A computer vision-based early identification method for crop diseases

Abstract

The application discloses a kind of crop disease early identification method based on computer vision, belong to image recognition technical field, specifically include: acquisition crop leaf under multiple continuous excitation period under the delayed fluorescence time series image;Extract the fluorescence intensity value of each grid unit in each cycle and construct response curve;Each curve is numerically fitted, and fluorescence kinetic characteristic vector is extracted;The characteristic vector of all grid units in the same cycle is arranged according to spatial position, and feature distribution diagram is formed;The trend of change of the characteristic value of each spatial position in the continuous cycle is tracked, and fluorescence kinetic abnormal point is identified;Abnormal points are spatially clustered and merged into abnormal regions, the average rate of change of the characteristic intensity of each region and the average rate of area expansion are calculated and combined into kinetic evolution characteristic vector;Evolution characteristic vector is matched with pre-constructed disease response mode library, and the corresponding disease type is output.The application realizes the non-invasive early identification of disease incubation period.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method for early identification of crop diseases based on computer vision. Background Technology

[0002] Crop diseases are one of the main factors restricting agricultural production, and the resulting yield reductions each year cause huge economic losses to agriculture. Timely and accurate disease detection is a key prerequisite for achieving precision pesticide application and reducing losses. Traditional disease monitoring mainly relies on manual field inspections and expert judgment. This method is not only inefficient and has limited coverage, but it also makes it difficult to detect diseases in the early stages of outbreaks, often missing the best opportunity for control.

[0003] With the development of machine vision and image processing technologies, automatic crop disease identification methods based on computer vision have gradually become a research hotspot. Existing methods can be mainly divided into two categories: one is based on visible light images, which identifies diseases by analyzing color changes, lesion morphology, or texture features on the leaf surface; the other is based on multispectral or hyperspectral imaging, which identifies disease-related physiological stress characteristics by capturing changes in the reflectance of crops in specific wavelength bands. Both of these methods rely on the formation of visually or optically discernible symptoms on the leaf surface before effective identification can be achieved.

[0004] However, there is usually an incubation period between pathogen infection and symptom appearance in crops. During this period, the pathogen has already colonized and spread within the plant, but no visible lesions have yet appeared on the leaves. Existing methods are ineffective in identifying diseases during this incubation period. By the time symptoms appear, the pathogen has often completed infection and begun sporulation and spread, at which point control measures are much less effective. Therefore, shifting the disease identification window forward to before symptom appearance is crucial for achieving true early warning.

[0005] In recent years, researchers have attempted to use chlorophyll fluorescence imaging technology to detect early stress responses in crops. Chlorophyll fluorescence is closely related to plant photosynthetic activity. When plants are infected by diseases, the function of the photosynthetic system changes, thus affecting fluorescence emission characteristics. Existing fluorescence detection methods are mainly divided into two categories: one is to measure steady-state fluorescence intensity, and the other is to measure fluorescence decay kinetic parameters after excitation. However, these methods usually only measure single leaf sites or only extract global statistical features of fluorescence intensity, failing to reflect the spatial distribution and dynamic evolution of subtle physiological changes formed in local areas of leaves during the early stages of disease infection. Furthermore, different types of pathogens (such as fungi, bacteria, and viruses) induce different plant defense response mechanisms, resulting in different response patterns in fluorescence kinetic parameters. Existing methods have failed to establish a correlation between these response patterns and disease types, making it difficult to practically achieve early identification of different types of diseases. Summary of the Invention

[0006] The purpose of this invention is to provide a method for early identification of crop diseases based on computer vision, and to solve the following technical problems: Currently, there is a lack of a method to identify different types of diseases in their early stages by capturing the spatiotemporal evolution characteristics of fluorescence dynamic parameters in local leaf areas before disease symptoms appear.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for early identification of crop diseases based on computer vision includes the following steps: S1, acquire delayed fluorescence time-series images of the target crop leaves under multiple consecutive excitation cycles, each excitation cycle including a pulse excitation and a subsequent fluorescence decay process; S2, divide the delayed fluorescence time-series image into leaf region grids, extract the fluorescence intensity value of each grid unit in each excitation cycle, and construct a fluorescence intensity response curve with the excitation cycle number as the independent variable; S3, numerically fit each fluorescence intensity response curve, extract the average slope of the rising phase, the fluctuation amplitude of the peak region and the fluctuation frequency of the decay phase, and combine them into the fluorescence dynamics feature vector of the corresponding grid cell. S4. Arrange the fluorescence dynamics characteristic vectors of all grid units within the same excitation cycle according to their spatial positions to form the excitation response rate distribution map, peak stability distribution map, and decay oscillation characteristic distribution map, respectively. S5, track the trend of feature value changes of each spatial location in continuous excitation cycles, and mark spatial locations where the feature value changes monotonically and the change amplitude exceeds the preset threshold as fluorescence dynamics abnormal points. S6, perform spatial clustering on the fluorescence dynamics anomalies, merge adjacent anomalies into anomaly regions, calculate the average rate of change of feature values ​​and the rate of expansion of region area for each anomaly region, and combine them into a dynamic evolution feature vector of the anomaly region. S7, match the dynamic evolution feature vector with the standard vector in the disease response pattern library, and output the corresponding disease type when the matching degree exceeds the preset threshold.

