Sugarcane pollen development state recognition method combining multispectral imaging and AI recognition

Abstract

The application discloses a sugarcane pollen development state recognition method combining multispectral imaging and AI recognition, relates to the technical field of agricultural detection, and specifically comprises the following steps: first, assembling equipment and preparing samples; then, collecting multimodal data; subsequently, calibrating signals and fusing features; then, calling an AI model for reasoning; finally, updating environmental parameters in real time, dynamically correcting signals, and outputting final detection results and reports after confidence verification; the application integrates portable equipment, exclusive parameter fitting and multimodal data acquisition technology, solves the problem that traditional detection relies on large instruments and is difficult to adapt to fields, improves data accuracy and stability, and meets real-time detection requirements; meanwhile, the application adopts multimodal feature fusion and a double-branch AI model, realizes cross-species accurate recognition, reduces the interference of variety differences, enhances anti-interference capability, and provides high-precision and strong-adaptability scientific support for pollen vitality identification in sugarcane breeding.

Description

Technical Field

[0001] This invention relates to the field of agricultural detection technology, specifically a method for identifying the developmental status of sugarcane pollen using multispectral imaging combined with AI recognition. Background Technology

[0002] Sugarcane, as a globally important sugar and energy crop, directly impacts the development of the sugar industry and biomass energy sector through its yield and quality. In sugarcane hybridization breeding, pollen development status is a key factor determining pollination success rate and the seed setting rate of hybrid offspring. Accurate detection of pollen development status can effectively screen for high-quality pollen for hybridization, significantly improving breeding efficiency and the probability of selecting superior varieties. With the development of agricultural detection technologies, fluorescence imaging technology, which reflects pollen activity through fluorescent staining, and multispectral imaging technology, which captures differences in pollen spectral characteristics, are increasingly being applied to crop pollen detection. Simultaneously, artificial intelligence recognition technology, with its automated and high-precision feature analysis capabilities, offers the possibility of quantitatively identifying pollen development status. However, current sugarcane pollen detection technologies largely remain at the level of single-technology applications, lacking a systematic solution combining fluorescence-multispectral fusion and AI recognition. Furthermore, there is a lack of specific adaptation designs for the pollen characteristics of different sugarcane varieties, making it difficult to meet the precise and efficient detection needs in breeding scenarios.

[0003] Traditional methods for detecting the developmental status of sugarcane pollen have significant technical shortcomings and are difficult to adapt to the demands of modern breeding for high efficiency and precision. First, manual observation relies on morphological observation under a microscope and visual judgment after staining. This is not only extremely inefficient, but also highly dependent on the operator's experience, resulting in significant subjective errors. It cannot quantify pollen activity and developmental stage, and is prone to misidentifying aborted or immature pollen, leading to low hybridization pollination efficiency. Second, single-modal detection technology has significant limitations. When using only fluorescence imaging, it is easily affected by ambient light intensity and temperature fluctuations, resulting in poor fluorescence signal stability and an inability to distinguish pollen with similar morphology but different activities. When using only multispectral imaging, it is difficult to capture the microscopic features related to pollen activity and cannot accurately identify early aborted pollen. Third, existing detection methods mostly rely on large laboratory instruments. These devices are bulky and not portable, requiring pollen samples to be brought back to the laboratory for testing. This makes it impossible to achieve real-time detection during the flowering period in the field, resulting in delayed results and missing the optimal pollination time. In addition, traditional methods do not have calibration mechanisms designed for the differences in pollen particle size and spectral characteristics among different sugarcane varieties. The accuracy drops significantly when detecting across varieties, making it difficult to meet the detection needs of multi-variety breeding bases. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for identifying the pollen development status of sugarcane using multispectral imaging combined with AI recognition. This method achieves real-time field detection through a portable device integrating fluorescence and multispectral imaging and edge computing functions. The method utilizes dedicated parameter fitting and simultaneous multimodal data acquisition, combined with signal calibration and feature fusion techniques to improve data accuracy. A dual-branch AI model is employed to achieve accurate cross-variety identification, reducing interference from varietal differences and providing a high-precision, highly adaptable pollen development status detection solution for sugarcane breeding.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for identifying the developmental state of sugarcane pollen using multispectral imaging combined with AI recognition, the specific steps of which are as follows:

