitude evidence is linearly normalized to 0 to 1 by the 99th percentile of the historical clear sky and the 50th percentile of the cloud sample. Values ​​exceeding the endpoints are truncated. When the directional consistency is less than 0, it is recorded as 0.

[0095] The cross-spectral texture module calculates the standard deviation of brightness temperature, the median of Sobel gradient magnitude, and the local binary mode histogram in each 7×7 grid neighborhood. The absolute differences of the statistics corresponding to 8.7 μm and 10.8 μm, and 10.8 μm and 12.0 μm are normalized to the 95th percentile of historical clear sky and truncated to 0 to 1. The mean of the three normalized values ​​is then taken as evidence of texture difference. If there are fewer than 30 effective grids in the neighborhood, this evidence is not output but written into the insufficient neighborhood flag.

[0096] The edge and time-series modules calculate the cross-union ratio (CU) within one grid cell tolerance for the Sobel edges of the three spectral bands after non-maximum suppression. The CU is used as evidence of cross-spectral edge coincidence. The cloud probability field of the previous cycle is adsorbed to the current cycle using the S2 sky flow field. The normalized cross-correlation between the adsorbed probability gradient and the current texture gradient is calculated. The cross-correlation is linearly mapped from negative 1 to 1 to 0 to 1 as evidence of temporal consistency. The evidence is invalid when the sky flow field cannot be estimated or there are fewer than 20 effective edge pixels.

[0097] The evidence construction module calculates the Euclidean mode according to the three components of the residual vector, calculates the cosine similarity between the unit residual vector and the historical principal direction, calculates the cross-spectral difference according to the texture statistics of the three spectral bands, calculates the degree of overlap according to the intersection and union of the edges, calculates the temporal correlation according to the gradient after advection and the current gradient, and adopts the corresponding calibration zero protection quantity when any denominator is 0. When the normalization result exceeds 0 to 1, it is truncated to the nearest endpoint.

[0098] The standard uncertainty of each evidence item is synthesized by the root of square of the robust standard deviation of the evidence value in the current 7×7 neighborhood, the time standard deviation of the evidence in the same grid in the last 5 observation periods, and the components obtained by sensitivity propagation of S1 calibration and attitude error and S2 compensation error. The uncertainty of probabilistic evidence is uniformly kept dimensionless and truncated to 0 to 1. If any input is invalid or the synthesis result is not a finite value, the evidence item is removed from the set of valid evidence.

[0099] For grid cells valid evidence set Using formula Determine the weights and use the formula. Obtain cloud probability, Represents grid cells The Middle The dimensionless normalized weights of the evidence items, Indicates a unified sky grid index. and Indicates the index of valid evidence. Represents grid cells The set of valid evidence indexes, Indicates the first The dimensionless standard uncertainty of the evidence. In the summation, the first... The dimensionless standard uncertainty of the evidence. This represents a dimensionless zero-protection quantity of 0.02, determined by the minimum stable variance of the validation set. This represents the probability of a cloud being between 0 and 1. This indicates the reliability of pixels 0 to 1 in the S2 output. Indicates the first The normalized cloud evidence is taken as 0 to 1, the evidence weights are summed to 1 and the smaller the uncertainty, the greater the weight;

[0100] When the set of valid evidence is empty, the fusion module writes the corresponding grid into the unusable cell mask and records the cloud discrimination state as uncertain. When the cell reliability is less than 0.45, the cloud probability value is retained but no definite cloud discrimination is output. For the remaining grids, the minimum weighted sum of the cloud missed detection cost and the clear sky false alarm cost on the historical independent annotation set is 0.58 as the segmentation threshold. When the cloud probability is not less than 0.58, the cloud mask is recorded as 1, otherwise it is recorded as 0. Isolated connected components with an area of ​​less than 4 grid cells and a continuous time of less than 2 cycles are deleted.

[0101] The fusion quality module calculates the evidence ranking according to the uncertainty of valid evidence, calculates the weighted evidence sum according to the normalized weight, calculates the cloud probability scaling value according to the pixel reliability, marks the fusion state as single evidence dominant when the maximum single weight exceeds 0.80, selects the cloud class when the cloud probability is equal to the segmentation threshold, and merges connected components when the distance between adjacent connected components is no more than 1 grid cell and the time displacement direction is consistent.

