Parameter self-consistent calibration method for rice seedling narrow-band light supplement based on spectral inverse design

Abstract

This invention discloses a self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design, comprising the following steps: Step 1: Acquire emission spectrum curves and perform dark field subtraction and wavelength axis alignment; Step 2: Integrate by band and perform index binding to obtain the seedling spatial spectral energy propagation operator; Step 3: Obtain the rice seedling absorption operator field through an improved DeepONet model; Step 4: Acquire various readings at fixed sampling intervals and perform time alignment and remove missing and outlier values; Step 5: Write various fitted curves, correction matrices, drift terms, and hysteresis terms into a correction parameter table, define the correction interface, and form a cross-modal observation self-consistent correction field; Step 6: Calculate the three-domain self-consistent error tensor and update the parameter set to obtain the final narrowband supplemental lighting parameter set. This invention constructs a three-domain self-consistent calibration chain through an improved DeepONet model to achieve convergence of narrowband supplemental lighting parameters.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design. Background Technology

[0002] In facility-based rice seedling cultivation, to meet the light quality and quantity requirements of rice seedlings, narrow-band supplemental lighting channels such as red and blue light are often used for timed and quantitative supplemental lighting control. Existing technologies typically rely on the nominal wavelength and power of the lamps, or simplified illuminance measurements, combined with empirical thresholds or fixed formulas to determine the driving current, photoperiod, and duty cycle, using illuminance or photosynthetically active radiation values ​​from a small number of points as parameter adjustment references. In this type of approach, spectral measurements are often limited to single calibrations or factory parameters, spatial distribution is roughly estimated using geometric distance attenuation, and seedling phenotypic data collection and environmental observation are often processed separately, making it difficult to establish a repeatable and traceable unified calibration chain.

[0003] However, the aforementioned existing technologies have significant drawbacks in practical seedling cultivation applications. First, narrowband channels exhibit center wavelength drift, half-width at half-maximum (HWHM) variations, and spectral tail morphology changes under different driving currents, heat dissipation temperatures, and operating durations. Relying solely on nominal parameters or single-spectrum calibration fails to reflect the dynamic changes in the actual emission spectrum, leading to inconsistencies between supplementary lighting parameters and the actual emitted spectrum. Second, differences in the arrangement of lighting fixtures, obstructions, reflective materials, and the position of the seedling tray grid in the seedling cultivation area can cause spatial superposition differences between the direct component and the primary reflection component. Using only a few points or average illuminance to replace the grid-level incident spectral flux will result in the spatial spectral energy distribution being ignored, making precise calibration difficult. Third, spectral observation, imaging observation, and environmental observation commonly suffer from error sources such as wavelength axis shift, nonlinear response, lens vignetting, drift, and hysteresis. Without a unified calibration interface and parameter table management, and with inconsistent observation apertures across different modes, the basis for parameter updates becomes unstable. Fourth, existing methods often lack a mechanism to establish an operator-based mapping between spatial incident spectral flux and rice seedling absorption characterization. This makes it difficult to utilize multi-source aligned phenotypic observations to form a computable error tensor and drive parameter set updates, resulting in supplementary lighting parameter iterations relying on manual experience and difficulty in converging to stable calibration results.

[0004] Therefore, how to provide a self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design. This invention uses a narrowband spectral fingerprint library as the real light emission benchmark, constructs a seedling raising spatial spectral energy propagation operator, combines it with an improved DeepONet model to form a rice seedling absorption operator field, aligns phenotypic observations, performs self-consistent correction on cross-modal observations, and further constructs a three-domain self-consistent error tensor to drive the parameter set update, thereby achieving stable convergence and reproducible calibration of narrowband supplemental lighting driving quantity, photoperiod, and duty cycle parameters.

[0006] The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to an embodiment of the present invention includes the following steps: Step 1: Acquire the emission spectrum curves of each narrowband supplementary light channel under multi-level driving, and perform dark field subtraction and wavelength axis alignment to obtain the narrowband spectral fingerprint library; Step 2: Based on the emission angle distribution parameters of each channel in the narrowband spectral fingerprint spectral library and the direct component and primary reflection component from each channel of each lamp, the grid point incident spectral flux expression is obtained by band integration, and index binding is performed to obtain the seedling raising spatial spectral energy propagation operator; Step 3: Input the seedling spatial spectral energy propagation operator into the improved DeepONet model, and obtain the rice seedling absorption operator field through the incident spectral energy sample construction module, tensor encoding module, physiological condition vector construction module and operator training and solidification module; Step 4: Determine fixed sampling points on the seedling tray grid and bind a grid index to each sampling point. Collect various readings at fixed sampling intervals, perform time alignment and remove missing and outlier values ​​to generate a rice seedling phenotypic observation alignment set. Step 5: Fit the wavelength axis offset curve, response nonlinearity curve, illuminance response curve and lens vignetting correction matrix, drift term and hysteresis term, and write them into the correction parameter table according to the equipment identification, define a unified correction interface, and form a cross-modal observation self-consistent correction field. Step 6: Based on the rice seedling phenotypic observation alignment set, cross-modal observation self-consistent correction field, and rice seedling absorption operator field, obtain the three-domain self-consistent error tensor, and update the parameter set to obtain the final narrowband supplementary lighting parameter set.

[0007] Optionally, step one specifically includes: Three drive current levels are set for each narrowband fill light channel, and a constant sampling duration is set for each drive current level. The channel identifier and drive current level identifier are recorded, and a drive sampling configuration table is generated. Within the sampling time corresponding to each drive current level, the heat sink temperature sequence and the channel working time sequence are collected synchronously. Temperature range identifiers are generated for the heat sink temperature sequence according to the preset temperature segmentation rules, and working time range identifiers are generated for the channel working time sequence according to the preset time segmentation rules. Use a spectral acquisition device to acquire the emission spectrum curve of the corresponding narrowband supplementary light channel at a preset fixed distance position at the light outlet of the lamp, and acquire the corresponding dark field spectrum curve. The emission spectrum curve is subjected to dark field subtraction using the dark field spectral curve to obtain the dark field subtracted spectral curve, and then wavelength axis alignment is performed on the dark field subtracted spectral curve to obtain the aligned spectral curve. For each aligned spectral curve, calculate the center wavelength, full width at half maximum (FWHM), spectral tail skew, peak radiant flux, and first-order difference peak position, and arrange them in a preset order to form a spectral fingerprint parameter vector. The temperature drift slope is calculated based on the radiator temperature sequence, the working time drift slope is calculated based on the channel working time sequence, and the temperature drift slope and the working time drift slope are incorporated into the spectral fingerprint parameter vector. The spectral fingerprint parameter vector is written into the same index table along with the channel identifier, drive current level identifier, temperature range identifier, and operating time range identifier. The aligned spectral curve is then bound one by one to the index table entries corresponding to the spectral fingerprint parameter vector to obtain the narrowband spectral fingerprint library.

[0008] Optionally, step two specifically involves: Geometric calibration of the seedling raising area is performed under a unified coordinate system to obtain the installation coordinates and orientation angle of each lamp, and a lamp pose table is generated. Divide the seedling tray area into grids according to a preset fixed grid size, write the coordinate index for each grid point and generate a seedling tray grid index table; Measure the height of each lamp from the seedling tray plane and record the boundary polygon of the obstruction in the seedling raising area. Write the height data and boundary polygon data into the obstruction geometry table. Collect the band reflectance curves of the seedling frame and surrounding materials, and map the band reflectance curves to a unified coordinate system according to the material area to generate a material reflectance map. The emission angle distribution parameters of each narrowband supplementary light channel are read from the narrowband spectral fingerprint spectrum library, and the emission angle distribution parameters are written together with the lamp pose table, seedling tray grid index table, occlusion geometry table and material reflectivity map into the optical propagation configuration file; Based on the optical propagation profile, the direct component and primary reflection component from each channel of each lamp are calculated for each grid point, and the direct component and primary reflection component are integrated by band to obtain the expression for the incident spectral flux of the grid point. By binding the expression for the incident spectral flux at the grid points with the seedling tray grid index table, the spatial spectral energy propagation operator for seedling raising is obtained.