[0008] As a further aspect of the present invention: the specific acquisition process of the delayed fluorescence time-series image in S1 is as follows: In a darkroom environment, a high-sensitivity camera and a pulsed excitation light source are fixed on a movable bracket, with the camera's optical axis aligned with the canopy of the crop to be tested, and the illumination range of the pulsed excitation light source covering the entire leaf area within the camera's field of view. The pulse excitation light source is set to repeatedly emit pulse light at a preset fixed time interval. The duration and intensity of each pulse light remain constant, and the time interval between adjacent pulses is greater than the time required for the fluorescence signal generated by the previous pulse excitation to completely decay below the camera's light sensitivity threshold. At the moment each pulse light emission ends, a high-sensitivity camera is activated to continuously acquire images of the blade surface within the current field of view. The camera captures images at a preset fixed frame rate until the intensity of the fluorescence signal generated by the current pulse excitation decays below the camera's photosensitive threshold, at which point acquisition stops. All images acquired in this acquisition are arranged in chronological order of capture time as a delayed fluorescence image sequence for that excitation cycle. After completing the acquisition of the current field of view, the movable support is moved to the next crop canopy location to be tested, and the pulse excitation and image acquisition process is repeated until the data acquisition of all preset acquisition points is completed. The delayed fluorescence image sequences obtained in all excitation cycles are arranged in the order of acquisition point location and excitation cycle to form a delayed fluorescence time series image set covering multiple acquisition points and multiple consecutive excitation cycles. Each frame in this delayed fluorescence time series image set is marked with its acquisition point identifier, excitation cycle number and shooting time number within that cycle.

[0009] As a further aspect of the present invention: in S3, the specific generation process of the fluorescence dynamics feature vector is as follows: Numerical fitting was performed on each fluorescence intensity response curve. The fitting parameters were calculated using a nonlinear regression model that included an exponential decay term. Based on the fitting parameters, the average rate of change of the curve from the starting point of the rising phase to the first peak point was calculated. This average rate of change was defined as the excitation response rate value of the grid cell. In the peak region of the fluorescence intensity response curve, the fluorescence intensity values ​​corresponding to the peak point and several consecutive moments before and after it are selected to form a local sequence. The standard deviation of all fluorescence intensity values ​​in the local sequence is calculated, and the standard deviation is defined as the peak fluctuation amplitude of the grid cell. Spectral analysis is performed on the decay phase of the fluorescence intensity response curve. The decay segment data from the peak point to the end point of the curve is converted into a frequency domain signal through Fourier transform. The frequency component corresponding to the maximum amplitude in the frequency domain signal is extracted and the frequency component is defined as the decay oscillation frequency value of the grid cell. The excitation response rate value, peak fluctuation amplitude value, and decay oscillation frequency value are combined in a preset order to generate a three-dimensional numerical vector as the fluorescence dynamics feature vector of the grid cell.

[0010] As a further aspect of the present invention: the specific process of forming the excitation response rate distribution map, the peak stability distribution map, and the damped oscillation characteristic distribution map in step S4 is as follows: Extract the fluorescence dynamics feature vectors of all grid cells within the same excitation cycle, extract the excitation response rate value from each feature vector, and fill it into the same row and column positions of a two-dimensional matrix according to the row and column numbers of the grid cell in the blade region. The number of rows of the two-dimensional matrix is ​​equal to the number of grid cells in the vertical direction of the blade region, and the number of columns is equal to the number of grid cells in the horizontal direction. The filled two-dimensional matrix is ​​defined as the excitation response rate distribution map of the excitation cycle. The peak fluctuation amplitude values ​​of the fluorescence dynamics eigenvectors of all grid cells within the same excitation cycle are extracted and filled into the corresponding positions of another two-dimensional matrix of the same size according to the same row and column numbers of each grid cell. The filled two-dimensional matrix is ​​defined as the peak stability distribution map of the excitation cycle. Take the decay oscillation frequency values ​​from the fluorescence dynamics eigenvectors of all grid cells within the same excitation period, and fill them into the corresponding positions of a third two-dimensional matrix of the same size according to the same row and column numbers of each grid cell. Define the filled two-dimensional matrix as the decay oscillation characteristic distribution map of the excitation period.

[0011] As a further aspect of the present invention: the specific identification process of fluorescence dynamics anomalies in S5 is as follows: For the excitation response rate distribution map, peak stability distribution map, and damped oscillation characteristic distribution map, a sequence of the same type of distribution map under multiple consecutive excitation cycles is obtained. Each distribution map in the sequence is aligned according to the same spatial coordinate system so that the same row number and column number position in the distribution map of different cycles corresponds to the same spatial position of the blade region. For each spatial location, feature values ​​are extracted from the excitation response rate distribution sequence, peak stability distribution sequence, and decay oscillation feature distribution sequence to form a time series of the location on the three features. The difference between adjacent periodic feature values ​​in each time series is calculated. When at least two consecutive periods of positive difference and absolute value exceed a preset threshold appear in the time series of any feature, the feature is determined to have a continuous upward trend. When at least two consecutive periods of negative difference and absolute value exceed a preset threshold appear, the feature is determined to have a continuous downward trend. Spatial locations where any feature shows a continuous upward or downward trend are marked as fluorescence kinetic anomalies, and the feature type corresponding to the anomaly is recorded.

[0012] As a further aspect of the present invention: in step S6, the specific process for obtaining the dynamic evolution feature vector of the abnormal region is as follows: Spatial clustering based on Euclidean distance is performed on the spatial coordinates of all fluorescence dynamics anomalies. Anomalies whose distance to each other is less than the preset neighborhood radius are grouped into the same anomaly region. Each anomaly region contains several anomalies and their corresponding feature types. For each abnormal region, the number of abnormal points in the region belonging to the excitation response rate abnormality type, peak stability abnormality type, and decay oscillation characteristic abnormality type is counted, and the abnormality type with the largest number is determined as the dominant abnormality type of the abnormal region. Extract the feature values ​​of all anomalies in the anomaly region at their respective marked times, calculate the arithmetic mean of these feature values ​​as the mean feature intensity of the anomaly region, extract the time interval between the earliest and latest marked times in the anomaly region, and divide the mean feature intensity by the time interval to obtain the average rate of change of feature intensity of the anomaly region. Obtain the spatial boundary of the anomalous region under continuous excitation cycles, calculate the number of grid cells covered by the anomalous region in each cycle as the region area of ​​that cycle, divide the difference in region area between adjacent cycles by the region area of ​​the previous cycle to obtain the area change rate of that cycle, and take the arithmetic mean of the area change rates of all cycles to obtain the average rate of area expansion of the anomalous region. The average rate of change of the characteristic intensity and the average rate of area expansion are combined in a preset order to generate a two-dimensional numerical vector as the dynamic evolution characteristic vector of the abnormal region.