[0102] The cloud motion module performs multi-scale phase correlation on the cloud probability field of two adjacent cycles and uses the S2 sky flow field as the initial displacement. It searches for the two-dimensional displacement that minimizes the sum of squares of the weighted probability difference within each 15°×15° sky partition. The search radius is 6 grid cells. When the ratio of the optimal value to the second-best value is less than 1.15, the motion of that partition is marked as uncertain. The determined position is removed and converted into angular velocity in 30 s, and written into the cloud motion field along with the timestamp, partition boundary and confidence level.

[0103] For grid cells with a cloud probability of not less than 0.70 and a pixel reliability of not less than 0.60, the inversion module constructs a formula based on the compensated radiance observed in the three spectral bands. , Indicates parameter combination The dimensionless weighted cost, This represents a parameter vector consisting of cloud effective temperature, cloud optical thickness, and effective particle size. Indicates the index of the three thermal infrared spectral bands. Indicates spectral segment The compensated observation radiance is expressed in W·m⁻²·sr⁻¹·μm⁻¹. This indicates the radiative transfer model in parameter combinations. The same unit forward radiance is calculated below. Indicates spectral segment The standard uncertainty of radiance, which is synthesized from calibration, compensation and atmospheric profile errors, and has the same unit, indicates that the smaller the cost, the more consistent the forward modeling results are with the observations.

[0104] The radiative transfer forward modeling uses line-by-line integration pre-calculation to find the library. The library is divided into three spectral bands of radiance: phase state, cloud effective temperature (180 K to 310 K with a step size of 1 K), cloud optical thickness (0.05 to 150 logarithmic nodes), particle effective size (2 μm to 120 μm), and observed zenith angle. During operation, trilinear interpolation is first performed using two libraries for water clouds and ice clouds respectively. Then, the bounded Levenberg-Marquardt method is used to minimize the weighted cost within the above physical interval. When the parameters exceed the limit, they are projected back to the nearest boundary. If the model fails to converge after 20 iterations or the minimum cost exceeds the 99th percentile of the historical effective inversion, the parameter state is recorded as uncertain.

[0105] The inversion solver matches the forward model bins according to the current observed zenith angle, calculates trilinear interpolation according to cloud effective temperature, cloud optical thickness and particle scale, selects the damping coefficient according to the weighted cost decrease, and obtains the parameter increment according to the normal equation. When the parameter increment crosses the physical boundary, it is projected to the boundary. When the cost decrease is less than 10^-4 for two consecutive times, the iteration stops. When it fails to converge within 20 iterations, an uncertain output branch is selected.

[0106] The parameter confidence is obtained by multiplying the inverse of the normal matrix of the weighted Jacobian matrix at the optimal point with the residual variance. The square root of the covariance diagonal is used as the standard uncertainty of cloud effective temperature, cloud optical thickness and particle effective size, respectively. When the condition number of the normal matrix is ​​greater than 10^8, the determined parameters are not output. The inversion record saves the optimal parameters, cost, number of iterations, covariance, spectral band used and search library version.

[0107] The phase discrimination module reads the atmospheric temperature and humidity profile that matches the observation time and location, and obtains the clear sky background temperature from the S1 clear sky reference. After correcting the effective cloud temperature by the profile transmittance, it forms a temperature difference relative to the clear sky background. This difference, along with the cloud optical thickness, is then input into the three-class support vector discrimination boundary trained from labeled samples of water clouds, ice clouds, and mixed clouds in the same season. The training samples are divided into a 70% training set, a 15% calibration set, and a 15% validation set according to the observation date. Platt scaling converts the boundary distance into three-class probabilities. When the maximum probability is not less than 0.65 and the difference between the maximum probability and the second largest probability is not less than 0.15, the corresponding water cloud, ice cloud, or mixed cloud is output. Otherwise, an uncertain phase is output.

[0108] The cloud height inversion module extracts the edges of connected cloud bodies from the cloud mask at observation positions A and B with a time difference of no more than 10 ms. The edge grid center is converted into a spatial line of sight using the corrected exterior orientation. Candidate pairs are established according to phase label, epipolar distance no more than 0.20°, local brightness temperature difference no more than 3 K, and edge tangential angle no more than 15°. For each pair, the midpoint of the shortest line segment between the two spatial lines of sight is taken as the intersection point. If the angular error after reprojection to the two observation positions is greater than 0.25°, the pair is deleted. The above error threshold is determined by the 99th percentile of the uncertainty jointly calibrated by baseline measurement and exterior orientation.