[0009] Optionally, the improved DeepONet model is specifically as follows: The seedling raising spatial spectral energy propagation operator is input into the incident spectral energy sample construction module. By running the seedling raising spatial spectral energy propagation operator under different narrowband channel driving combinations, the incident spectral flux samples of each grid point of the seedling tray grid are obtained. Each set of incident spectral flux samples is bound to its corresponding channel drive combination identifier and grid index to form an incident spectral energy sample set containing multiple sets of grid-level incident spectral fluxes. The incident spectrum energy sample set is input into the tensor encoding module. By dividing each set of grid-level incident spectral flux samples in the incident spectrum energy sample set into multi-resolution segments according to spectral bands, a multi-resolution spectral band sequence is obtained. The multi-resolution spectral band sequence is encoded into an input tensor according to a preset tensor organization rule, while retaining the one-to-one correspondence between the grid index and the channel drive combination identifier and the input tensor, thus forming the spectral band input tensor. The spectral input tensor is input into the physiological condition vector construction module to collect and align the chlorophyll index, leaf temperature, water content surrogate quantity and seedling age stage label corresponding to the same sampling time as the incident spectral energy sample set. The aligned chlorophyll index, leaf temperature, water content proxy quantity and seedling age stage label are combined into a condition vector according to the preset splicing rules. The condition vectors are bound to the spectral input tensors according to the sample identifier and the grid index, resulting in a set of condition vectors that correspond one-to-one with each spectral input tensor. The conditional vector set is input into the operator training and solidification module, the operator training input, operator training conditions and operator training output are set, the data is normalized and the normalization coefficient is recorded, and the data fitting error and physical constraint error are calculated to update the model parameters and generate the rice seedling absorption operator field.

[0010] Optionally, the steps of setting the operator training input, operator training conditions, and operator training output, performing data normalization processing and recording the normalization coefficients, calculating the data fitting error and physical constraint error to update the model parameters, and generating a rice seedling absorption operator field are as follows: The operator training input is set as the spectral input tensor set, the operator training condition is set as the condition vector set, and the operator training output is set as the grid-level effective absorption spectral energy sequence and the key absorption band proportion sequence. Establish a training sample index table to establish a one-to-one correspondence between spectral input tensors, conditional vectors, and output sequences; For each sample in the training sample index table, data normalization is performed and the normalization coefficients are recorded. The normalization process includes: Normalize the spectral input tensor by band energy scale and by grid flux scale; normalize the condition vector by dimension grouping; normalize the output sequence by absorbed energy scale. Set the number of iteration training rounds, the number of samples per batch, and the update step size. Read the spectral input tensor, conditional vector, and output sequence from the training sample index table in batches and calculate the predicted value of the output sequence. The data fitting error is calculated based on the predicted value and the true value of the output sequence, and the physical constraint error is calculated based on the predicted value of the output sequence. The physical constraint error includes the non-negative constraint error applied to the effective absorption spectrum energy sequence, the energy upper bound constraint error applied to the effective absorption spectrum energy sequence, and the spectral smoothing constraint error applied to the key absorption band proportion sequence. The data fitting error and the physical constraint error are combined into a total error according to a preset weight; Based on the total error, the model parameters are updated and the updated model parameters are written into the parameter snapshot table. At the same time, the corresponding spectral input tensor and condition vector are read from the validation samples to calculate the validation error and write the validation error into the validation error record table. Training ends when the verification error meets the preset stopping condition. The model parameter corresponding to the minimum verification error is selected from the parameter snapshot table as the fixed parameter, and its input and output interface is defined as an operator to obtain the rice seedling absorption operator field.

[0011] Optionally, step four specifically involves: Select a preset number of fixed sampling points on the seedling tray grid corresponding to the seedling tray grid index table, assign a sampling point identifier to each fixed sampling point and bind it to the corresponding grid index, and generate a sampling point index table. According to the preset fixed sampling interval, chlorophyll index readings, leaf temperature readings, seedling height readings, leaf area readings, dry matter sampling and weighing values, and stomatal conductance proxy readings are collected for each fixed sampling point in the sampling point index table. Each reading record is written with the sampling point identifier and sampling interval identifier to form a set of original readings for multiple indicators. The raw reading set of multiple indicators is aligned with the sampling point identifier and sampling interval identifier by time, and an alignment record is generated. Missing value removal and outlier removal are performed on the alignment record. For each sampling point, within each sampling interval, the retained chlorophyll index readings, leaf temperature readings, seedling height readings, leaf area readings, dry matter sampling and weighing values, and stomatal conductance proxy readings are combined into a multi-index observation vector according to a preset dimension order. The environmental sensor readings corresponding to the current sampling interval are read and packaged with the multi-index observation vector to generate a sampling point-level phenotypic aligned sample. All sample point-level phenotypic aligned samples are indexed and stored according to the sample point identifier, sample interval identifier, and grid index to obtain the rice seedling phenotypic observation alignment set.

[0012] Optionally, step five specifically includes: Based on a preset standard light source reference spectrum, baseline modeling is performed on the dark field spectrum. The baseline modeling involves locating characteristic peaks in the standard light source reference spectrum and fitting a wavelength axis offset curve based on the characteristic peak location results. The response nonlinearity curve is fitted based on the nominal radiation distribution and the measured radiation distribution of the standard light source reference spectrum, and the wavelength axis offset curve, the response nonlinearity curve and the spectrometer identification are written into the spectrometer correction parameter entry; Under uniform light field conditions, the imaging device acquires a multi-exposure image sequence, groups the multi-exposure image sequence according to the exposure parameters, and calculates the pixel brightness statistics of each exposure group; Illumination response curves are fitted based on pixel brightness statistics and uniform light field illuminance benchmarks, and lens vignetting correction matrices are fitted based on the spatial brightness distribution of multi-exposure images; the illuminance response curves, lens vignetting correction matrices, and imaging device identifiers are then written into the imaging correction parameter entries. Under stable environmental conditions, a reference reading sequence is acquired from the environmental sensor. A time series fitting is performed on the reference reading sequence to obtain the drift term. A piecewise fitting is performed on the rising and falling edges of the reference reading sequence to obtain the hysteresis term. The drift term and the hysteresis term, along with the environmental sensor identifier, are written into the environmental correction parameter entry. The spectral correction parameter entries, imaging correction parameter entries, and environmental correction parameter entries are written into a unified correction parameter table, and a unified correction interface is defined for the correction parameter table. The unified correction interface includes wavelength axis offset correction and response nonlinearity correction for spectral observation, illuminance response correction and lens vignetting correction for imaging observation, and drift correction and hysteresis correction for environmental observation. When inputting spectral observations, imaging observations, and environmental observations, a unified calibration interface matching the equipment identifier is invoked to output the calibrated physical quantities. The calibrated physical quantities are then organized according to time slice identifiers and grid indices to form a cross-modal observation self-consistent calibration field.

[0013] Optionally, step six specifically includes: Read the multi-index observation vector corresponding to each sampling interval, as well as the corresponding grid index and sampling point identifier, from the rice seedling phenotypic observation alignment set to generate a sampling interval index list; Read the spectral observations, imaging observations, and environmental observations corresponding to the sampling interval index list from the cross-modal observation self-consistent calibration field, and call the calibration interface that matches the device identifier to perform spectral observation calibration, imaging observation calibration, and environmental observation calibration respectively, so as to obtain the calibrated spectral observations, calibrated imaging observations, and calibrated environmental observations. The candidate values ​​of the supplementary lighting parameters to be determined are organized into a channel drive candidate table according to the narrowband supplementary lighting channel identifier. The candidate values ​​of the supplementary lighting parameters to be determined include the narrowband supplementary lighting channel drive amount, optical period parameter, and pulse duty cycle parameter. The channel-driven candidate list is input into the seedling raising spatial spectral energy propagation operator to obtain the grid-level incident spectral flux corresponding to the grid index; The energy distribution difference, center wavelength difference, half-peak width difference, and spectral tail difference are calculated by combining the grid-level incident spectral flux with the corrected spectral observations, corrected imaging observations, and corrected environmental observations, respectively, to obtain the three-domain self-consistency error tensor. The parameter set is updated based on the candidate values ​​of the supplementary lighting parameters to be determined formed by the three-domain self-consistent error tensor. Repeat the calculation of the three-domain self-consistent error tensor and the update of the parameter set until the number of iterations reaches the set value or the error convergence condition is met, to obtain the final narrowband supplementary lighting parameter set.