[0006] S1, Equipment Setup and Sample Preparation: A portable device integrating fluorescence imaging, multispectral imaging, optical synchronization and edge computing functions was set up. Phosphate buffer, FDA and PI mixed staining agent and agarose fixation solution were prepared. Pollen samples were prepared through anther peeling, pollen separation, agarose film fixation, light-protected staining and rinsing. Sugarcane pollen-specific parameters were fitted.

[0007] S2, Multimodal Data Acquisition: Place the pollen sample on the stage of the device and simultaneously acquire fluorescence images, multispectral images, background fluorescence intensity, dark field gray values, record the actual detection temperature and ambient light intensity, and obtain the original fluorescence intensity and original multispectral gray values.

[0008] S3, Signal Calibration and Feature Fusion: Based on the original data and specific parameters of S2, the fluorescence signal and multispectral signal are calibrated by the fluorescence intensity environment adaptive calibration formula and the multispectral reflectance particle size correction calibration formula, respectively. Fluorescence features, spectral features and morphological features are extracted. After processing by the multimodal feature weighted normalization formula, the fused feature vector is generated by the attention mechanism multimodal feature fusion formula.

[0009] S4, AI Model Recognition and Inference: Calls a dual-branch AI recognition model with feature enhancement capabilities. This model is trained with a cross-variety domain adaptive loss function as the optimization objective. The fused feature vector from S3 is input into the model, and the preliminary recognition result is output.

[0010] S5, Environmental Adaptation and Result Output: Real-time updates of environmental parameters, dynamic invocation of the two types of calibration formulas in S3 to correct signals, combined with equipment protection design and environmental adaptive adjustment technology, after verifying the confidence level of the preliminary identification results, outputs the final detection results and corresponding detection reports.

[0011] Furthermore, the fluorescence imaging function in step 1 is achieved through a dual-wavelength excitation component and a high-pixel imaging component, with adjustable light intensity and the imaging component supporting high-bit format data storage; the multispectral imaging function covers the visible to near-infrared spectral range, includes multiple characteristic bands, and has a fast band switching speed.

[0012] Furthermore, the mixed staining agent in step 1 is prepared by FDA, PI and phosphate buffer with a pH of 7.0 to 7.4, wherein the volume ratio of FDA to PI is 3:1; and the concentration of agarose fixative is 0.6% to 1.0%.

[0013] Furthermore, the formula for calculating the fluorescence intensity environmental adaptive calibration in S3 is as follows: ,in, The fluorescence intensity after calibration; The original fluorescence intensity is obtained directly from the fluorescence imaging component; The mean fluorescence intensity of 100% active pollen from 10 core sugarcane varieties was calculated by collecting pollen samples of these 10 varieties with an in vitro germination rate of not less than 90%. The temperature correction coefficient was obtained by fitting the fluorescence signal gradient experiment of 10 sugarcane varieties in the range of 5℃ to 40℃. This is the difference between the actual detected temperature and 25℃. The actual detected temperature is measured by the built-in temperature sensor of the device. The background fluorescence intensity was obtained by averaging the fluorescence intensity at three different locations in the pollen-free area surrounding the pollen sample. The ambient light suppression factor was obtained by experimental fitting of the ambient light gradient in the range of 0 to 10000 lux. The ambient light intensity is measured by the device's built-in ambient light sensor.