[0109] The parallax processing module calculates the spatial line of sight according to the coordinates of the observation position and the exterior orientation, filters edge candidates according to the epipolar distance, matches similar edges according to the phase label, finds the intersection midpoint according to the shortest distance between two lines of sight, deletes mismatched points according to the reprojection angle error of two observation positions, does not perform height inversion when there is only one observation position as a candidate, and selects the uncertain cloud base branch when the number of pairs is less than 30.

[0110] The cloud layers are grouped according to the DBSCAN rule that the intersection points are retained based on cloud mask connectivity, phase continuity, and vertical coordinate spacing not exceeding 300 m. Each group retains at least 30 pairs. The median of the lowest 10% of the intersection points is used as the vertical coordinate of the cloud base. The 2.5% and 97.5% quantiles are obtained by performing 1000 resampling cycles with replacement on the retained pairs as disparity confidence intervals. If there are fewer than 30 pairs or the vertical intervals of two adjacent groups overlap by more than 200 m, the corresponding cloud base is marked as uncertain.

[0111] The cloud top inversion module reads the extinction profile shape from the vertical distribution parameter library of water content or ice content established by the joint samples of radiosonde and cloud radar in the same season based on the optimal cloud optical thickness, phase state and particle effective scale. It uses the height interval constrained profile in the atmospheric temperature and humidity profile at the observation time that meets the stable temperature conditions of water phase or ice phase. The minimum vertical thickness that reaches the inverted cloud optical thickness by the extinction integral is taken as the cloud geometric thickness. The cloud bottom vertical coordinates are added to the cloud geometric thickness to obtain the cloud top vertical coordinates. The parameter library is established by phase state, optical thickness and particle scale and is updated every quarter with the effective samples of cloud radar in the same period.

[0112] When the phase is determined to be a mixed cloud, the cloud top inversion module normalizes the water phase probability and ice phase probability output by the phase module to obtain the water-ice ratio. It then calls the vertical distribution parameter library of water content and ice content to calculate two extinction profiles and weights them by height according to the ratio to form a mixed extinction profile. The standard uncertainty of the water-ice ratio is simultaneously propagated to the cloud geometric thickness and cloud top confidence interval. When the phase is determined to be uncertain, the module performs candidate calculations for water phase, ice phase, and mixed phase when both types of probabilities are valid. If the union width of each valid candidate cloud top confidence interval is not greater than 800 m, the union is output as the envelope confidence interval and the phase state is retained as uncertain. If the union width is greater than 800 m or any necessary profile is missing, only the cloud bottom result is retained, the cloud top state is written as uncertain, the profile branch inconsistency reason code is written, and the prohibited benchmark update set is added.

[0113] The phase confidence level, cloud base vertical coordinates, and cloud top vertical coordinates are calculated using the formula. Propagation uncertainty, Indicates output quantity The standard uncertainty is consistent with the output quantity. When representing phase probability, it is a dimensionless value; when representing height, the unit is meters (m). This represents the complete processing function from the input error source to the output quantity. Indicates the first There are several input error sources, which in turn include the reference source response, mirror temperature, pollution category quantification term, occlusion probability, inertial attitude residual, anchor point residual, calibration error, compensation error, evidence error, radiative transfer residual, parallax error, and atmospheric temperature and humidity profile error. Indicates the source of input error The standard uncertainty is calculated using the input's own units. Indicates the error source index. It represents the number of error sources participating in the propagation of the current output. The partial derivative sensitivity is obtained by applying a central difference of a standard uncertainty to each input and completing the unit conversion.

[0114] The uncertainty processing module determines the difference step size according to the unit of the input quantity, calculates the central difference according to one standard uncertainty, calculates the contribution by multiplying the square of the sensitivity by the input variance, calculates the output variance by summing all contributions, calculates the 95% confidence interval according to the output standard uncertainty, adds a covariance term when the absolute value of the correlation coefficient between the input error sources is greater than 0.20, and truncates the output interval when it exceeds the physical value range while retaining the truncation flag.