[0014] The beneficial effects of this invention are: This method acquires the emission spectra of each narrowband supplementary lighting channel under multiple driving conditions, performs dark field subtraction and wavelength axis alignment, and establishes a narrowband spectral fingerprint library. This makes the supplementary lighting parameters no longer dependent on the nominal wavelength and nominal power, but rather bound to the true spectral morphology and its characteristics of variation with temperature and working time, reducing the impact of channel emission uncertainty on calibration results from the source. On this basis, the emission angle distribution parameters of the channel, along with the lamp pose, seedling tray grid index, occlusion geometry, and material reflectivity map, are incorporated into the optical propagation configuration. The expression of the incident spectral flux at the grid point is obtained by integrating the direct component and the first reflection component by band, and then indexed and solidified into the seedling raising spatial spectral energy propagation operator. This transforms the supplementary lighting calibration from "point-to-point averaging" to "grid-level computability," enabling a consistent description of spatial spectral energy differences under the same coordinate system.

[0015] The incident spectral energy samples output from the seedling spatial spectral energy propagation operator are divided into multi-resolution spectral bands and tensor encoded. These samples are then bound to a conditional vector composed of chlorophyll index, leaf temperature, water content surrogate, and seedling age stage labels, and input into an improved DeepONet model. Through operator training, a rice seedling absorption operator field is obtained, creating a reusable operator-based mapping between the incident spectral flux and the rice seedling absorption characterization. This provides a stable computational foundation for subsequent error construction and parameter updates. Simultaneously, chlorophyll index, leaf temperature, seedling height, leaf area, dry matter, and stomatal conductance surrogates are collected at fixed sampling points and intervals, and time alignment, missing data, and anomaly removal are performed to form a rice seedling phenotypic observation alignment set. This ensures traceable consistency of multi-index observations in terms of grid index and time slice identification. Furthermore, through simulation... The wavelength axis offset curve, response nonlinearity curve, illuminance response curve, lens vignetting correction matrix, and environmental drift and hysteresis terms are written into a unified correction parameter table. A unified correction interface is defined to form a cross-modal observation self-consistent correction field, enabling spectral observation, imaging observation, and environmental observation to complete same-aperture correction before entering error calculation, reducing artifacts introduced by inconsistencies in multi-source observations. With the support of the above links, a three-domain self-consistent error tensor is constructed based on the rice seedling phenotypic observation alignment set, the cross-modal observation self-consistent correction field, and the rice seedling absorption operator field. This error tensor drives the set update of the supplementary lighting parameter candidate set until the iteration termination condition is met, thereby achieving systematic convergence of the narrowband supplementary lighting channel driving quantity, optical period parameter, and pulse duty cycle parameter, and obtaining an executable and reproducible final narrowband supplementary lighting parameter set. Attached Figure Description

[0016] 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: Figure 1 This is an overall flowchart of the self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design proposed in this invention. Figure 2 This is a schematic diagram of the improved DeepONet model processing of the self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design proposed in this invention. Figure 3 This is a flowchart of the operator training and solidification module processing steps for the self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-3A self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design includes the following steps: Step 1: Acquire the emission spectrum curves of each narrowband supplementary light channel under multi-level driving, and perform dark field subtraction and wavelength axis alignment to obtain the narrowband spectral fingerprint library; Step 2: Based on the emission angle distribution parameters of each channel in the narrowband spectral fingerprint spectral library and the direct component and primary reflection component from each channel of each lamp, the grid point incident spectral flux expression is obtained by band integration, and index binding is performed to obtain the seedling raising spatial spectral energy propagation operator; Step 3: Input the seedling spatial spectral energy propagation operator into the improved DeepONet model, and obtain the rice seedling absorption operator field through the incident spectral energy sample construction module, tensor encoding module, physiological condition vector construction module and operator training and solidification module; Step 4: Determine fixed sampling points on the seedling tray grid and bind a grid index to each sampling point. Collect various readings at fixed sampling intervals, perform time alignment and remove missing and outlier values ​​to generate a rice seedling phenotypic observation alignment set. Step 5: Fit the wavelength axis offset curve, response nonlinearity curve, illuminance response curve and lens vignetting correction matrix, drift term and hysteresis term, and write them into the correction parameter table according to the equipment identification, define a unified correction interface, and form a cross-modal observation self-consistent correction field. Step 6: Based on the rice seedling phenotypic observation alignment set, cross-modal observation self-consistent correction field, and rice seedling absorption operator field, obtain the three-domain self-consistent error tensor, and update the parameter set to obtain the final narrowband supplementary lighting parameter set.

[0019] This step involves fingerprinting and database construction of the actual emission spectrum of the narrowband supplementary lighting channel, and introducing spectral energy propagation calculations for direct and primary reflections in the spatial dimension. This establishes a traceable correlation between the supplementary lighting parameters and the actual emission spectral shape and grid-level incident spectral energy distribution, avoiding deviations caused by relying solely on nominal parameters or a small number of point measurements. The spatial spectral energy propagation results of the seedling raising area are input into the improved DeepONet in the form of operators to obtain the rice seedling absorption operator field. This establishes a reusable mapping relationship between the incident spectral energy and the rice seedling absorption characterization, and together with the phenotypic observation alignment set and the cross-modal observation self-consistent correction field, forms a unified data foundation, reducing the interference of errors from spectroscopic equipment, imaging equipment, and environmental sensors on the calibration results. Based on this, a three-domain self-consistent error tensor is constructed to drive the iterative update of the parameter set, enabling parameters such as the narrowband supplementary lighting driving quantity, optical period, and pulse duty cycle to form a stable and convergent final set, improving the repeatability, portability, and consistency of the supplementary lighting parameter calibration.

[0020] In this embodiment, step one specifically includes: Three drive current levels are set for each narrowband supplementary light channel, and a constant sampling duration is set for each drive current level. The channel identifier and drive current level identifier are recorded, and a drive sampling configuration table is generated. The drive current level is used to cover the low-load, medium-load, and high-load operating range of the channel. The constant sampling duration is used to ensure that the spectral sampling of each level is completed within the same time window and can be compared. The drive sampling configuration table includes at least the channel identifier field, drive current level identifier field, level setting value field, sampling duration field, sampling order field, and sampling parameter field consistent with the spectral acquisition device. The sampling parameter field includes integration time, averaging times, sampling interval, and wavelength sampling resolution to ensure that the subsequently acquired data has repeatability and traceability.

[0021] Within the sampling duration corresponding to each drive current level, the heat sink temperature sequence and the channel operating time sequence are simultaneously acquired. Temperature range identifiers are generated for the heat sink temperature sequence according to preset temperature segmentation rules, and operating time range identifiers are generated for the channel operating time sequence according to preset duration segmentation rules. The heat sink temperature sequence and the channel operating time sequence are timestamped using the same time base. Synchronous acquisition ensures that the influence of temperature and duration on spectral morphology changes can be referenced by the same spectral curve. Preset temperature segmentation rules are used to map continuous temperature values ​​to discrete temperature range identifiers, and preset duration segmentation rules are used to map continuous operating time to discrete duration range identifiers. Temperature range identifiers and operating time range identifiers are used to establish operating condition dimensions in the index table and support subsequent retrieval and comparison by operating condition.

[0022] The emission spectrum curve of the corresponding narrowband supplementary light channel is acquired at a preset fixed distance position at the light outlet of the lamp using a spectral acquisition device, and the corresponding dark field spectrum curve is also acquired. The geometric relationship between the preset fixed distance position and the light outlet remains unchanged to reduce measurement geometric errors. The emission spectrum curve and the dark field spectrum curve are acquired using the same sampling parameters to maintain consistent observation aperture. The dark field spectrum curve is used to characterize the dark current and background noise of the spectral acquisition device under conditions of no effective light input, and the emission spectrum curve is used to characterize the actual output spectrum shape of the channel under the current drive current level, current heat dissipation temperature, and working time.

[0023] Dark field subtraction is performed on the emission spectrum curve using the dark field spectral curve to obtain the dark field subtracted spectral curve, and wavelength axis alignment is performed on the dark field subtracted spectral curve to obtain the aligned spectral curve. Dark field subtraction is to subtract the dark field spectral curve from the emission spectrum curve point by point according to the wavelength sampling point to eliminate dark current and background noise components. Wavelength axis alignment is to map the wavelength sampling axis of the dark field subtracted spectral curve to a unified wavelength reference axis to eliminate peak position shift caused by device wavelength drift. The unified wavelength reference axis is used to ensure that the spectral curves of different channels, different levels, and different operating conditions can be directly compared in the same wavelength coordinate system. The aligned spectral curve serves as the only input spectrum for subsequent spectral parameter calculations.