[0014] Furthermore, the multispectral reflectance particle size correction calibration calculation formula in S3 is as follows: ,in, The reflectivity after calibration for the i-th band; For multispectral band wavelengths, the preset wavelengths are 450nm, 550nm, 650nm, 800nm, and 950nm. The original grayscale value of the i-th band is obtained by the multispectral imaging component sequentially according to the bands. The dark field gray value of the i-th band is obtained by taking the average of three images of this band under conditions of no light source illumination and lens shading. The gray value of the standard gray card in the i-th band is obtained by placing the standard gray card on the stage and collecting data with the same parameters. is the nominal reflectance of the i-th band of the standard gray card, and is the factory calibration value of the standard gray card; The particle size correction coefficient was obtained by measuring the equivalent diameter of pollen from 10 core sugarcane varieties and fitting it with the corresponding band reflectance deviation. The current equivalent diameter of the pollen is derived by extracting the pollen contour from the multispectral 650nm band image and calculating the area. The pollen grain size was obtained by taking the average of the equivalent diameters of pollen from 10 core sugarcane varieties.

[0015] Furthermore, the multimodal feature weighted normalization calculation formula in S3 is as follows: ; ,in, This represents the standardized result of the j-th feature; j ranges from 1 to 14, corresponding to 3 fluorescence features, 7 spectral features, and 4 morphological features. The j-th original feature value is extracted from the calibrated fluorescence signal multispectral signal and image analysis. The sample mean of the j-th feature is calculated by collecting and detecting 5000 labeled sugarcane pollen samples; The sample standard deviation of the j-th feature is calculated statistically from 5000 labeled sugarcane pollen samples. The weight coefficient for the j-th feature; The signal-to-noise ratio of the j-th feature is obtained by calculating the ratio of the signal difference of this feature in normal pollen and aborted pollen samples to the fluctuation value within the sample. is the correlation coefficient between the j-th feature and pollen development status, obtained through Pearson correlation analysis; For summation index; For the first The signal-to-noise ratio of each feature; For the first The correlation coefficients between these characteristics and pollen development status.

[0016] Furthermore, the multimodal feature fusion calculation formula for the attention mechanism in S3 is as follows: ; ,in, To fuse feature vectors; Let be the attention weight for the j-th feature; The attention activation coefficients were obtained through optimization on an AI model validation set. γ is the mean of the j-th feature of aborted pollen samples, calculated by collecting and testing 2000 aborted pollen samples; γ is a smoothing factor, fixed at 0.001 to avoid the denominator approaching 0; The weight coefficient for the j-th feature; This is the standardized result for the j-th feature.

[0017] Furthermore, the cross-variety domain adaptive loss function in S4 is: ,in, This is the total loss function; The classification loss is calculated using the cross-entropy algorithm to determine the deviation between the model's predictions and the true labels. The domain adversarial loss weights are determined through model validation set tuning. The domain adversarial loss is calculated from the output of the domain discriminator. The variety difference penalty coefficient was determined through model validation set tuning. The variance loss of accuracy among varieties is calculated by subtracting 1 from the number of varieties in this method. The number of varieties is 10. The variance loss is calculated by the deviation between the accuracy of each variety and the average accuracy.

[0018] Furthermore, the two branches of the dual-branch artificial intelligence recognition model with feature enhancement function in S4 correspond to the category distinction and feature enhancement of multimodal fusion features, respectively. The feature enhancement function is realized through the attention mechanism. The model training uses cross-variety pollen labeled samples to construct a training set, covering samples from different developmental stages and abortion states.

[0019] Furthermore, the environmental adaptive adjustment technology in S5 dynamically adjusts imaging-related parameters and signal calibration strategies based on real-time detection data from ambient light and temperature sensors to adapt to changes in the field environment; the confidence verification threshold is set to 0.7 to 0.95, and the detection report includes sample information, imaging parameters, and recognition results.