[0013] As a further aspect of the present invention: the specific construction process of the disease response pattern library in S7 is as follows: Multiple groups of healthy plants were selected and inoculated with a single type of known disease. After inoculation, each plant was placed in a dark room and steps S1 to S6 were performed to obtain the dynamic evolution feature vector of each abnormal region on each inoculated plant. At the same time, the disease type label corresponding to each abnormal region was recorded. Each dynamic evolution feature vector and its corresponding disease type label were combined into a sample, and all samples constituted the training dataset. Each dynamic evolution feature vector is used as network input, and the corresponding disease type label is converted into a one-hot encoding as the training target. The neural network parameters are initialized, and samples from the training dataset are input into the neural network in batches for forward propagation. The predicted probability distribution is obtained by normalizing the exponential function at the output layer. The cross-entropy loss between the predicted probability distribution and the true one-hot encoded label is calculated. The gradient of the loss with respect to the neural network parameters is calculated using the backpropagation algorithm, and the neural network parameters are updated using the gradient descent algorithm. The above process is repeated until the classification accuracy of the neural network on the validation set tends to stabilize. The neural network parameters at this time are saved as a disease response pattern library.

[0014] As a further aspect of the present invention, if multiple mutually separated anomalous regions appear on the same blade, then the dynamic evolution feature vector of each anomalous region is extracted, wherein the dynamic evolution feature vector is composed of the average rate of change of the feature intensity and the average rate of area expansion of the region. Calculate the Euclidean distance between the dynamic evolution feature vectors of any two anomalous regions, and mark two anomalous regions whose Euclidean distance is less than a preset similarity threshold as a pair of similar anomalous regions; For each pair of similar anomalous regions, the earliest and latest labeled times of all fluorescence dynamics anomalous points in the two anomalous regions are extracted, the time interval overlap length of the labeled times of the two anomalous regions is calculated, and similar anomalous region pairs with an overlap length greater than a preset overlap threshold are marked as spatiotemporal synchronization anomalous region pairs. All anomalous regions connected by spatiotemporal synchronization anomalous regions are grouped into the same associated anomalous region group, which contains multiple spatially separated anomalous regions. Anomalies that do not form a spatiotemporally synchronized anomaly pair with any other anomalies are retained as independent anomalies.

[0015] The beneficial effects of this invention are: 1) This can be understood as a subtle change in the function of a plant's photosynthetic system even before any visible symptoms appear on the leaves after infection by pathogens. Chlorophyll fluorescence, as a natural probe of photosynthesis, can sensitively reflect the electron transport state, light conversion efficiency, and energy dissipation process of photosystem II. This invention acquires complete fluorescence decay process data by collecting delayed fluorescence time-series images of crop leaves under multiple consecutive excitation cycles in a darkroom environment. The images are gridded, and the fluorescence intensity response curve of each grid cell is extracted. Three characteristic parameters are extracted from the curves: excitation response rate, peak fluctuation amplitude, and decay oscillation frequency. These parameters correspond to the light energy capture efficiency, photoprotection regulation intensity, and oscillation characteristics of the electron transport chain, respectively. When pathogens infect, even if the infected area is small and no lesions have formed, abnormal photosynthetic function in the local tissue will cause these characteristic parameters to deviate from normal values. By capturing this early physiological change, this invention significantly advances the disease identification window from the symptom manifestation period of traditional methods to the incubation period, gaining crucial early intervention time for disease control and achieving true non-invasive early warning.

[0016] 2) Disease infection typically begins in a localized area on the leaf surface and then gradually expands. This spatial dynamic process is crucial information for early disease identification. This invention arranges the fluorescence dynamics feature vectors of all grid cells within the same excitation cycle according to their spatial location, forming an excitation response rate distribution map, a peak stability distribution map, and a decay oscillation feature distribution map, transforming microscopic physiological parameters into intuitive spatial distribution maps. By tracking the changing trends of feature values ​​at each spatial location in consecutive excitation cycles, spatial points with continuously abnormal changes in feature values ​​are identified and marked as fluorescence dynamics anomalies. These anomalies are potential targets for disease infection. Spatial clustering of anomalies is performed, merging adjacent points into anomaly regions. Simultaneously, the average rate of change of feature values ​​and the area expansion rate of each region are calculated to construct a spatiotemporal evolution feature vector that reflects the dynamics of disease development. When multiple separated anomaly regions appear on the same leaf, this invention further uses feature similarity and temporal synchronization analysis to determine whether they belong to the same infection event, thereby accurately locating the source of infection and dynamically tracking the disease expansion trajectory, providing spatial and temporal decision-making basis for precise pesticide application.