[0024] For each aligned spectral curve, the center wavelength, half-maximum width (HWHM), spectral tail skew, peak radiant flux, and first-order difference peak position are calculated and arranged in a preset order to form a spectral fingerprint parameter vector. The center wavelength is determined by the position of the main peak of the aligned spectral curve, the HWHM is determined by the wavelength difference between the left and right sides corresponding to the HWHM of the main peak, the spectral tail skew is determined by the asymmetry of the energy distribution of the spectral tails on both sides of the main peak, the peak radiant flux is determined by the peak amplitude of the main peak or the integral amplitude of the preset band, and the first-order difference peak position is determined by the extreme value position of the wavelength dimension first-order difference sequence of the aligned spectral curve. The preset order is used to fix the field arrangement rules of the parameter vector and ensure that the meaning of the fields is consistent when writing to the index table, so that the spectral fingerprint parameter vector can be directly used for subsequent retrieval, comparison, and modeling.

[0025] The temperature drift slope is calculated based on the radiator temperature sequence, and the working time drift slope is calculated based on the channel working time sequence. The temperature drift slope and the working time drift slope are then incorporated into the spectral fingerprint parameter vector. The temperature drift slope is used to characterize the rate of change of the center wavelength or key peak position of the aligned spectral curve with the radiator temperature, and the working time drift slope is used to characterize the rate of change of the center wavelength or key peak position of the aligned spectral curve with the channel working time. The drift slope is calculated based on the temperature sequence and the time sequence within the same sampling time as the aligned spectral curve. The temperature drift slope field and the working time drift slope field are appended to the preset position of the spectral fingerprint parameter vector to form an extended spectral fingerprint parameter vector that simultaneously contains spectral features and working condition drift features.

[0026] The spectral fingerprint parameter vector, along with the channel identifier, drive current level identifier, temperature range identifier, and operating time range identifier, are written into the same index table. The aligned spectral curves are then bound one-to-one with the corresponding index table entries for the spectral fingerprint parameter vectors, resulting in a narrowband spectral fingerprint library. The index table contains at least the following fields: channel identifier, drive current level identifier, temperature range identifier, operating time range identifier, spectral fingerprint parameter vector, and aligned spectral curve storage pointer. These fields are bound one-to-one so that the aligned spectral curves and their corresponding spectral fingerprint parameter vectors share the same index key and establish a bidirectional search relationship. This allows for searching aligned spectral curves by channel and operating condition, and searching for corresponding aligned spectral curves by spectral fingerprint parameter vectors, ultimately forming a narrowband spectral fingerprint library that can be used in subsequent steps.

[0027] In this embodiment, step two specifically includes: Geometric calibration of the seedling raising area is performed under a unified coordinate system to obtain the installation coordinates and orientation angle of each lamp, generating a lamp pose table. The unified coordinate system is used to describe the relative positional relationship between the lamps, seedling trays, and obstructions. Geometric calibration includes determining the origin, coordinate axis directions, and scale reference of the coordinate system, mapping the installation position of each lamp to a coordinate triplet, and mapping the orientation of each lamp to a set of angle parameters. The lamp pose table includes at least a lamp identification field, an installation coordinate field, an orientation angle field, an installation method field, and a lamp geometric parameter field associated with subsequent optical propagation calculations to support pose retrieval and updating by lamp identification.

[0028] The seedling tray area is divided into grids according to a preset fixed grid size. A coordinate index is written for each grid point, and a seedling tray grid index table is generated. The preset fixed grid size is used to determine the spatial discretization accuracy. The grid point is a set of sampling positions generated according to the grid size within the seedling tray area. The coordinate index is a unique number of the grid point and corresponds one-to-one with the coordinate value of the grid point. The seedling tray grid index table contains at least a grid index field, a grid point coordinate field, a seedling tray identifier field to which the grid belongs, and an index key field bound to the subsequent incident spectral flux expression, so as to realize the addressable calculation of the incident spectral flux at the grid level.

[0029] Measure the height of each lamp from the seedling tray plane and record the boundary polygon of the obstruction in the seedling raising area. Write the height data and boundary polygon data into the obstruction geometry table. The height data is used to determine the vertical distance parameter between the lamp and the seedling tray plane. The boundary polygon of the obstruction is used to describe the occupied area and projection range of the obstruction in a unified coordinate system. The obstruction geometry table includes at least the obstruction identification field, the boundary polygon vertex sequence field, the height data field, and the obstruction determination field related to the subsequent direct component calculation, so that the optical path obstruction status from each grid point to each lamp can be determined according to the obstruction geometry table.

[0030] The reflectance curves of the seedling rack and surrounding materials are collected, and the reflectance curves are mapped onto a unified coordinate system according to the material region to generate a material reflectance mapping chart. The reflectance curves are used to characterize the proportion of incident light reflected by different materials in each band, and the material region is used to describe the spatial distribution boundary of the material in the unified coordinate system. The mapping establishes a relationship between the material region and the corresponding reflectance curve and writes it into the region index. The material reflectance mapping chart should include at least a material identifier field, a material region boundary field, a reflectance curve field, and a region index field to support the query of reflectance by spatial location when calculating the primary reflection component.

[0031] The emission angle distribution parameters of each narrowband supplementary lighting channel are read from the narrowband spectral fingerprint database and written together with the luminaire pose table, seedling tray mesh index table, occlusion geometry table, and material reflectivity map into the optical propagation configuration file. The emission angle distribution parameters are used to describe the radiation intensity distribution of each narrowband supplementary lighting channel in different emission angle directions. The reading is performed by extracting the emission angle distribution parameters from the narrowband spectral fingerprint database according to the channel identifier and forming a channel angle distribution entry. The optical propagation configuration file is used to uniformly encapsulate the luminaire pose, mesh index, occlusion geometry, material reflectivity, and channel angle distribution information required for optical propagation calculation. The optical propagation configuration file contains at least a luminaire pose segment, a mesh definition segment, an occlusion definition segment, a material map segment, and a channel angle distribution segment to enable reproducible incident spectral flux calculation for each mesh point in the future.

[0032] Based on the optical propagation configuration file, the direct component and primary reflection component from each channel of each luminaire are calculated for each grid point. The direct component and primary reflection component are integrated by band to obtain the expression for the incident spectral flux of the grid point. The calculation of the direct component includes establishing the geometric optical path from the luminaire to the grid point based on the luminaire pose segment and the grid definition segment, determining whether the optical path is blocked based on the occlusion definition segment, and determining the emission weight of the optical path direction based on the channel angle distribution segment under the condition that it is not blocked. The calculation of the primary reflection component includes determining the band reflectivity of the reflection area based on the material map segment and establishing the two-segment optical path relationship from the luminaire to the reflection area to the grid point, determining the occlusion state of the two optical paths based on the occlusion definition segment, and calculating the primary reflection contribution when the condition is met. The band integration is to sum the direct component and primary reflection component on the preset band set to obtain the incident spectral flux of each band. The expression for the incident spectral flux of the grid point is an expression or function that can output the incident spectral flux with the grid index, luminaire identifier, and channel identifier as independent variables.

[0033] By binding the grid point incident spectral flux expression with the seedling tray grid index table, a seedling raising spatial spectral energy propagation operator is obtained. Binding involves establishing a mapping relationship between the input / output interface of the grid point incident spectral flux expression and the grid index key of the seedling tray grid index table. This allows any grid index to call the corresponding incident spectral flux expression to obtain the incident spectral flux result for that grid point. The seedling raising spatial spectral energy propagation operator is a callable computational object driven by an optical propagation configuration file and addressed by the grid index. It is used in subsequent steps to generate grid-level incident spectral flux samples under given channel driving conditions.