[0115] When any propagation uncertainty exceeds the corresponding output tolerance, the output module retains the value and confidence interval but sets the quality flag to uncertain, and adds the observation position, period, grid, spectral segment and cause code to the prohibited benchmark update set. The cloud phase probability tolerance of 0.25, the cloud bottom confidence interval half tolerance of 500 m and the cloud top confidence interval half tolerance of 800 m are jointly determined by the historical effective sample error distribution and the lower limit of the effective observation rate of the operation. When the proportion of unusable pixels or the proportion of uncertain output exceeds 30% within 10 consecutive observation periods, a maintenance alarm is generated.

[0116] S3 ultimately outputs the cloud probability field, cloud mask, cloud motion field, cloud effective temperature, cloud optical thickness, effective particle size, cloud phase state and its confidence level, vertical coordinates and confidence intervals of the cloud base and cloud top of each cloud layer, and simultaneously writes the prohibited reference update set, output uncertainty and maintenance alarm into the periodic product record, so that S4 can determine the cloud detection access, compensation parameters and clear sky reference update status for the next observation period.

[0117] In this specific embodiment, S4 includes:

[0118] S4 reads the quality record of S2 cycle and the product record of S3 cycle at the end of each 30-second observation cycle, and establishes the control record for the next cycle. The control record includes the observation position identifier, spectral band number, compensation parameter version, optical path transmittance estimate, cloud detection access status, clear sky baseline version, update freeze flag, freeze reason, maintenance alarm and effective cycle number, and forms a continuous version chain with the hash value of the current record and the hash value of the previous cycle record.

[0119] The spectral compensation controller uses the median of the last 5 effective cycles as the parameters for the next cycle for the transmittance, self-emission correction and response drift of the optical path through the S2 access. When the parameter changes by more than 10% relative to the previous effective version, the previous version is retained and marked as pending review. When the pollution category changes from condensation or frost to unobstructed and the occlusion probability is less than 0.30 for 3 consecutive cycles, the recovery lock is released. When dust accumulation, unclassified pollution or maintenance alarms are valid, the lower limit of transmittance is not automatically relaxed.

[0120] The cloud detection access controller performs deterministic state transitions based on grid cells: when the unavailable cell mask is 1, the cell reliability is less than 0.45, or the device attitude reliability is less than 0.40, the state is "deterministic discrimination prohibited"; when the cell reliability is not less than 0.45 and less than 0.60, the state is "only output probability and prohibits microphysics and cloud hyperspectral inversion"; when the cell reliability is not less than 0.60, the device attitude reliability is not less than 0.40, and the maintenance alarm is invalid, the state is "full-process access". The state change is written to the mask cache of the next cycle and takes effect after the acquisition is completed and before the evidence fusion begins.

[0121] The status arbitration module matches the set of allowed actions according to the pollution category, selects the cloud detection access level according to the pixel reliability, restricts the cloud high inversion status according to the device attitude reliability, freezes the access of the corresponding area when the condensation or frost heating status is valid, selects the full area restricted status when the maintenance alarm is valid, and selects the more restrictive status and records the triggering conditions when different conditions give conflicting statuses.

[0122] The clear sky baseline update module selects update samples only from grid cells that have a cloud probability of less than 0.20, a pixel reliability of not less than 0.80, an unavailable pixel mask of 0, an equipment attitude reliability of not less than 0.80, and have not entered the S3 prohibited baseline update set. All five conditions must be met simultaneously. The threshold is determined by the weighted sum and minimization of the cost of writing contaminated samples to the baseline from the independent clear sky annotation set and the equipment calibration set, the cost of rejecting real clear sky samples, and the lower limit of the number of effective samples.

[0123] The sample manager deduplicates candidate update samples according to observation position, zenith angle bin, and spectral band number, sorts them by cloud probability from smallest to largest, resolves pairs of samples with the same probability by pixel reliability from largest to smallest, retains only the latest record when the conditions are met for 5 consecutive cycles in the same grid, deletes the earliest record with the lowest reliability when the number of bin samples exceeds 5000, and writes the primary key of the retained sample into the baseline update queue.

[0124] Qualified updated samples were grouped according to zenith angle (2°), precipitation potential (2 mm), equipment operating temperature (5 K), and spectral band number, and entered into the conditional statistical queue. The formula was then used to... Update the time decay mean. Indicates conditional binning and spectral segments Updated average brightness temperature of clear skies in Kelvin. This indicates a conditional binning index determined by zenith angle, water vapor characterization, and equipment operating parameters. Indicates the thermal infrared spectral band index. This represents the dimensionless time decay weight, which is set to 0.02 in this specific implementation. This represents the average brightness temperature of clear skies before the update, in Kelvin. Indicates a qualified grid cell In the spectrum The compensated brightness temperature is expressed in Kelvin. This indicates a unified sky grid index, and the update module simultaneously updates the variance and inter-band brightness temperature difference statistics with the same attenuation weight.