[0020] Compared with existing technologies, this method for identifying the developmental status of sugarcane pollen using multispectral imaging combined with AI recognition has the following advantages:

[0021] I. This invention solves the technical problems of traditional sugarcane pollen development status detection, which relies on large instruments, is complex to operate, and is difficult to adapt to field scenarios, by constructing a portable device integrating fluorescence imaging, multispectral imaging, optical synchronization, and edge computing functions, combined with dedicated parameter fitting and multimodal data synchronous acquisition technology. Through a standardized pollen sample preparation process, coupled with fluorescence intensity environmental adaptive calibration formulas and multispectral reflectance particle size correction calibration formulas, it effectively counteracts signal interference caused by differences in temperature, ambient light, and pollen particle size, significantly improving the accuracy and stability of the raw data. The device's portable design and rapid sample processing flow allow for detection without complex laboratory conditions, greatly shortening the detection cycle and meeting the needs of real-time field detection, providing an efficient and feasible technical solution for rapid screening of sugarcane pollen development status.

[0022] II. This invention achieves accurate identification of the developmental status of sugarcane pollen across varieties by integrating multimodal feature weighted standardization processing with an attention mechanism, combined with a dual-branch AI recognition model featuring feature enhancement. The model uses a cross-variety domain adaptive loss function as its optimization objective and is trained with a large number of cross-variety labeled samples, effectively reducing interference from feature differences between different sugarcane varieties and improving the model's adaptability to pollen from different varieties. Simultaneously, the combination of environmental adaptive adjustment technology and equipment protection design further enhances the method's resistance to interference in complex field environments, and confidence level verification ensures the reliability of the identification results. Compared to traditional manual observation or single-modal detection techniques, this invention offers higher accuracy and stronger adaptability, accurately distinguishing pollen at different developmental stages and in aborted states, providing scientific data support for pollen viability identification in sugarcane breeding, and contributing to the cultivation and promotion of superior sugarcane varieties.

[0023] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0025] Figure 1 A flowchart of a method for identifying the developmental status of sugarcane pollen by combining multispectral imaging with AI recognition;

[0026] Figure 2 A framework diagram for identifying the developmental status of sugarcane pollen using multispectral imaging combined with AI recognition. Detailed Implementation

[0027] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0028] Example 1: Real-time monitoring of flowering period in sugarcane breeding base fields

[0029] This embodiment was applied to the spring sugarcane flowering period at a sugarcane breeding base, specifically from March to April. Real-time monitoring of pollen development status was conducted on two main varieties, Guitang 42 and Xintaitang 22, with the aim of selecting highly active mature pollen for hybridization pollination on the same day. The operational steps are as follows: Figure 1 As shown.

[0030] S1: Equipment Setup and Sample Preparation

[0031] A portable device integrating fluorescence imaging, multispectral imaging, optical synchronization, and edge computing was constructed, weighing no more than 5 kg for easy carrying by breeders in the field. Phosphate buffer, a mixed staining agent of FDA and PI, and agarose fixative were prepared, with the mixed staining agent mixed according to a preset volume ratio of FDA and PI, and the phosphate buffer adjusted to a suitable pH range. Fresh anthers were collected from the flower spikes at the middle of the main stem of sugarcane plants of the two varieties in full bloom. The anther shells were removed using an anther peeling tool to obtain dispersed pollen grains. The pollen grains were evenly spread on a glass slide, covered with agarose fixative to form a thin film, and stained with the mixed staining agent under a dark environment. After staining, the slides were rinsed 2-3 times with phosphate buffer to remove excess staining agent, completing the pollen sample preparation. Simultaneously, pollen samples with known viability from both varieties (specifically, pollen samples with an in vitro germination rate of no less than 90%) were collected, and parameters specific to the sugarcane pollen of these two varieties were obtained through equipment detection and fitting.

[0032] S2: Multimodal data acquisition

[0033] Place the prepared pollen sample on the stage of the device and adjust the stage height so that the sample is centered in the imaging field of view. Activate the device's multimodal acquisition function to simultaneously acquire fluorescence images and multispectral images, as well as the background fluorescence intensity of the pollen-free area surrounding the sample and the dark-field grayscale value under no light source illumination. Record the actual field temperature using the device's built-in sensors; the temperature range for the day's testing period is between 22℃ and 28℃. Simultaneously record the ambient light intensity, which is the outdoor light intensity on a sunny day and fluctuates over time. The device automatically stores the original fluorescence intensity and original multispectral grayscale value data.