[0017] 3) Different pathogens (such as fungi, bacteria, and viruses) infect plants with fundamentally different pathogenic mechanisms and the resulting plant defense responses. Fungal infections are usually accompanied by cell wall degradation and toxin secretion, mainly affecting the integrity of the photosynthetic membrane system; bacterial infections often lead to vascular bundle blockage and water transport obstacles, thereby interfering with the supply of raw materials for photosynthesis; viral infections systematically affect photosynthetic enzyme activity by interfering with host gene expression and protein synthesis. These different physiological perturbations leave unique "fingerprint" patterns on fluorescence kinetic parameters. This invention, through the extraction of spatiotemporal evolution feature vectors of abnormal regions, not only includes the abnormal amplitude of fluorescence parameters but also integrates the rate of disease development and spatial expansion characteristics, comprehensively characterizing the physiological impact patterns of diseases. Through the pre-constructed disease response pattern library, it can be understood that this disease response pattern library consists of feature vectors and their type labels obtained by processing known single disease infection samples through the same process. The feature vectors of unknown samples are matched with standard vectors in the library. The disease type is output based on the comparison results of feature similarity, so that the identification process no longer relies on subjective symptom judgment, but is based on the inherent physiological specific response of the disease. This significantly improves the accuracy and specificity of early disease identification and effectively avoids interference from environmental factors and non-disease stress. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating a method for early identification of crop diseases based on computer vision according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, this invention is a method for early identification of crop diseases based on computer vision, comprising the following steps: S1, acquire delayed fluorescence time-series images of the target crop leaves under multiple consecutive excitation cycles, each excitation cycle including a pulse excitation and a subsequent fluorescence decay process; When plant leaves are excited by pulsed light, they emit fluorescence signals that decay over time. The intensity and decay characteristics of this signal are closely related to the functional state of the leaf's photosynthetic system. This invention collects data in a darkroom environment to avoid interference from ambient light on the weak fluorescence signal. A high-sensitivity camera and a pulsed excitation light source are fixed on a movable support, with the camera's optical axis aligned with the canopy of the crop under test. The illumination range of the pulsed excitation light source covers the entire leaf area within the camera's field of view. The pulsed excitation light source is set to repeatedly emit pulsed light at a preset fixed time interval. The duration and intensity of each pulse remain constant, and the time interval between adjacent pulses is greater than the time required for the fluorescence signal generated by the previous pulse to completely decay below the camera's sensitivity threshold, ensuring that the fluorescence signals between excitation cycles do not interfere with each other. At the moment each pulse ends, the high-sensitivity camera is activated to continuously acquire images of the leaf surface within the current field of view. The camera captures images at a preset fixed frame rate until the intensity of the fluorescence signal generated by the current pulse decays below the camera's sensitivity threshold, at which point acquisition stops. All images acquired in this acquisition are arranged in chronological order of capture time to form a delayed fluorescence image sequence for that excitation cycle. After completing the acquisition of the current field of view, move the movable support to the next crop canopy location to be tested, and repeat the pulse excitation and image acquisition process until data acquisition is completed for all preset acquisition points. Arrange the delayed fluorescence image sequences obtained from all excitation cycles in order of acquisition point location and excitation cycle sequence to form a delayed fluorescence time-series image set covering multiple acquisition points and multiple consecutive excitation cycles. Each frame in this delayed fluorescence time-series image set is labeled with its corresponding acquisition point identifier, excitation cycle number, and shooting time number within that cycle.

[0022] S2, divide the delayed fluorescence time-series image into leaf region grids, extract the fluorescence intensity value of each grid unit in each excitation cycle, and construct a fluorescence intensity response curve with the excitation cycle number as the independent variable; The obtained delayed fluorescence time-series images were divided into leaf regions using a grid. First, the pixel region containing the leaf was identified in each frame, and the background was removed. The leaf region was then uniformly divided into multiple grid cells in both the vertical and horizontal directions, with each grid cell corresponding to a small area on the leaf surface. For each grid cell, the fluorescence intensity value was extracted for each excitation cycle. Since multiple frames reflecting the fluorescence decay process are contained within the same excitation cycle, the average grayscale value of the grid cell in each frame within that cycle was taken as the fluorescence intensity value for that grid cell in that cycle. Then, a response curve reflecting the change in fluorescence intensity with the number of excitation cycles was constructed for each grid cell, with the excitation cycle number as the x-axis and the fluorescence intensity value as the y-axis. This curve reflects the dynamic changes in the functional state of the photosynthetic system in this small area under multiple consecutive excitations.

[0023] S3, numerically fit each fluorescence intensity response curve, extract the average slope of the rising phase, the fluctuation amplitude of the peak region and the fluctuation frequency of the decay phase, and combine them into the fluorescence dynamics feature vector of the corresponding grid cell. A nonlinear regression model incorporating an exponential decay term was used to fit the response curve, yielding fitting parameters that describe the curve's shape. Based on these fitting parameters, the average rate of change of the curve from the starting point of the rising phase to the first peak point was calculated. This rate reflects the speed at which the photosystem II reaction center opens and the initial response capability of the electron transport chain in this region after excitation. This average rate of change was defined as the excitation response rate value of this grid cell. In the peak region of the curve, fluorescence intensity values ​​corresponding to the peak point and several consecutive moments before and after it were selected to form a local sequence. The standard deviation of all fluorescence intensity values ​​in this local sequence was calculated. This standard deviation reflects the degree of fluctuation in fluorescence intensity near the peak, and is related to the activity of the photoprotective regulation mechanism and the stability of excitation energy distribution. This standard deviation was defined as the peak fluctuation amplitude value of this grid cell. Spectral analysis is performed on the decay phase of the curve. The decay segment data from the peak point to the end point is converted into a frequency domain signal using Fourier transform. The frequency component corresponding to the maximum amplitude in the frequency domain signal is extracted. This frequency component reflects the periodic oscillation characteristics during the fluorescence decay process and is related to the feedback regulation of the electron transport chain and the dynamic balance of energy dissipation. This frequency component is defined as the decay oscillation frequency value of the grid cell. The excitation response rate value, peak fluctuation amplitude value, and decay oscillation frequency value are combined in a preset order to generate a three-dimensional numerical vector as the fluorescence dynamics characteristic vector of the grid cell.

[0024] S4. Arrange the fluorescence dynamics characteristic vectors of all grid units within the same excitation cycle according to their spatial positions to form the excitation response rate distribution map, peak stability distribution map, and decay oscillation characteristic distribution map, respectively. The fluorescence kinetic feature vectors of all grid cells within the same excitation cycle are extracted. The excitation response rate value of each feature vector is then extracted and filled into the same row and column positions of a two-dimensional matrix according to the row and column numbers of that grid cell in the leaf region. The number of rows in this two-dimensional matrix equals the number of grid cells in the vertical direction of the leaf region, and the number of columns equals the number of grid cells in the horizontal direction. The filled two-dimensional matrix is ​​defined as the excitation response rate distribution map for that excitation cycle. Using the same method, the peak fluctuation amplitude values ​​of the fluorescence kinetic feature vectors of all grid cells within the same excitation cycle are extracted and filled into the corresponding positions of another two-dimensional matrix of the same size according to the same row and column numbers of each grid cell, forming a peak stability distribution map. The decay oscillation frequency values ​​of the fluorescence kinetic feature vectors of all grid cells within the same excitation cycle are extracted and filled into the corresponding positions of a third two-dimensional matrix of the same size according to the same row and column numbers, forming a decay oscillation feature distribution map. These three distribution maps depict the spatial distribution of photosynthetic physiological functions on the leaf surface at different times during that excitation cycle from different perspectives.