[0125] To ensure consistency with the S1 reading caliber, the conditional statistics queue also maintains online 25th, 50th, and 75th percentile estimators with time decay weights for the compensated brightness temperature of each spectral band and the brightness temperature difference between the three spectral band pairs. New samples update the three quantiles with the same time decay weights. The 50th percentile is written into the median field of clear sky brightness temperature and the median field of brightness temperature difference between spectral bands, respectively. The difference between the 75th percentile and the 25th percentile is written into the interquartile range field. Each field simultaneously stores the conditional bin index, spectral band number or spectral band pair, number of valid samples, calibration batch, compensation version, and statistical version. When there are fewer than 50 samples, the previous valid robust statistical field is kept read-only and the median and interquartile range are not replaced by the mean or variance.

[0126] If the reference source response drift is greater than 0.5%, the equipment attitude reliability is less than 0.80, the maintenance alarm is valid, the number of qualified samples in the current bin is less than 50, or the mean of the bin after the update under the same conditions deviates from the previous reference by more than 0.5 times the standard deviation of the historical clear sky, the reference manager freezes the update of the corresponding spectral segment or the corresponding time period and retains the previous valid version. After freezing, the reference source drift is not greater than 0.3% for 20 consecutive periods, the equipment attitude reliability is not less than 0.90, the maintenance alarm is cleared and the number of qualified samples is not less than 200. The first version after the freeze is lifted is still limited to a single period mean change of no more than 0.1 K.

[0127] The freeze controller compares the spectral segment status with the reference source response drift and the 0.5% drift tolerance, compares the time period status with the equipment attitude reliability and the 0.80 threshold, and selects the freeze range according to the valid bits of the maintenance alarm. When the freeze condition only involves one spectral segment, the remaining spectral segments are kept up-to-read updates. When the freeze condition involves attitude, all spectral segments at the same observation position are frozen. When the release condition is not met for 20 consecutive cycles, the freeze state is maintained.

[0128] At 00:00 UTC each day, the benchmark manager writes the sample size, mean, variance, median of clear sky brightness temperature for each spectral band, median brightness temperature difference for three spectral band pairs, corresponding interquartile range, input calibration batch, compensation version, and previous version number into the candidate clear sky benchmark. In the next cycle, S1 directly reads the above median and interquartile range fields according to the condition bins and statistical versions. It calculates the cloud misclassification rate of the candidate benchmark using the clear sky period manually reviewed in the most recent 7 days. When the misclassification rate is not higher than the previous version and all bins pass the 0.2 K stability check, a new version is released. Otherwise, the previous version is rolled back and the reason for failure is retained. Any rollback does not modify the content of the historical version.

[0129] The version validator calculates the bin mean difference between the candidate baseline and the previous version, calculates the cloud false positive rate based on the clear sky period during manual review, traces and calibrates batches and compensation versions based on the sample primary key, rejects release when the mean difference of any bin exceeds 0.2 K, selects rollback when the candidate false positive rate is higher than the previous version, and restores the previous read-only version and generates a version integrity alarm when the release file verification value is inconsistent.

[0130] Before the next cycle control record is released, the controller calculates the check hash according to the control record fields, selects the loading time according to the effective cycle number, matches the device cache according to the observation bit identifier, and resends the control record once if the device does not confirm the loading. If it is still not confirmed after resending, the parameters of the previous cycle are retained and clear sky benchmark updates are prohibited. After the device confirms, the loading time and version number are written to the operation log.

[0131] At the start of the next observation period, the acquisition and calibration module loads the spectral compensation parameter version according to the control record. S3 decides whether to output only the probability or perform a complete inversion based on the cloud detection access status. The S1 clear sky reference module reads the conditional statistics according to the published version. When the control record verification fails, the version file is missing, or the hash chain is interrupted, the most recently verified read-only version is loaded and the reference update for this period is prohibited. This makes pollution identification, pixel reliability, attitude correction, compensation access, and clear sky reference update form a closed loop in the continuous observation period.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0133] This invention integrates multispectral brightness temperature difference, crossspectral texture, and cloud advection time series information into a unified sky grid for collaborative processing. It also transfers the uncertainties of attitude correction, optical path occlusion decomposition, and spectral band compensation to cloud probability, phase state, and altitude inversion through pixel reliability, ensuring that the observation quality control at each stage is consistent with the final cloud parameter output.