[0034] S3: Signal Calibration and Feature Fusion

[0035] The device's edge computing module calls the fluorescence intensity environmental adaptive calibration formula. Based on the original fluorescence intensity, background fluorescence intensity, actual detection temperature, and specific parameters collected by S2, it calibrates the fluorescence signal to offset the interference of field temperature fluctuations and ambient light changes on the fluorescence signal. The fluorescence intensity environmental adaptive calibration calculation formula is as follows: ,in, The fluorescence intensity after calibration; Original fluorescence intensity; The average fluorescence intensity of 100% active pollen from 10 core sugarcane varieties; This is a temperature correction factor; This is the difference between the actual measured temperature and 25℃. Background fluorescence intensity; It is an ambient light inhibitor; The ambient light intensity is used. Simultaneously, the multispectral reflectance particle size correction calibration formula is applied, combining the original multispectral grayscale values, dark-field grayscale values, and data collected from the standard gray card to calibrate the multispectral signal, correcting the reflectance deviation caused by pollen particle size differences. The multispectral reflectance particle size correction calibration calculation formula is as follows: ,in, The reflectivity after calibration for the i-th band; Multispectral wavelengths; This represents the original grayscale value of the i-th band; The grayscale value of the dark field in the i-th band; The grayscale value of the standard gray card in the i-th band; The nominal reflectance of the i-th band of the standard gray card; This is the particle size correction factor; This is the current pollen equivalent diameter; For reference pollen grain size, after calibration, the device automatically extracts the fluorescence characteristics of the pollen, including fluorescence intensity distribution and fluorescence lifetime-related characteristics; spectral characteristics, including reflectance peaks and inter-band reflectance ratios in different wavelength bands; and morphological characteristics, including pollen grain roundness and area. Subsequently, a multimodal feature weighted normalization formula is used to normalize the extracted 14 features, which include 3 fluorescence features, 7 spectral features, and 4 morphological features, thereby eliminating the influence of differences in the dimensions of different features. The multimodal feature weighted normalization calculation formula is as follows: ; ,in, The standardized result for the j-th feature; This is the j-th original feature value; Let be the sample mean of the j-th feature; Let be the sample standard deviation of the j-th feature; The weight coefficient for the j-th feature; Let be the signal-to-noise ratio of the j-th feature; Let be the correlation coefficient between the j-th feature and the pollen development status; For summation index; For the first The signal-to-noise ratio of each feature; For the first The correlation coefficients between each feature and pollen development status were then used. Finally, the attention mechanism multimodal feature fusion formula was applied to fuse the standardized features, generating a fused feature vector. The attention mechanism multimodal feature fusion calculation formula is as follows: ; ,in, To fuse feature vectors; Let be the attention weight for the j-th feature; Attention activation coefficient; is the mean of the j-th feature of the aborted pollen samples; γ is the smoothing factor; The weight coefficient for the j-th feature; This is the standardized result for the j-th feature.

[0036] S4: AI Model Recognition and Reasoning

[0037] The device utilizes a dual-branch AI recognition model with feature enhancement capabilities. This model has been trained using pollen-labeled samples from 10 sugarcane varieties, including Guitang 42 and Xintaitang 22, covering three developmental stages: mature, immature, and aborted. The training process optimizes the model using a cross-variety domain adaptive loss function to ensure its suitability for different pollen varieties. The cross-variety domain adaptive loss function is as follows: ,in, This is the total loss function; For classification loss; The domain adversarial loss weights are determined through model validation set tuning. Losses due to domain confrontation; This is the penalty coefficient for variety differences; The variance loss in accuracy is calculated between varieties. The fused feature vector generated by S3 is input into the model. The model performs feature enhancement and category differentiation through two branches, and outputs the preliminary identification results of pollen development status, including three categories: mature and highly active, immature, and aborted. At the same time, the confidence scores of the corresponding identification results are also output.