[0025] S5, track the trend of feature value changes of each spatial location in continuous excitation cycles, and mark spatial locations where the feature value changes monotonically and the change amplitude exceeds the preset threshold as fluorescence dynamics abnormal points. For excitation response rate distribution maps, peak stability distribution maps, and decay oscillation characteristic distribution maps, sequences of the same type of distribution map were obtained under multiple consecutive excitation cycles. Each distribution map in the sequence was aligned using the same spatial coordinate system, ensuring that the same row and column numbers in the distribution maps of different cycles corresponded to the same spatial location within the leaf region. For each spatial location, feature values ​​were extracted from the excitation response rate distribution map sequence, peak stability distribution map sequence, and decay oscillation characteristic distribution map sequence, constructing a time series for that location on the three features. The difference between the feature values ​​of adjacent cycles in each time series was calculated; this difference reflects the direction and magnitude of the change in the feature value at that location with the number of excitation cycles. Spatial locations in the time series of any feature where two or more consecutive differences are positive and their absolute values ​​exceed a preset threshold, or where two consecutive differences are negative and their absolute values ​​exceed a preset threshold, were marked as fluorescence kinetic anomalies, and the corresponding feature type was recorded. Continuous positive differences indicate a sustained increase in the feature value at that location, while continuous negative differences indicate a sustained decrease. This continuous monotonic trend eliminates the interference of random fluctuations and reflects a possible persistent physiological change at that location.

[0026] S6, perform spatial clustering on the fluorescence dynamics anomalies, merge adjacent anomalies into anomaly regions, calculate the average rate of change of feature values ​​and the rate of expansion of region area for each anomaly region, and combine them into a dynamic evolution feature vector of the anomaly region. Spatial clustering based on Euclidean distance was performed on the spatial coordinates of all fluorescence kinetic anomalies. Anomalies whose distance to each other was less than a preset neighborhood radius were grouped into the same anomaly region. Each anomaly region contained several anomalies and their corresponding feature types. For each anomaly region, the number of anomalies belonging to the excitation response rate anomaly, peak stability anomaly, and decay oscillation feature anomaly types within the region was counted. The anomaly type with the most occurrences was determined as the dominant anomaly type of the region, reflecting the main photosynthetic function anomaly pattern of the region. The feature values ​​of all anomalies within the region at their respective labeled times were extracted, and the arithmetic mean of these feature values ​​was calculated as the mean feature intensity of the region. The time interval between the earliest and latest labeled times within the region was extracted, and the mean feature intensity was divided by this time interval to obtain the average rate of change of feature intensity of the region, which reflects the severity of the anomaly development. The spatial boundary of the anomaly region under continuous excitation cycles was obtained, and the number of grid cells covered by the anomaly region in each cycle was calculated as the region area for that cycle. The area change rate for a given period is obtained by dividing the difference in area between adjacent periodic regions by the area of ​​the previous periodic region. The arithmetic mean of the area change rates for all periods is then taken to obtain the average rate of area expansion of the anomalous region, which reflects the speed of the anomalous range expansion. The average rate of change of characteristic intensity and the average rate of area expansion are combined in a preset order to generate a two-dimensional numerical vector as the dynamic evolution feature vector of the anomalous region. If multiple isolated anomalous regions appear on the same leaf, further processing of the relationships between these regions is required. The dynamic evolution feature vector of each anomalous region is extracted, and the Euclidean distance between any two anomalous regions' dynamic evolution feature vectors is calculated. Two anomalous regions with an Euclidean distance less than a preset similarity threshold are marked as similar anomalous region pairs. For each similar anomalous region pair, the earliest and latest labeling times of all fluorescence dynamic anomalies within both regions are extracted, and the time interval overlap length between the labeling times of the two anomalous regions is calculated. Similar anomalous region pairs with an overlap length greater than a preset overlap threshold are marked as spatiotemporally synchronized anomalous region pairs. All anomalous regions connected by spatiotemporally synchronized anomalous region pairs are grouped into the same associated anomalous region group. The anomalous regions in this group are spatially separated but similar in their characteristic evolution patterns and synchronized in their active time, collectively representing multiple infection points of the same disease type on different parts of the leaf. Anomalous regions that do not form spatiotemporally synchronized anomalous region pairs with any other anomalous regions are retained as independent anomalous regions.

[0027] S7, match the dynamic evolution feature vector with the standard vector in the disease response pattern library, and output the corresponding disease type when the matching degree exceeds the preset threshold.

[0028] Multiple groups of healthy plants were selected and inoculated with a single known disease. After inoculation, each plant was placed in a darkroom environment and steps S1 to S6 were performed to obtain the dynamic evolution feature vectors of each abnormal region or associated abnormal region group on each inoculated plant. Simultaneously, the disease type label corresponding to each abnormal region or associated abnormal region group was recorded. Each dynamic evolution feature vector and its corresponding disease type label were combined into a sample, and all samples constituted the training dataset. Each dynamic evolution feature vector was used as the network input, and the corresponding disease type label was converted into a one-hot encoding as the training target. The neural network parameters were initialized, and samples from the training dataset were input into the neural network in batches for forward propagation. The predicted probability distribution was obtained at the output layer using a normalized exponential function. The cross-entropy loss between the predicted probability distribution and the true one-hot encoded label was calculated. The gradient of the loss with respect to the neural network parameters was calculated using the backpropagation algorithm, and the neural network parameters were updated using the gradient descent algorithm. The above process was repeated until the classification accuracy of the neural network on the validation set stabilized. The neural network parameters at this point were saved as a disease response pattern library. In actual detection, the dynamic evolution feature vectors of each associated abnormal region group and each independent abnormal region obtained through steps S1 to S6 are input into the disease response pattern library. The predicted probability distribution is calculated through neural network forward propagation, and the disease type corresponding to the maximum probability is selected as the identification result output.