[0034] In this embodiment, the improved DeepONet model is specifically as follows: The seedling raising spatial spectral energy propagation operator is input into the incident spectral energy sample construction module. By running the seedling raising spatial spectral energy propagation operator under different narrowband channel driving combinations, the incident spectral flux samples of each grid point of the seedling tray grid are obtained. Each set of incident spectral flux samples is bound to its corresponding channel drive combination identifier and grid index to form an incident spectral energy sample set containing multiple sets of grid-level incident spectral fluxes. The incident spectrum energy sample set is input into the tensor encoding module. By dividing each set of grid-level incident spectral flux samples in the incident spectrum energy sample set into multi-resolution segments according to spectral bands, a multi-resolution spectral band sequence is obtained. The multi-resolution spectral band sequence is encoded into an input tensor according to a preset tensor organization rule, while retaining the one-to-one correspondence between the grid index and the channel drive combination identifier and the input tensor, thus forming the spectral band input tensor. The spectral input tensor is input into the physiological condition vector construction module to collect and align the chlorophyll index, leaf temperature, water content surrogate quantity and seedling age stage label corresponding to the same sampling time as the incident spectral energy sample set. The aligned chlorophyll index, leaf temperature, water content proxy quantity and seedling age stage label are combined into a condition vector according to the preset splicing rules. The condition vectors are bound to the spectral input tensors according to the sample identifier and the grid index, resulting in a set of condition vectors that correspond one-to-one with each spectral input tensor. The conditional vector set is input into the operator training and solidification module, the operator training input, operator training conditions and operator training output are set, the data is normalized and the normalization coefficient is recorded, and the data fitting error and physical constraint error are calculated to update the model parameters and generate the rice seedling absorption operator field.

[0035] The improved DeepONet model proposed in this step shares similarities with the traditional DeepONet model in its basic modeling paradigm. Both models focus on "operator learning" as their core objective, mapping input information to output results and learning this mapping relationship through a sample-driven approach. Both require constructing pairs of training samples, ensuring each sample contains an input representation that the model can read, along with its corresponding output representation. During training, the model parameters are updated by calculating the difference between predicted and true values, allowing the model to gradually approximate the target operator. Both models feature an iterative training process oriented towards batch samples, including dividing the training and validation sets, reading samples in batches for prediction, updating parameters based on error, monitoring the training process using validation error, and fixing parameters after a stopping condition is met. Both follow the common path of "input organization - sample correspondence - error calculation - parameter update - model solidification" and emphasize the one-to-one correspondence between input and output in the sense of index to ensure the correctness of sample pairing in operator learning, thereby supporting the output of corresponding prediction results when given new input, and meeting the requirements for use as a callable operator.

[0036] The difference lies in the improved DeepONet model proposed in this step, which is reflected in the organization of the entire process of input construction, condition injection, index binding, and training constraints. Instead of directly using a single-resolution continuous spectral sequence as the input, the grid-level incident spectral flux samples are divided into multi-resolution spectral bands according to wavebands, forming a multi-resolution spectral band sequence. This sequence is then encoded into spectral band input tensors according to preset tensor organization rules, ensuring that the input carries structural information across waveband scales and maintains a one-to-one correspondence with the grid index and channel-driven combination identifier. During training, the model introduces chlorophyll index, leaf temperature, water content surrogate quantity, and seedling age stage label corresponding to the same sampling time as the incident spectral energy sample. These are then combined according to preset splicing rules to form a condition vector. The condition vector and the spectral band input tensor are bound to the sample identifier and grid index to form a condition vector set, thus expanding operator learning from "output determined solely by incident spectral energy" to "output jointly determined by incident spectral energy and physiological state." The training and solidification phase not only establishes a training sample index table to achieve a one-to-one correspondence between the spectral input tensor, condition vector, and output sequence, but also performs grouping normalization on the spectral input tensor, condition vector, and output sequence respectively and records the normalization coefficients to unify the dimensions and numerical scale. In addition to the data fitting error, the error composition also explicitly calculates the physical constraint error and incorporates it into the total error with a preset weight for parameter update. Furthermore, the parameter with the smallest verification error is selected as the solidified parameter through the parameter snapshot table and the verification error record table, forming a rice seedling absorption operator field that can be directly called.

[0037] The beneficial effects of the improvements are as follows: by dividing the grid-level incident spectral flux into multi-resolution segments by band and encoding it as a spectral input tensor, the model input can simultaneously retain narrowband peak details and cross-band energy distribution patterns, reducing the omission of key band information by single-resolution representation and ensuring stable representation of incident spectral energy characteristics during training and inference. Furthermore, by forming conditional vectors from chlorophyll index, leaf temperature, and water content surrogates with seedling age stage labels and binding them to the spectral input tensor at the sample and grid levels, the same incident spectral energy can achieve differentiated absorption outputs under different physiological states, avoiding misattribution of physiological differences to spectral differences. This improves the rice seedling absorption operator field across seedling age, water state, and temperature states. The system ensures usability and consistency. By establishing a training sample index table, a strict one-to-one correspondence is achieved between the input tensor, condition vector, and output sequence. The three types of data are grouped and normalized according to their dimensions, and the normalization coefficients are recorded. This makes the numerical scale of the training process more controllable and reduces the instability of parameter updates. By introducing physical constraint errors corresponding to non-negativity constraints, energy upper bound constraints, and spectral smoothing constraints into the total error, the output sequence meets the basic boundary conditions of absorption and the requirements of spectral continuity. This reduces the probability of negative absorption, overbound absorption, or unreasonable spectral oscillations. Combined with the parameter solidification strategy that minimizes verification error, the final rice seedling absorption operator field is more easily called by the subsequent three-domain self-consistent error tensor calculation link.

[0038] In this embodiment, the operator training input, operator training conditions, and operator training output are set. Data normalization is performed and the normalization coefficients are recorded. The data fitting error and physical constraint error are calculated to update the model parameters, generating a rice seedling absorption operator field. Specifically: The operator training input is set as the spectral input tensor set, the operator training condition is set as the condition vector set, and the operator training output is set as the grid-level effective absorption spectral energy sequence and the key absorption band proportion sequence. Establish a training sample index table to establish a one-to-one correspondence between spectral input tensors, conditional vectors, and output sequences; For each sample in the training sample index table, data normalization is performed and the normalization coefficients are recorded. The normalization process includes: Normalize the spectral input tensor by band energy scale and by grid flux scale; normalize the condition vector by dimension grouping; normalize the output sequence by absorbed energy scale. Set the number of iteration training rounds, the number of samples per batch, and the update step size. Read the spectral input tensor, conditional vector, and output sequence from the training sample index table in batches and calculate the predicted value of the output sequence. The data fitting error is calculated based on the predicted value and the true value of the output sequence, and the physical constraint error is calculated based on the predicted value of the output sequence. The physical constraint error includes the non-negative constraint error applied to the effective absorption spectrum energy sequence, the energy upper bound constraint error applied to the effective absorption spectrum energy sequence, and the spectral smoothing constraint error applied to the key absorption band proportion sequence. The data fitting error and the physical constraint error are combined into a total error according to a preset weight; Based on the total error, the model parameters are updated and the updated model parameters are written into the parameter snapshot table. At the same time, the corresponding spectral input tensor and condition vector are read from the validation samples to calculate the validation error and write the validation error into the validation error record table. Training ends when the verification error meets the preset stopping condition. The model parameter corresponding to the minimum verification error is selected from the parameter snapshot table as the fixed parameter, and its input and output interface is defined as an operator to obtain the rice seedling absorption operator field.

[0039] In this embodiment, step four specifically includes: Select a preset number of fixed sampling points on the seedling tray grid corresponding to the seedling tray grid index table, assign a sampling point identifier to each fixed sampling point and bind it to the corresponding grid index, and generate a sampling point index table. According to the preset fixed sampling interval, chlorophyll index readings, leaf temperature readings, seedling height readings, leaf area readings, dry matter sampling and weighing values, and stomatal conductance proxy readings are collected for each fixed sampling point in the sampling point index table. Each reading record is written with the sampling point identifier and sampling interval identifier to form a set of original readings for multiple indicators. The raw reading set of multiple indicators is aligned with the sampling point identifier and sampling interval identifier by time, and an alignment record is generated. Missing value removal and outlier removal are performed on the alignment record. For each sampling point, within each sampling interval, the retained chlorophyll index readings, leaf temperature readings, seedling height readings, leaf area readings, dry matter sampling and weighing values, and stomatal conductance proxy readings are combined into a multi-index observation vector according to a preset dimension order. The environmental sensor readings corresponding to the current sampling interval are read and packaged with the multi-index observation vector to generate a sampling point-level phenotypic aligned sample. All sample point-level phenotypic aligned samples are indexed and stored according to the sample point identifier, sample interval identifier, and grid index to obtain the rice seedling phenotypic observation aligned set.