[0038] S5: Environment Adaptation and Result Output

[0039] The equipment updates field ambient light intensity and temperature data in real time. When the ambient light intensity exceeds a set threshold, it automatically calls the fluorescence intensity environmental adaptive calibration formula and the multispectral reflectance particle size correction calibration formula in S3 to recalibrate subsequent acquired signals. Simultaneously, the sealed outer casing and dustproof structure of the platform prevent interference from field dust and dew. The initial identification results output by S4 are validated for confidence level, retaining only those with a confidence level of at least 0.8, and a final detection report is generated. The report includes the sample variety, collection time, imaging parameters, identification results, and confidence level. Breeders use the report to select mature, highly active pollen for immediate use in sugarcane hybridization pollination operations that day, improving pollination success rates.

[0040] In summary, this embodiment addresses the real-time monitoring of flowering in sugarcane breeding bases. It utilizes a portable and multifunctional detection device, follows standardized procedures for pollen sample preparation, simultaneously collects multimodal data, and incorporates adaptive calibration formulas for fluorescence intensity and multispectral reflectance particle size correction to eliminate interference from environmental and pollen variations. The data is then processed using a multimodal feature weighting standardization formula and an attention mechanism multimodal feature fusion formula to generate a fused feature vector. Finally, a dual-branch AI model adapted to multiple varieties outputs high-confidence recognition results. The entire process requires no complex laboratory conditions, enabling rapid and accurate detection of pollen development in the field. This effectively selects mature, highly active pollen for same-day hybridization and pollination, significantly improving the success rate of sugarcane hybridization and meeting the real-time flowering monitoring needs of breeding bases.

[0041] Example 2: Laboratory monitoring of pollen development dynamics across multiple varieties

[0042] This embodiment was applied in a sugarcane research institute laboratory to dynamically monitor the entire pollen development cycle of 10 new sugarcane varieties to be screened, numbered Y1 to Y10. The entire pollen development cycle, from the initial opening of the inflorescence to full bloom and then to its decline, lasts for 7 days. The purpose of the monitoring is to analyze the pollen development patterns of different varieties and to screen superior varieties with longer pollen development cycles and longer periods of high pollen activity. The operation steps are as follows: Figure 2 As shown.

[0043] S1: Equipment Setup and Sample Preparation

[0044] A detection platform integrating fluorescence imaging, multispectral imaging, optical synchronization, and edge computing was built in the laboratory. Compared to field equipment, this platform adds an automatic sample transfer module to support batch testing. Phosphate buffer, FDA and PI mixed staining agent, and agarose fixative were prepared according to the same formula as in Example 1. Anthers from the middle of the flower spikes were collected from sugarcane plants of 10 new varieties at the same time every day at 9:00 AM, with 3 replicate samples collected from each variety. Pollen samples for each variety were prepared in batches according to the procedure of anther peeling, pollen separation, agarose film fixation, staining in the dark, and rinsing. At the same time, highly active pollen samples of each variety were collected during the full bloom period. The specific parameters of sugarcane pollen for each variety were obtained by detection and fitting, and stored in the parameter database of the equipment.

[0045] S2: Multimodal data acquisition

[0046] Prepared pollen samples were placed on the automated conveyor stage of the device, with three replicates per strain, totaling 30 samples. The conveying interval was set to 1 minute per sample. The device automatically conveyed each sample to the imaging area according to a preset program, simultaneously acquiring fluorescence and multispectral images, as well as the background fluorescence intensity and dark-field grayscale value for each sample. The laboratory ambient temperature was controlled at 25℃, with an allowable error of 1℃. The ambient light intensity was recorded by the device's sensors; this ambient light intensity was from a stable LED light source in the laboratory. The device automatically stored the original fluorescence intensity and original multispectral grayscale value data for each sample, and associated them with the sample strain number and collection date.