[0029] In a preferred embodiment of the present invention, the specific acquisition process of the delayed fluorescence time-series image in step S1 is as follows: A high-sensitivity camera and a pulsed excitation light source are mounted together on a movable stand. The height and angle of the stand are adjusted so that the camera's optical axis is perpendicularly aligned with the canopy of the crop under test. Simultaneously, the direction of the light source is adjusted to cover the entire leaf area within the camera's field of view, ensuring that each leaf is uniformly excited and clearly imaged. Based on the typical duration of leaf fluorescence decay, the pulsed excitation light source is set to repeatedly emit pulses at fixed time intervals, such as each pulse lasting several milliseconds, with the intensity stable at a preset value. The interval between adjacent pulses is sufficiently long to ensure that the fluorescence signal generated by the previous excitation completely decays below the camera's sensitivity threshold before the next excitation begins, avoiding signal overlap between cycles. At the instant each pulse emission ends, a synchronous trigger signal activates the camera to begin continuous acquisition. The camera captures images at a fixed frame rate at high speed, continuously recording the dynamic decay process of fluorescence intensity on the leaf surface until the signal intensity decays to an undetectable level. All images captured during this period are arranged chronologically to form a delayed fluorescence image sequence for that excitation cycle. After acquiring data for one field of view, the movable stand is moved to the next canopy location to be tested, and the above pulse excitation and image acquisition process is repeated until data from all preset acquisition points has been obtained. Finally, all image sequences obtained from the excitation cycles were organized and labeled according to the acquisition point location, excitation cycle number, and acquisition time number within the cycle, forming a structured delayed fluorescence time series image set. Each frame of the image is accompanied by a clear acquisition point identifier, cycle number, and time number, providing an accurate data index for subsequent spatial alignment and temporal analysis.

[0030] By completely eliminating the interference of ambient light on weak fluorescence signals in a darkroom environment, the collected fluorescence intensity accurately reflects the excited-state decay characteristics of the leaf photosynthetic system. The use of a movable support allows for efficient sequential detection of multiple plants or leaves under the same equipment, significantly increasing data throughput. Acquisition at fixed time intervals and throughout the complete decay process ensures the independence and integrity of data from each excitation cycle, enabling the extracted fluorescence response curves to accurately depict the dynamic changes in photosynthetic function under multiple excitations. Multiple labels attached to the images provide a foundation for precise spatial alignment and temporal series construction of images from different periods and locations, ensuring reliable spatiotemporal correspondence in fluorescence dynamics analysis from the pixel level to the region level. Ultimately, these refined acquisition operations lay a high-quality data foundation for extracting physiologically significant feature parameters from fluorescence time-series images, enabling this invention to capture weak anomalies in photosynthetic function in localized areas of leaves during the early stages of disease infection, thereby achieving early identification of the latent period.

[0031] In another preferred embodiment of the present invention, the specific process for generating the fluorescence dynamics feature vector in step S3 is as follows: First, numerical fitting is performed on each fluorescence intensity response curve. Taking a specific curve as an example, this curve records the fluorescence intensity change of a certain grid cell under multiple consecutive excitation cycles. The horizontal axis represents the excitation cycle number, and the vertical axis represents the corresponding fluorescence intensity value. A nonlinear regression model including an exponential decay term is used to fit this curve. The fitting parameters in the model are determined through an iterative optimization algorithm to minimize the error between the fitted curve and the actual data points. Based on these fitting parameters, the overall shape and change law of the curve can be accurately described. On this basis, the first peak point of the curve is located from the fitting results, that is, the turning point where the fluorescence intensity changes from rising to falling. The average rate of change from the starting point of the rising phase of the curve to the peak point is calculated. Specifically, the fluorescence intensity value at the peak point is subtracted from the fluorescence intensity value at the starting point, and then divided by the number of excitation cycles between the two to obtain the average increase in fluorescence intensity per unit cycle. This value is defined as the excitation response rate value of the grid cell. Subsequently, in the peak region of the fluorescence intensity response curve, the peak point itself and several consecutive moments before and after it are selected. Fluorescence intensity values, for example, are taken from five points—two periods before and two periods after the peak—to form a local sequence. The standard deviation of all fluorescence intensity values ​​in this sequence is calculated. This standard deviation reflects the dispersion of fluorescence intensity near the peak, i.e., the magnitude of the fluctuation. This standard deviation is defined as the peak fluctuation amplitude of this grid cell. Next, the decay phase of the fluorescence intensity response curve is analyzed. Data from the decay segment after the peak point to the end of the curve is extracted. This segment of data is transformed from the time domain to the frequency domain using Fourier transform to obtain the amplitude distribution corresponding to different frequency components. The frequency component corresponding to the maximum amplitude value is found in the frequency domain signal. This frequency represents the most important oscillation period during the decay process. This frequency component is defined as the decay oscillation frequency of this grid cell. Finally, the excitation response rate value, peak fluctuation amplitude value, and decay oscillation frequency value are combined in a pre-set order to form a vector containing three values, which is the fluorescence dynamics characteristic vector of this grid cell. This vector comprehensively characterizes the response speed, stability, and oscillation characteristics of the photosynthetic system in this small region under multiple excitations.