[0040] In this embodiment, step five specifically includes: Based on a preset standard light source reference spectrum, baseline modeling is performed on the dark field spectrum. Baseline modeling involves locating characteristic peaks in the standard light source reference spectrum and fitting a wavelength axis offset curve based on the characteristic peak location results. The response nonlinearity curve is fitted based on the nominal radiation distribution and the measured radiation distribution of the standard light source reference spectrum, and the wavelength axis offset curve, the response nonlinearity curve and the spectrometer identification are written into the spectrometer correction parameter entry; Under uniform light field conditions, the imaging device acquires a multi-exposure image sequence, groups the multi-exposure image sequence according to the exposure parameters, and calculates the pixel brightness statistics of each exposure group; Illumination response curves are fitted based on pixel brightness statistics and uniform light field illuminance benchmarks, and lens vignetting correction matrices are fitted based on the spatial brightness distribution of multi-exposure images; the illuminance response curves, lens vignetting correction matrices, and imaging device identifiers are then written into the imaging correction parameter entries. Under stable environmental conditions, a reference reading sequence is acquired from the environmental sensor. A time series fitting is performed on the reference reading sequence to obtain the drift term. A piecewise fitting is performed on the rising and falling edges of the reference reading sequence to obtain the hysteresis term. The drift term and the hysteresis term, along with the environmental sensor identifier, are written into the environmental correction parameter entry. The spectral correction parameter entries, imaging correction parameter entries, and environmental correction parameter entries are written into a unified correction parameter table, and a unified correction interface is defined for the correction parameter table. The unified correction interface includes wavelength axis offset correction and response nonlinearity correction for spectral observation, illuminance response correction and lens vignetting correction for imaging observation, and drift correction and hysteresis correction for environmental observation. When inputting spectral observations, imaging observations, and environmental observations, a unified calibration interface matching the equipment identifier is invoked to output the calibrated physical quantities. The calibrated physical quantities are then organized according to time slice identifiers and grid indices to form a cross-modal observation self-consistent calibration field.

[0041] In this embodiment, step six specifically includes: Read the multi-index observation vector corresponding to each sampling interval, as well as the corresponding grid index and sampling point identifier, from the rice seedling phenotypic observation alignment set to generate a sampling interval index list; Read the spectral observations, imaging observations, and environmental observations corresponding to the sampling interval index list from the cross-modal observation self-consistent calibration field, and call the calibration interface that matches the device identifier to perform spectral observation calibration, imaging observation calibration, and environmental observation calibration respectively, so as to obtain the calibrated spectral observations, calibrated imaging observations, and calibrated environmental observations. The candidate values ​​of the supplementary lighting parameters to be determined are organized into a channel drive candidate table according to the narrowband supplementary lighting channel identifier. The candidate values ​​of the supplementary lighting parameters to be determined include the narrowband supplementary lighting channel drive amount, optical period parameter, and pulse duty cycle parameter. The channel-driven candidate list is input into the seedling raising spatial spectral energy propagation operator to obtain the grid-level incident spectral flux corresponding to the grid index; The energy distribution difference, center wavelength difference, half-peak width difference, and spectral tail difference are calculated by combining the grid-level incident spectral flux with the corrected spectral observations, corrected imaging observations, and corrected environmental observations, respectively, to obtain the three-domain self-consistency error tensor. The parameter set is updated based on the candidate values ​​of the supplementary lighting parameters to be determined formed by the three-domain self-consistent error tensor. Repeat the calculation of the three-domain self-consistent error tensor and the update of the parameter set until the number of iterations reaches the set value or the error convergence condition is met, to obtain the final narrowband supplementary lighting parameter set.

[0042] Example 1: To verify the feasibility of this invention in practice, it was applied to a factory-style rice seedling raising center using a three-dimensional, multi-layered seedling raising rack system. Each layer of the rack holds a standard seedling tray, and a multi-channel narrowband supplemental lighting fixture is installed at the top, with channels covering narrowband outputs of red and blue light. A prominent problem encountered by this center during continuous production was inconsistent growth of the same batch of seedling trays in different layers and corners, with some areas exhibiting excessive growth, pale leaf color, or excessively high leaf temperature. Maintenance personnel attempted to adjust the drive current, photoperiod, and duty cycle parameters using a fixed formula, but the adjustment effect fluctuated with operating time, especially after the lights had been operating continuously for a period of time, resulting in unstable seedling growth even with the same settings. On-site investigation revealed that the lamp's heat dissipation and continuous operating time caused slight shifts in the narrowband peak position and changes in the half-peak width. Furthermore, the reflective materials and shielding structures around the seedling trays resulted in spatially uneven superposition of direct and primary reflected light. Simultaneously, the spectrometer, camera, and environmental sensors exhibited their own offsets, nonlinearities, and drifts, leading to discrepancies between the "measured data" and the "actual spectral energy reaching the seedling tray," further amplifying the uncertainty of manual parameter adjustments. Based on these issues, the center chose to conduct a comparative application on two adjacent seedling trays. One set used the self-consistent calibration method of this invention to form a set of supplemental lighting parameters and then implemented it. The other set maintained the commonly used practice in existing technologies, namely, empirically setting parameters based on nominal parameters and a small number of point illuminance measurements, serving as a control.

[0043] In practical applications, the process begins by setting three driving current levels for each narrowband supplementary lighting channel while maintaining a constant sampling duration. Emission spectral curves for each level are collected at a fixed distance from the light outlet of the lamp. Simultaneously, dark-field spectra are acquired, and dark-field subtraction and wavelength axis alignment are performed. The center wavelength, half-maximum width, spectral tail skew, peak radiant flux, and first-order difference peak position are calculated. Temperature drift slope and working time drift slope are incorporated into the spectral fingerprint parameter vector, ultimately forming a searchable narrowband spectral fingerprint library. Subsequently, the seedling raising area is geometrically calibrated in a unified coordinate system. The installation coordinates and orientation angle of each lamp are recorded. The seedling tray area is divided into grids of fixed size and numbered. The height of the lamps from the seedling tray plane is measured, and the boundary polygons of obstructions are recorded. Simultaneously, the band reflectance curves of the seedling frame and surrounding materials are collected and mapped onto the coordinate system.

[0044] The system reads the emission angle distribution parameters of each channel from a narrowband spectral fingerprint database, and writes the lamp pose, grid index, occlusion geometry, and material reflectivity into an optical propagation configuration file. Based on the direct component and the first reflection component, it calculates the incident spectral flux expression for each grid point and binds it to the grid index to obtain the seedling raising space spectral energy propagation operator. To further transform the "grid incident spectral energy" into "rice seedling absorption side characterization," the system runs the propagation operator under different channel driving combinations to generate incident spectral flux samples, and divides them into spectral input tensors according to multi-resolution bands. Simultaneously, it collects the chlorophyll index, leaf temperature, water content surrogate quantity, and seedling age stage label corresponding to the samples, concatenates them into a conditional vector, binds it to the spectral input tensor, and inputs the operator training and solidification process. After normalization, error calculation, and parameter updates, and combined with the verification error stopping condition, the rice seedling absorption operator field is solidified.

[0045] During the seedling raising process, fixed sampling points are set on the seedling tray grid and bound to the grid index. Chlorophyll index, leaf temperature, seedling height, leaf area, dry matter sampling and weighing values, and stomatal conductance are collected at fixed sampling intervals. At the same time, environmental sensor data is read. The readings are time-aligned and missing and outlier values ​​are removed to form a rice seedling phenotypic observation alignment set. The system also collects the wavelength axis offset curve and response nonlinearity curve fitted between the dark field of the spectral equipment and the reference spectrum of the standard light source. It collects the illuminance response curve and lens vignetting correction matrix fitted between the multi-exposure image of the uniform light field. It collects the drift and hysteresis terms fitted to the benchmark reading sequence under stable conditions and writes them into the correction parameter table according to the equipment identification. A unified correction interface is defined to form a cross-modal observation self-consistent correction field. Finally, within the same sampling interval, the system calls the calibration interface to unify the spectral, imaging, and environmental observation apertures. It then constructs a three-domain self-consistent error tensor by combining the rice seedling phenotypic observation alignment set with the rice seedling absorption operator field. The system also updates the candidate set of supplementary lighting parameters, converging to obtain an executable set of narrowband supplementary lighting parameters, which is then sent to the luminaire controller. The controller executes the drive current, light period, and pulse duty cycle settings according to this set during the daily prescribed supplementary lighting period. The control group still uses an empirical setting method, mainly relying on the nominal wavelength and a small number of point illuminance values ​​for manual correction to maintain the same supplementary lighting period and similar energy consumption levels.