[0047] S3: Signal Calibration and Feature Fusion

[0048] The device uses an environmentally adaptive calibration formula for fluorescence intensity, combining the original fluorescence intensity, background fluorescence intensity, laboratory temperature, and specific parameters for each strain (with the laboratory temperature set at 25℃) to calibrate the fluorescence signal. It then uses a multispectral reflectance particle size correction calibration formula, based on the original multispectral grayscale values, dark-field grayscale values, and standard grayscale data, to calibrate the multispectral signal. After calibration, the device automatically extracts the fluorescence, spectral, and morphological features of each sample. A multimodal feature weighted normalization formula is used to standardize the features, and then an attention-based multimodal feature fusion formula is used to generate a fused feature vector for each sample. These vectors are then categorized and stored according to strain number and collection date.

[0049] S4: AI Model Recognition and Reasoning

[0050] The same dual-branch AI recognition model with feature enhancement as in Example 1 was invoked. This model has been imported into the parameter database of 10 new varieties and can automatically match the variety-specific parameters corresponding to the samples. The fused feature vector of each sample was input into the model, and the model outputs the pollen development status of each sample, including mature and highly active, mature and lowly active, immature, and aborted, while also outputting the corresponding confidence level. For three replicate samples of each variety, the average of the recognition results was taken as the pollen development status data of that variety on that day.

[0051] S5: Environment Adaptation and Result Output

[0052] The laboratory maintains stable ambient light intensity and temperature, and the equipment continuously monitors environmental parameters in real time. When the ambient light intensity fluctuates slightly due to light source aging, the system automatically invokes the fluorescence intensity environmental adaptive calibration formula and the multispectral reflectance particle size correction calibration formula in S3 to correct the signals of subsequent samples. The equipment's protective design effectively prevents laboratory dust from contaminating the imaging components. After each day's testing, the equipment generates a daily dynamic monitoring report, which includes the distribution of pollen development status for each strain and the proportion of highly active samples. After 7 days of monitoring, a full-cycle dynamic analysis report is generated, showing the number of days the pollen in a mature, highly active state persists from initial blooming to decay, and the daily trend of the proportion of highly active samples. Researchers use the reports to select strains that maintain a mature, highly active state for more than 4 days, such as Y3 and Y7, as superior strains for subsequent breeding.

[0053] In summary, this embodiment addresses the need for dynamic monitoring of pollen development in multiple sugarcane varieties in the laboratory. It establishes a detection platform with an automated transport module to enable batch preparation and data acquisition of pollen samples from multiple strains. A series of calibration and feature processing formulas ensure data accuracy and feature validity. A dual-branch AI model with pre-integrated multi-strain parameters is used to obtain the daily pollen development status of each strain. After a 7-day full-cycle monitoring period, a dynamic report covering the pollen development patterns of each strain is generated, successfully screening out superior strains with high and long-lasting pollen activity. This process achieves systematic monitoring of pollen development dynamics across multiple strains, avoiding the fragmented and delayed nature of traditional laboratory testing. It provides scientific data support for the screening of new sugarcane strains, contributing to the breeding and promotion of superior sugarcane varieties.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for identifying the developmental state of sugarcane pollen using multispectral imaging combined with AI recognition, characterized in that, The specific steps of this method are as follows: S1, Equipment Setup and Sample Preparation: A portable device integrating fluorescence imaging, multispectral imaging, optical synchronization and edge computing functions was set up. Phosphate buffer, FDA and PI mixed staining agent and agarose fixation solution were prepared. Pollen samples were prepared through anther peeling, pollen separation, agarose film fixation, light-protected staining and rinsing. Sugarcane pollen-specific parameters were fitted. S2, Multimodal Data Acquisition: Place the pollen sample on the stage of the device and simultaneously acquire fluorescence images, multispectral images, background fluorescence intensity, dark field gray values, record the actual detection temperature and ambient light intensity, and obtain the original fluorescence intensity and original multispectral gray values. S3, Signal Calibration and Feature Fusion: Based on the original data and specific parameters of S2, the fluorescence signal and multispectral signal are calibrated by the fluorescence intensity environment adaptive calibration formula and the multispectral reflectance particle size correction calibration formula, respectively. Fluorescence features, spectral features and morphological features are extracted. After processing by the multimodal feature weighted normalization formula, the fused feature vector is generated by the attention mechanism multimodal feature fusion formula. S4, AI model recognition and inference: Call the dual-branch artificial intelligence recognition model with feature enhancement function. The model is trained with cross-variety domain adaptive loss function as the optimization objective. Input the fused feature vector from S3 into the model and output the preliminary recognition result. S5, Environmental Adaptation and Result Output: Real-time updates of environmental parameters, dynamic invocation of the two types of calibration formulas in S3 to correct signals, combined with equipment protection design and environmental adaptive adjustment technology, after verifying the confidence level of the preliminary identification results, outputs the final detection results and corresponding detection reports.

2. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The fluorescence imaging function in step 1 is achieved through a dual-wavelength excitation component and a high-pixel imaging component. The light intensity is adjustable, and the imaging component supports high-bit format data storage. The multispectral imaging function covers the visible to near-infrared spectral range, includes multiple characteristic bands, and has a fast band switching speed.

3. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The mixed staining agent in step 1 is prepared by FDA, PI and phosphate buffer with a pH of 7.0 to 7.4, wherein the volume ratio of FDA to PI is 3:1; and the concentration of agarose fixative is 0.6% to 1.0%.

4. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The formula for calculating the adaptive calibration of fluorescence intensity in S3 is as follows: ,in, The fluorescence intensity after calibration; Original fluorescence intensity; The average fluorescence intensity of 100% active pollen from 10 core sugarcane varieties; This is a temperature correction factor; This is the difference between the actual measured temperature and 25℃. Background fluorescence intensity; It is an ambient light inhibitor; This refers to the ambient light intensity.

5. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The multispectral reflectance particle size correction calibration calculation formula in S3 is as follows: ,in, The reflectivity after calibration for the i-th band; Multispectral wavelengths; This represents the original grayscale value of the i-th band; The grayscale value of the dark field in the i-th band; The grayscale value of the standard gray card in the i-th band; The nominal reflectance of the i-th band of the standard gray card; This is the particle size correction factor; This is the current pollen equivalent diameter; For reference pollen grain size.

6. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The multimodal feature weighted normalization calculation formula in S3 is as follows: ; ,in, The standardized result for the j-th feature; This is the j-th original feature value; Let be the sample mean of the j-th feature; Let be the sample standard deviation of the j-th feature; The weight coefficient for the j-th feature; Let be the signal-to-noise ratio of the j-th feature; Let be the correlation coefficient between the j-th feature and the pollen development status; For summation index; For the first The signal-to-noise ratio of each feature; For the first The correlation coefficients between these characteristics and pollen development status.

7. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The multimodal feature fusion calculation formula for the attention mechanism in S3 is as follows: ; ,in, To fuse feature vectors; Let be the attention weight for the j-th feature; Attention activation coefficient; is the mean of the j-th feature of the aborted pollen samples; γ is the smoothing factor; The weight coefficient for the j-th feature; This is the standardized result for the j-th feature.

8. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The cross-variety domain adaptive loss function in S4 is: ,in, This is the total loss function; For classification loss; The domain adversarial loss weights are determined through model validation set tuning. Losses due to domain confrontation; This is the penalty coefficient for variety differences; This represents the variance loss in accuracy across varieties.

9. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The two branches of the dual-branch AI recognition model with feature enhancement function in S4 correspond to the category distinction and feature enhancement of multimodal fusion features, respectively. The feature enhancement function is realized through the attention mechanism. The model training uses cross-variety pollen labeled samples to construct a training set, covering samples of different developmental stages and abortion states.

10. The method for identifying the developmental state of sugarcane pollen by combining multispectral imaging with AI recognition according to claim 1, characterized in that, The environmental adaptive adjustment technology in S5 uses real-time detection data from ambient light and temperature sensors to dynamically adjust imaging-related parameters and signal calibration strategies to adapt to changes in the field environment. The confidence verification threshold is set to 0.7 to 0.95, and the detection report includes sample information, imaging parameters, and recognition results.

Images