[0032] By mathematically modeling and extracting features from the original fluorescence response curves, complex time-series signals are transformed into parameter combinations with clear physiological significance, enabling subsequent analysis to focus on key indicators closely related to photosynthetic functional status. It can be understood that the excitation response rate reflects the efficiency of the photosystem II reaction center opening and the initial response capability of the electron transport chain; this rate often changes when pathogen infection impairs local tissue photosynthetic function. The peak fluctuation amplitude is related to the activity of the photoprotective regulatory mechanism; pathogen infection may interfere with the stability of the energy dissipation process, leading to abnormal fluctuation amplitudes. The decay oscillation frequency is related to the feedback regulation of the electron transport chain; different pathogens cause varying degrees of change in photosynthetic function. Physiological disorders may exhibit specific differences in oscillation characteristics. By combining these three parameters into a feature vector, the multifaceted functional information of the photosynthetic system is preserved, while data dimensionality reduction and structuring are achieved. This allows for the quantitative comparison of the state changes of the same grid cell under different excitation cycles, and provides standardized input for subsequent identification of abnormal regions from the perspective of spatial distribution and temporal evolution. This feature extraction method based on photosynthetic physiological mechanisms enables the present invention to capture functional changes in local areas of leaves in the early stages of disease infection. Even when no symptoms are visible, potential infection points can be detected through abnormal fluorescence kinetic parameters, laying the foundation for accurate identification of disease types.

[0033] In another preferred embodiment of the present invention, the specific process of forming the excitation response rate distribution map, the peak stability distribution map, and the damped oscillation characteristic distribution map in step S4 is as follows: First, obtain the fluorescence dynamics feature vectors of all grid cells within a given excitation cycle. Assume the blade region within this cycle is divided into 150 grid cells (10 rows, 15 columns). Each grid cell corresponds to a three-dimensional vector containing the excitation response rate value, peak fluctuation amplitude value, and decay oscillation frequency value. For constructing the excitation response rate distribution map, iterate through all grid cells, extracting the excitation response rate value from the feature vector of each grid cell. Based on the row and column numbers of the grid cell within the blade region, fill this value into the corresponding position in a pre-established blank two-dimensional matrix. For example, the excitation response rate value of the grid cell located in the third row and fifth column is filled into the third row and fifth column position of this matrix. After all grid cells' excitation response rate values ​​are filled in this way, each position in this two-dimensional matrix corresponds to a specific... The values ​​are calculated such that the number of rows equals the number of grid cells in the vertical direction (ten rows) and the number of columns equals the number of grid cells in the horizontal direction (fifteen columns). This filled two-dimensional matrix constitutes the excitation response rate distribution map for that excitation cycle. The peak fluctuation amplitude value is processed in the same way. All grid cells are traversed again, and the peak fluctuation amplitude value in the characteristic vector of each cell is extracted and filled into another blank two-dimensional matrix of the same size according to the same row and column number rules, to obtain the peak stability distribution map for that cycle. The same operation is then performed on the damped oscillation frequency value. The damped oscillation frequency value of each cell is filled into a third blank two-dimensional matrix according to the row and column positions, forming the damped oscillation characteristic distribution map for that cycle. At this point, all three distribution maps for the same excitation cycle are completed, and they have the same matrix size and spatial correspondence.

[0034] The discrete grid cell feature values ​​extracted from the original image are reorganized into a spatially continuous two-dimensional matrix, enabling the recovery of the true spatial position of the originally isolated data points on the leaf surface. This transforms the microscopic fluorescence dynamics parameters into a macroscopically visible spatial distribution image of physiological functions. By constructing distribution maps with three different parameters, the information of the photosynthetic system's response rate, stability, and oscillation characteristics is separated, facilitating subsequent independent analysis of its spatial distribution patterns and evolution laws for each dimension. This spatialization process reveals spatial clustering features that were originally hidden in a one-dimensional numerical list. For example, the excitation response rate values ​​in a certain region may be generally high or low, or the peak fluctuation amplitudes of multiple adjacent grid cells may show similar trends. These spatial correlations can only be effectively identified after restoring the positional information. Furthermore, the three distribution maps use identical grid divisions and row-column correspondences, ensuring precise spatial alignment between different parameters and different periods. This provides a unified spatial benchmark for tracking the changes of feature values ​​at each spatial location over time and identifying persistently abnormal local areas. Ultimately, this invention enables the comprehensive capture of physiological abnormal signals in the early stages of disease infection from both spatial and temporal dimensions.

[0035] In another preferred embodiment of the present invention, the specific identification process of fluorescence dynamics anomalies in step S5 is as follows: First, for the excitation response rate distribution map, peak stability distribution map, and damped oscillation characteristic distribution map, we obtain their distribution map sequences over multiple consecutive excitation cycles. For example, we obtain ten excitation response rate distribution maps from the first to the tenth cycle, and similarly, ten peak stability distribution maps and ten damped oscillation characteristic distribution maps for these ten cycles. We then align these thirty distribution maps using the same spatial coordinate system. Since all distribution maps use the same grid division method (i.e., each map has the same number of rows and columns), we only need to ensure that the positions of the same row and column numbers in the same type of distribution maps across different cycles correspond to the blade table. The same physical region on the surface, for example, the grid cell in the fifth row and eighth column of the excitation response rate distribution diagram of the first period and the grid cell in the fifth row and eighth column of the excitation response rate distribution diagram of the tenth period represent the same point on the blade; for each spatial location on the blade, such as the grid cell in the fifth row and eighth column, the excitation response rate values ​​of that location from the first period to the tenth period are extracted from the excitation response rate distribution diagram sequence to form the excitation response rate time series of that location. Similarly, the peak fluctuation amplitude value of that location is extracted from the peak stability distribution diagram sequence to form the peak stability time series, and the peak fluctuation amplitude value of that location is extracted from the damped oscillation characteristic distribution diagram sequence. The decaying oscillation frequency values ​​constitute a decaying oscillation frequency time series. For each time series, the difference between adjacent periods is calculated sequentially. Taking the excitation response rate time series as an example, the first difference is obtained by subtracting the value of the first period from the value of the second period, the second difference is obtained by subtracting the value of the second period from the value of the third period, and so on to obtain nine differences. When two or more consecutive differences are positive and the absolute values ​​of these differences exceed a preset threshold in the time series of a certain feature, for example, three consecutive differences are positive and each difference is greater than 0.15, it indicates that the feature value at that position has been continuously enhanced in multiple consecutive periods, or that two or more consecutive differences have appeared. When all values ​​are negative and their absolute values ​​exceed a preset threshold, for example, two consecutive negative differences with absolute values ​​greater than 0.12, it indicates that the feature value at that location is continuously weakening. For the spatial location in the fifth row and eighth column, check whether the time series of its three features meet the above conditions. If the time series of any one of the features—excitation response rate, peak fluctuation amplitude, or decay oscillation frequency—has two or more consecutive positive or negative differences that exceed the threshold, mark the spatial location as a fluorescence kinetic anomaly point. At the same time, record which feature has shown this continuous change, for example, record it as "excitation response rate anomaly point" or "peak stability anomaly point".