[0134] The sensor domain-sky domain occlusion factor decomposition and dew point and spectral recovery closed loop of the present invention change the processing structure that triggers defrosting only based on dew point risk or fixed image anomalies. By combining evidence to control heating, recovery verification, fault transfer and subsequent periodic benchmark updates, it can reduce reference source drift caused by ineffective heating while suppressing pollution misjudgment, and increase the proportion of usable data in long-term unattended operation.

Claims

1. A method for all-sky cloud detection using multi-band brightness temperature difference collaborative inversion, characterized in that, include: S1. Acquire all-sky radiation data of at least three thermal infrared bands, as well as equipment attitude and environmental observation data. Perform radiometric calibration and attitude correction on the all-sky radiation data. Map the calibrated data of each band to a unified sky grid to obtain multi-spectral brightness temperature time series data and multi-brightness temperature difference residual vectors that have been differentially processed and error normalized based on clear sky brightness temperature reference. S2. Extract the fixed-position observation components from the sensor coordinate time series and the observation components that move with the sky flow field from the sky coordinate time series. Identify optical path contamination based on the inconsistency of the motion of the two types of observation components. Combine the dew point status determined by the environmental observation data and the contamination identification results to control the optical path recovery. Perform spectral compensation on observations that are not determined to be unusable to obtain the contamination mask, pixel reliability and reliable observation feature set. S3. Based on the pixel reliability fusion reliable observation feature set, multi-brightness temperature difference residuals, cross-spectral textures and cloud advection time series features are obtained to obtain cloud probability field and cloud mask; S4. Based on the identification results of optical path contamination, pixel reliability, and attitude correction results, control the spectral compensation, cloud detection admission, and clear sky brightness temperature benchmark update for subsequent observation cycles, and use the obtained spectral compensation parameters, cloud detection admission status, and updated clear sky brightness temperature benchmark for the next observation cycle.

2. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 1, characterized in that, S1 includes: The response of the reference source in each thermal infrared spectral band is converted into a radiance correction value, and combined with the mirror temperature correction of the instrument's self-emission; the initial exterior orientation given by the device's inertial attitude and the anchor point residual formed by the measured position of the sun, moon, or star in the fisheye image and the corresponding astronomically calculated position are optimized, and the device attitude reliability is determined based on the combined uncertainty of the inertial attitude residual and the anchor point residual; each spectral band pixel is projected onto a unified sky grid divided by zenith angle and azimuth angle according to the corrected exterior orientation; the conditional statistics of clear sky brightness temperature for each spectral band are determined from historical clear sky samples matched with zenith angle, water vapor characterization, and device operating parameters, and the multi-brightness temperature difference residual vector is generated by subtracting the corresponding clear sky brightness temperature difference from the brightness temperature difference between spectral bands of the same grid pixel and dividing by the combined uncertainty of calibration error and clear sky fluctuation.

3. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 2, characterized in that, S2 includes: The continuous multispectral brightness temperature frames are respectively retained in the sensor coordinate system and mapped to the sky coordinate system according to the corrected exterior orientation; In the sensor coordinate system, continuously decaying components with cross-frame position changes less than a fixed threshold in the sensor domain are extracted. In the sky coordinate system, the cloud advection position is predicted using the sky flow field determined by cloud texture displacement. Attenuating components that do not move with the sky flow field and remain fixed in the sensor coordinate system are identified as candidate components for optical path occlusion. When the absolute value of the formation time slope of the candidate component is not greater than the dust accumulation slope threshold and the dew point margin is greater than the dew point risk threshold, it is identified as dust accumulation. Dust accumulation occurs when the dew point margin is not greater than the dew point risk threshold, the absolute value of the formation time slope is greater than the liquid adhesion slope threshold, and the mirror temperature is higher than that of water. Condensation is defined as the point of freezing. Frost is defined as the point of formation when the dew point margin is not greater than the dew point risk threshold, the absolute value of the formation time slope is greater than the solid adhesion slope threshold, and the mirror temperature is not higher than the freezing point of water. All other cases are defined as unclassified contamination. The aforementioned dust accumulation, condensation, frost, or unclassified contamination are used as the contamination categories. The occlusion probability and pixel reliability are generated based on the common attenuation degree of each spectral band, the point diffusion blur increment, and the degree to which the classification conditions are met. Contamination masks and unusable pixel masks are generated based on the aforementioned contamination categories, occlusion probabilities, and pixel reliability. Each threshold is determined by the preset quantile of the corresponding statistics of the unoccluded historical samples.

4. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 3, characterized in that, The optical path recovery in step S2 includes generating a heating control state as follows: when the dew point margin is not greater than the dew point risk threshold, the pollution category is condensation or frost, and the occlusion probability is not less than the occlusion initiation threshold, a heating pulse is output; during the heating process, the absolute value statistics of the multi-brightness temperature difference residual of the unoccupied sky grid pixels before and after heating, the cross-spectral texture clarity, and the reference source response drift are used as recovery indicators; when the absolute value statistics of the multi-brightness temperature difference residual and the cross-spectral texture clarity are both restored to their respective historical unoccupied intervals, and the reference source response drift is not greater than the drift tolerance, heating is stopped; when the cumulative heating time reaches the safety time limit and the recovery indicators still do not meet the stop conditions, or the reference source response drift is greater than the drift tolerance, heating is stopped, the corresponding pixels are written into the unusable pixel mask, and a maintenance alarm is generated, wherein the occlusion initiation threshold, historical unoccupied interval, drift tolerance, and safety time limit are determined by the equipment calibration data and the unoccupied operation samples.

5. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 3, characterized in that, S3 includes: calculating the consistency of amplitude and direction of multi-brightness temperature difference residual vectors, the difference of local texture statistics of each spectral band, the degree of overlap of cross-spectral edge positions, and the degree of conformity between continuous frame texture displacement and the sky flow field within the same pixel neighborhood of a unified sky grid; determining the corresponding evidence uncertainty based on the spatiotemporal fluctuations of each feature and the propagation results of prior errors; identifying only evidence whose corresponding spectral band pixels are not marked by the unusable pixel mask and whose evidence uncertainty is finite as valid evidence; fusing evidence weights that are negatively correlated with the uncertainty of valid evidence and sum to one after normalization; and multiplying the pixel reliability into the fusion result to obtain the cloud probability field; when there is no valid evidence, writing the corresponding pixel into the unusable pixel mask and not outputting a definitive cloud judgment; not outputting a definitive cloud judgment for pixels whose pixel reliability is less than the reliability threshold; generating the cloud mask for the remaining pixels according to the segmentation threshold determined by the cloud probability distribution of historical labeled samples; and estimating the cloud motion field from the displacement of the cloud probability field of adjacent frames.

6. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 5, characterized in that, The cloud microphysical parameters and phase inversion based on the cloud probability field include: for pixels with a cloud probability not less than the inversion probability threshold and a pixel reliability not less than the reliability admission threshold, using only the observed radiance corresponding to the multispectral brightness temperature after spectral compensation and written into the reliable observation feature set, constructing a cost function based on the difference between the observed radiance and the radiative transfer forward modeled radiance, and solving for the parameter combination that minimizes the cost function within the physical value range of cloud effective temperature, cloud optical thickness, and particle effective scale; and based on the residual of the parameter combination... The covariance determines the confidence level of the parameters; the atmospheric temperature and humidity profile corresponding to the observation time is obtained, and the clear sky background temperature is obtained according to the clear sky brightness temperature benchmark. The difference between the cloud effective temperature corrected by the atmospheric temperature and humidity profile and the clear sky background temperature, as well as the cloud optical thickness, are input into the phase discrimination boundary established by water cloud samples, ice cloud samples, and mixed cloud samples. When the parameter falls into the non-overlapping region corresponding to water cloud, ice cloud, or mixed cloud, it is determined as the corresponding phase. When the parameter falls into the boundary overlap region or the parameter confidence level is less than the inversion confidence threshold, it is determined as an uncertain phase.

7. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 6, characterized in that, Cloud base and cloud top height inversion includes: establishing line-of-sight intersection relationships for cloud edge points obtained from at least two observation positions within the time synchronization error tolerance, based on the baselines between observation positions and their respective corrected exterior orientations; eliminating edge point pairs with reprojection errors greater than the parallax error threshold; grouping the retained point pairs into cloud layers according to the clustering results of connected clouds, phase continuity, and intersection vertical coordinates in the cloud mask; determining the lower boundary points of each group along the vertical direction as cloud base edge points; and determining the cloud base vertical coordinates and parallax confidence intervals of each cloud layer using the median of the corresponding intersection vertical coordinates; when the number of retained point pairs for any cloud layer is less than the pairing number threshold or the overlap of adjacent cloud layer clustering intervals is greater than the stratification tolerance, the cloud layer is removed. The cloud base is marked as uncertain; the vertical distribution parameters of cloud water content or cloud ice content are determined based on the cloud optical thickness, phase state, and effective particle size. The cloud geometric thickness is estimated by combining the vertical coordinate intervals that meet the corresponding phase stability conditions in the atmospheric temperature and humidity profile. The sum of the cloud base vertical coordinates and the cloud geometric thickness is determined as the cloud top vertical coordinates. Parallax error, radiometric inversion residuals, and profile errors are propagated to generate the cloud top vertical coordinate confidence interval. The time synchronization error tolerance is determined by the time synchronization calibration results of the observation positions. The parallax error threshold is determined by the baseline measurement uncertainty between the exterior orientation and the observation positions. The pairing number threshold and the stratification tolerance are determined by the historical effective edge point pairing statistics.

8. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 4, characterized in that, S4 includes: selecting only grid cells with a cloud probability less than the clear sky update threshold, a cell reliability not less than the reliability update threshold, not marked by the unavailable cell mask, and a device attitude reliability not less than the attitude update threshold as clear sky brightness temperature baseline update samples; grouping the update samples according to zenith angle, water vapor characterization quantity, and device operating parameters, and updating the clear sky brightness temperature condition statistics of each group with time decay weight; freezing the baseline update of the corresponding spectral segment or the corresponding time period when the reference source response drift is greater than the drift tolerance, the device attitude reliability is less than the attitude update threshold, or the maintenance alarm is in a valid state, wherein the clear sky update threshold, the reliability update threshold, and the attitude update threshold are determined by the misjudgment cost constraint of historical clear sky samples and calibration samples.

9. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 3, characterized in that, The spectral compensation in step S2 includes determining the spectral compensation coefficients, which are determined as follows: Based on the reference source response, mirror temperature, and contamination category, the instrument response drift, instrument self-emission correction, and optical path transmittance attenuation are estimated respectively; error propagation is performed on these three factors to obtain the compensation uncertainty; the compensation uncertainty and the occlusion probability are jointly mapped to the attenuation of pixel reliability; only when the compensated reference source response is within the calibration response range and the compensation uncertainty is not greater than the compensation threshold, the compensated multispectral brightness temperature is written into the reliable observation feature set; otherwise, the corresponding spectral pixel is not written into the reliable observation feature set, but rather written into the unusable pixel mask, and is prohibited from being used as input for cloud detection and subsequent parameter inversion; wherein, the calibration response range and the compensation threshold are determined by equipment calibration data and unoccupied operation samples.

10. The all-sky cloud detection method based on multi-band brightness temperature difference collaborative inversion according to claim 7, characterized in that, When outputting the cloud phase confidence level and the confidence intervals of the cloud base vertical coordinates and cloud top vertical coordinates, the optical path occlusion compensation uncertainty is determined based on the reference source response, specular temperature, pollution category, and occlusion probability. The inertial attitude residual and anchor point residual, the calibration error, the optical path occlusion compensation uncertainty, the cloud probability fusion uncertainty obtained from the propagation of evidence uncertainty and pixel reliability, the radiative transfer inversion residual, the parallax error of at least two observation positions, and the atmospheric temperature and humidity profile error are propagated along each parameter processing chain. When the propagation uncertainty corresponding to any pixel, spectral band, or time period exceeds the corresponding output tolerance, the corresponding output parameter is marked as uncertain, and its confidence level or confidence interval is retained. The observation samples of that pixel, spectral band, or time period are prohibited from being used for clear sky brightness and temperature benchmark updates. When the proportion of unusable pixels or uncertain output parameters exceeds the maintenance threshold within a consecutive preset observation period, a maintenance alarm is output. The output tolerance, preset observation period, and maintenance threshold are determined by the parameter error distribution of historical valid observation samples and the lower limit of the effective observation rate.

Images