[0046] To verify the beneficial effects, data were recorded on-site focusing on "spatial consistency, spectral consistency, and seedling condition consistency within the same seedling raising cycle." The locations were the same-layer areas of two adjacent seedling raising racks in the seedling raising center. Supplemental lighting was provided continuously from early morning to evening each day, and the evaluation period covered the critical growth stages from seedling emergence to transplanting. The status was retested after the lamps had been running continuously for hundreds of hours. Data shows that the method of this invention can maintain consistency between the calculated incident spectral energy and the observation aperture against the background of spectral shape changes caused by lamp temperature rise and aging, providing a stable basis for parameter updates. Simultaneously, because the contribution of direct and primary reflections is calculated at the grid level and participates in the error tensor, insufficient supplemental lighting in spatial corner areas is specifically corrected, and the seedling condition dispersion at different locations in the seedling tray is significantly reduced. The table below presents key comparative data for the two seedling raising racks. All indicators are from on-site records and sampling statistics. In the table, "mean" is the statistical mean of the sampling points or grid, and "coefficient of variation" is the ratio of standard deviation to mean, used to reflect the degree of spatial non-uniformity.

[0047] Table 1. Comparison Data of Self-Conformance Calibration for Narrow-Strip Supplemental Lighting in Factory Seedling Raising As shown in Table 1, the control method exhibits more pronounced center wavelength drift and half-width variation after continuous lamp operation, leading to a greater gap between the nominal setting and the actual light output. This, coupled with the unevenness caused by spatial reflection and shading, amplifies the incident spectral energy deviation in corner areas, directly reflecting lower similarity in illuminance distribution structure and higher fitting error in spectral energy distribution. Correspondingly, the control method shows a lower mean chlorophyll index and a larger coefficient of variation, increased seedling height dispersion accompanied by an increase in the proportion of abnormal seedlings, indicating that the differences in light exposure at different locations within the same seedling tray are amplified in growth indicators. The proposed method integrates spectral fingerprinting, spatial propagation operators, absorption operator fields, phenotypic alignment observations, and cross-modal correction into a single computational chain, and then uses a three-domain self-consistent error tensor to drive parameter set updates. This significantly reduces the spatial dispersion coefficient of the incident spectral energy, narrows the relative deviation in corner areas to a smaller range, reduces the spatial fluctuations of leaf temperature and chlorophyll index, and increases the mean and dispersion of dry matter accumulation. Ultimately, this results in a decrease in the proportion of abnormal seedlings and an increase in the seedling tray qualification rate. It is worth noting that, under the condition that the power consumption of the unit seedling tray is basically the same, the method of the present invention achieves more stable seedling condition consistency and more repeatable parameter convergence results, which can adapt to the actual production fluctuations caused by differences in seedling rack structure and changes in lamp status, and meet the requirements of industrial seedling raising for stability and traceability.

[0048] 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.

Claims

1. A self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design, characterized in that, Includes the following steps: Step 1: Acquire the emission spectrum curves of each narrowband supplementary light channel under multi-level driving, and perform dark field subtraction and wavelength axis alignment to obtain the narrowband spectral fingerprint library; Step 2: Based on the emission angle distribution parameters of each channel in the narrowband spectral fingerprint spectral library and the direct component and primary reflection component from each channel of each lamp, the grid point incident spectral flux expression is obtained by band integration, and index binding is performed to obtain the seedling raising spatial spectral energy propagation operator; Step 3: Input the seedling spatial spectral energy propagation operator into the improved DeepONet model, and obtain the rice seedling absorption operator field through the incident spectral energy sample construction module, tensor encoding module, physiological condition vector construction module and operator training and solidification module; Step 4: Determine fixed sampling points on the seedling tray grid and bind a grid index to each sampling point. Collect various readings at fixed sampling intervals, perform time alignment and remove missing and outlier values ​​to generate a rice seedling phenotypic observation alignment set. Step 5: Fit the wavelength axis offset curve, response nonlinearity curve, illuminance response curve and lens vignetting correction matrix, drift term and hysteresis term, and write them into the correction parameter table according to the equipment identification, define a unified correction interface, and form a cross-modal observation self-consistent correction field. Step 6: Based on the rice seedling phenotypic observation alignment set, cross-modal observation self-consistent correction field, and rice seedling absorption operator field, obtain the three-domain self-consistent error tensor, and update the parameter set to obtain the final narrowband supplementary lighting parameter set.

2. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 1, characterized in that, Step one specifically involves: Three drive current levels are set for each narrowband fill light channel, and a constant sampling duration is set for each drive current level. The channel identifier and drive current level identifier are recorded, and a drive sampling configuration table is generated. Within the sampling time corresponding to each drive current level, the heat sink temperature sequence and the channel working time sequence are collected synchronously. Temperature range identifiers are generated for the heat sink temperature sequence according to the preset temperature segmentation rules, and working time range identifiers are generated for the channel working time sequence according to the preset time segmentation rules. Use a spectral acquisition device to acquire the emission spectrum curve of the corresponding narrowband supplementary light channel at a preset fixed distance position at the light outlet of the lamp, and acquire the corresponding dark field spectrum curve. The emission spectrum curve is subjected to dark field subtraction using the dark field spectral curve to obtain the dark field subtracted spectral curve, and then wavelength axis alignment is performed on the dark field subtracted spectral curve to obtain the aligned spectral curve. For each aligned spectral curve, calculate the center wavelength, full width at half maximum (FWHM), spectral tail skew, peak radiant flux, and first-order difference peak position, and arrange them in a preset order to form a spectral fingerprint parameter vector. The temperature drift slope is calculated based on the radiator temperature sequence, the working time drift slope is calculated based on the channel working time sequence, and the temperature drift slope and the working time drift slope are incorporated into the spectral fingerprint parameter vector. The spectral fingerprint parameter vector is written into the same index table along with the channel identifier, drive current level identifier, temperature range identifier, and operating time range identifier. The aligned spectral curve is then bound one by one to the index table entries corresponding to the spectral fingerprint parameter vector to obtain the narrowband spectral fingerprint library.

3. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 1, characterized in that, Step two specifically involves: Geometric calibration of the seedling raising area is performed under a unified coordinate system to obtain the installation coordinates and orientation angle of each lamp, and a lamp pose table is generated. Divide the seedling tray area into grids according to a preset fixed grid size, write the coordinate index for each grid point and generate a seedling tray grid index table; Measure the height of each lamp from the seedling tray plane and record the boundary polygon of the obstruction in the seedling raising area. Write the height data and boundary polygon data into the obstruction geometry table. Collect the band reflectance curves of the seedling frame and surrounding materials, and map the band reflectance curves to a unified coordinate system according to the material area to generate a material reflectance map. The emission angle distribution parameters of each narrowband supplementary light channel are read from the narrowband spectral fingerprint spectrum library, and the emission angle distribution parameters are written together with the lamp pose table, seedling tray grid index table, occlusion geometry table and material reflectivity map into the optical propagation configuration file; Based on the optical propagation profile, the direct component and primary reflection component from each channel of each lamp are calculated for each grid point, and the direct component and primary reflection component are integrated by band to obtain the expression for the incident spectral flux of the grid point. By binding the expression for the incident spectral flux at the grid points with the seedling tray grid index table, the spatial spectral energy propagation operator for seedling raising is obtained.

4. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 1, characterized in that, The improved DeepONet model is specifically as follows: The seedling raising spatial spectral energy propagation operator is input into the incident spectral energy sample construction module. By running the seedling raising spatial spectral energy propagation operator under different narrowband channel driving combinations, the incident spectral flux samples of each grid point of the seedling tray grid are obtained. Each set of incident spectral flux samples is bound to its corresponding channel drive combination identifier and grid index to form an incident spectral energy sample set containing multiple sets of grid-level incident spectral fluxes. The incident spectrum energy sample set is input into the tensor encoding module. By dividing each set of grid-level incident spectral flux samples in the incident spectrum energy sample set into multi-resolution segments according to spectral bands, a multi-resolution spectral band sequence is obtained. The multi-resolution spectral band sequence is encoded into an input tensor according to a preset tensor organization rule, while retaining the one-to-one correspondence between the grid index and the channel drive combination identifier and the input tensor, thus forming the spectral band input tensor. The spectral input tensor is input into the physiological condition vector construction module to collect and align the chlorophyll index, leaf temperature, water content surrogate quantity and seedling age stage label corresponding to the same sampling time as the incident spectral energy sample set. The aligned chlorophyll index, leaf temperature, water content proxy quantity and seedling age stage label are combined into a condition vector according to the preset splicing rules. The condition vectors are bound to the spectral input tensors according to the sample identifier and the grid index, resulting in a set of condition vectors that correspond one-to-one with each spectral input tensor. The conditional vector set is input into the operator training and solidification module, the operator training input, operator training conditions and operator training output are set, the data is normalized and the normalization coefficient is recorded, and the data fitting error and physical constraint error are calculated to update the model parameters and generate the rice seedling absorption operator field.

5. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 4, characterized in that, The process involves setting the operator training input, operator training conditions, and operator training output, performing data normalization processing and recording the normalization coefficients, calculating the data fitting error and physical constraint error to update the model parameters, and generating a rice seedling absorption operator field. Specifically: The operator training input is set as the spectral input tensor set, the operator training condition is set as the condition vector set, and the operator training output is set as the grid-level effective absorption spectral energy sequence and the key absorption band proportion sequence. Establish a training sample index table to establish a one-to-one correspondence between spectral input tensors, conditional vectors, and output sequences; For each sample in the training sample index table, data normalization is performed and the normalization coefficients are recorded. The normalization process includes: Normalize the spectral input tensor by band energy scale and by grid flux scale; normalize the condition vector by dimension grouping; normalize the output sequence by absorbed energy scale. Set the number of iteration training rounds, the number of samples per batch, and the update step size. Read the spectral input tensor, conditional vector, and output sequence from the training sample index table in batches and calculate the predicted value of the output sequence. The data fitting error is calculated based on the predicted value and the true value of the output sequence, and the physical constraint error is calculated based on the predicted value of the output sequence. The physical constraint error includes the non-negative constraint error applied to the effective absorption spectrum energy sequence, the energy upper bound constraint error applied to the effective absorption spectrum energy sequence, and the spectral smoothing constraint error applied to the key absorption band proportion sequence. The data fitting error and the physical constraint error are combined into a total error according to a preset weight; Based on the total error, the model parameters are updated and the updated model parameters are written into the parameter snapshot table. At the same time, the corresponding spectral input tensor and condition vector are read from the validation samples to calculate the validation error and write the validation error into the validation error record table. Training ends when the verification error meets the preset stopping condition. The model parameter corresponding to the minimum verification error is selected from the parameter snapshot table as the fixed parameter, and its input and output interface is defined as an operator to obtain the rice seedling absorption operator field.

6. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 1, characterized in that, Step four specifically involves: Select a preset number of fixed sampling points on the seedling tray grid corresponding to the seedling tray grid index table, assign a sampling point identifier to each fixed sampling point and bind it to the corresponding grid index, and generate a sampling point index table. According to the preset fixed sampling interval, chlorophyll index readings, leaf temperature readings, seedling height readings, leaf area readings, dry matter sampling and weighing values, and stomatal conductance proxy readings are collected for each fixed sampling point in the sampling point index table. Each reading record is written with the sampling point identifier and sampling interval identifier to form a set of original readings for multiple indicators. The raw reading set of multiple indicators is aligned with the sampling point identifier and sampling interval identifier by time, and an alignment record is generated. Missing value removal and outlier removal are performed on the alignment record. For each sampling point, within each sampling interval, the retained chlorophyll index readings, leaf temperature readings, seedling height readings, leaf area readings, dry matter sampling and weighing values, and stomatal conductance proxy readings are combined into a multi-index observation vector according to a preset dimension order. The environmental sensor readings corresponding to the current sampling interval are read and packaged with the multi-index observation vector to generate a sampling point-level phenotypic aligned sample. All sample point-level phenotypic aligned samples are indexed and stored according to the sample point identifier, sample interval identifier, and grid index to obtain the rice seedling phenotypic observation aligned set.

7. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 1, characterized in that, Step five specifically involves: Based on a preset standard light source reference spectrum, baseline modeling is performed on the dark field spectrum. The baseline modeling involves locating characteristic peaks in the standard light source reference spectrum and fitting a wavelength axis offset curve based on the characteristic peak location results. The response nonlinearity curve is fitted based on the nominal radiation distribution and the measured radiation distribution of the standard light source reference spectrum, and the wavelength axis offset curve, the response nonlinearity curve and the spectrometer identification are written into the spectrometer correction parameter entry; Under uniform light field conditions, the imaging device acquires a multi-exposure image sequence, groups the multi-exposure image sequence according to the exposure parameters, and calculates the pixel brightness statistics of each exposure group; Illumination response curves are fitted based on pixel brightness statistics and uniform light field illuminance benchmarks, and lens vignetting correction matrices are fitted based on the spatial brightness distribution of multi-exposure images; the illuminance response curves, lens vignetting correction matrices, and imaging device identifiers are then written into the imaging correction parameter entries. Under stable environmental conditions, a reference reading sequence is collected from an environmental sensor, and a time series fitting is performed on the reference reading sequence to obtain the drift term. A rising edge and falling edge segmented fitting is performed on the reference reading sequence to obtain the hysteresis term, and the drift term and hysteresis term, along with the environmental sensor identifier, are written into the environmental correction parameter entry; The spectral correction parameter entries, imaging correction parameter entries, and environmental correction parameter entries are written into a unified correction parameter table, and a unified correction interface is defined for the correction parameter table. The unified correction interface includes wavelength axis offset correction and response nonlinearity correction for spectral observation, illuminance response correction and lens vignetting correction for imaging observation, and drift correction and hysteresis correction for environmental observation. When inputting spectral observations, imaging observations, and environmental observations, a unified calibration interface matching the equipment identifier is invoked to output the calibrated physical quantities. The calibrated physical quantities are then organized according to time slice identifiers and grid indices to form a cross-modal observation self-consistent calibration field.

8. The self-consistent calibration method for narrowband supplemental lighting parameters of rice seedlings based on spectral inverse design according to claim 1, characterized in that, Step six specifically involves: Read the multi-index observation vector corresponding to each sampling interval, as well as the corresponding grid index and sampling point identifier, from the rice seedling phenotypic observation alignment set to generate a sampling interval index list; Read the spectral observations, imaging observations, and environmental observations corresponding to the sampling interval index list from the cross-modal observation self-consistent calibration field, and call the calibration interface that matches the device identifier to perform spectral observation calibration, imaging observation calibration, and environmental observation calibration respectively, so as to obtain the calibrated spectral observations, calibrated imaging observations, and calibrated environmental observations. The candidate values ​​of the supplementary lighting parameters to be determined are organized into a channel drive candidate table according to the narrowband supplementary lighting channel identifier. The candidate values ​​of the supplementary lighting parameters to be determined include the narrowband supplementary lighting channel drive amount, optical period parameter, and pulse duty cycle parameter. The channel-driven candidate list is input into the seedling raising spatial spectral energy propagation operator to obtain the grid-level incident spectral flux corresponding to the grid index; The energy distribution difference, center wavelength difference, half-peak width difference, and spectral tail difference are calculated by combining the grid-level incident spectral flux with the corrected spectral observations, corrected imaging observations, and corrected environmental observations, respectively, to obtain the three-domain self-consistency error tensor. The parameter set is updated based on the candidate values ​​of the supplementary lighting parameters to be determined formed by the three-domain self-consistent error tensor. Repeat the calculation of the three-domain self-consistent error tensor and the update of the parameter set until the number of iterations reaches the set value or the error convergence condition is met, to obtain the final narrowband supplementary lighting parameter set.

Images