[0036] By tracking the trend of characteristic value changes at each spatial location during continuous excitation cycles, the persistent changes in local tissue photosynthetic function in the early stages of disease infection are distinguished from random fluctuations or transient disturbances. Disease infection is a progressive biological process, and the resulting changes in physiological parameters typically exhibit a trend of continuous enhancement or weakening, while environmental noise or equipment vibration often leads to irregular random fluctuations. By setting the condition that multiple consecutive cycles maintain the same direction of change and the amplitude of change exceeds a threshold, the randomness of single fluctuations is effectively eliminated, making the identified anomalies more biologically reliable. The three characteristics are analyzed independently. This is because different types or stages of disease infection may primarily affect different aspects of the photosynthetic system. For example, some diseases may initially manifest as a decrease in the excitation response rate, while others may first cause an increase in the peak fluctuation amplitude. Retaining this characteristic type differentiation information helps in subsequent judgment of the nature of the abnormality. Marking the abnormal points at specific spatial locations and recording their corresponding characteristic types provides accurate raw data for the next step of spatial clustering and region merging. This invention enables the invention to accurately locate those tiny regions with truly persistent physiological abnormalities from massive pixel-level data, laying the foundation for the final realization of early disease localization and type identification.

[0037] In another preferred embodiment of the present invention, the specific process for obtaining the dynamic evolution feature vector of the abnormal region in step S6 is as follows: First, obtain the spatial coordinates and corresponding feature types of all points marked as fluorescence kinetic anomalies. Assume there are 120 anomalies on the blade, each containing row coordinates, column coordinates, and anomaly type records. For example, an anomaly might be located in row 5, column 8 and belong to the excitation response rate anomaly type. Perform spatial clustering analysis based on Euclidean distance on the spatial coordinates of all anomalies. Iterate through each anomaly, calculating the straight-line distance between it and every other anomaly. When the distance between two anomalies is less than a preset neighborhood radius (e.g., a neighborhood radius of three grid cells), these two points are considered adjacent and assigned to the same set to be merged. Through iterative processing, merge all interconnected point sets. Finally... Multiple anomalous regions are formed. For example, fifteen anomalous points close together near the base of the blade are grouped into anomalous region A, eight anomalous points close together near the tip of the blade are grouped into anomalous region B, and the remaining isolated points are grouped into separate regions. For each anomalous region, the characteristic type distribution of all anomalous points within the region is statistically analyzed. Taking anomalous region A as an example, this region has fifteen anomalous points, of which ten belong to the excitation response rate anomalous type, three belong to the peak stability anomalous type, and two belong to the damped oscillation characteristic anomalous type. The excitation response rate anomalous type, which has the largest number of points, is determined as the dominant anomalous type of this region. Then, the characteristic values ​​of all anomalous points within this region at their respective labeled times are extracted. For example, for these fifteen anomalous points... In the excitation response rate analysis, some regions were marked in the third period with a rate of 0.82, while others were marked in the fifth period with a rate of 0.94. All these characteristic values ​​were extracted and their arithmetic mean was calculated, yielding a mean characteristic intensity of 0.88 for the anomalous region. The earliest marked time within this region was identified as the third period, and the latest marked time as the ninth period, with a time interval of six periods. Dividing the mean characteristic intensity of 0.88 by the six-period interval yielded an average characteristic intensity change rate of 0.147 per period for the anomalous region. Then, the spatial boundary of the anomalous region was obtained under continuous excitation cycles, from the third to the ninth period, showing the number of grid cells covered by the anomalous region in each cycle. The number may change. For example, in the third cycle, the area covers only five grid cells, in the fourth cycle it expands to seven, in the fifth cycle to nine, in the sixth cycle to eleven, in the seventh cycle to twelve, in the eighth cycle to thirteen, and in the ninth cycle to fourteen. These values ​​are used as the area of ​​the region in each cycle. The rate of change of area between adjacent cycles is calculated. For example, subtracting the area of ​​the third cycle from the area of ​​the fourth cycle gives two grid cells, and dividing by the area of ​​the third cycle (five grid cells) gives 0.4. Subtracting the area of ​​the fourth cycle from the area of ​​the fifth cycle gives two grid cells, and dividing by the area of ​​the fourth cycle (seven grid cells) gives 0.286. And so on, six rates of change of area are obtained. The arithmetic mean of these rates of change is taken to obtain the average rate of area expansion of the abnormal region.Finally, the average rate of change of characteristic intensity (0.147 per cycle) and the average rate of area expansion are combined in a preset order, for example, by placing the average rate of change of characteristic intensity first and the average rate of area expansion second, to generate a two-dimensional numerical vector as the characteristic vector of the dynamic evolution of the anomalous region.

[0038] Spatial clustering integrates discrete anomalies into regions with practical biological significance. Disease infection is typically not an isolated point event, but rather forms a continuous affected area within a local tissue. Merging adjacent anomalies restores this spatial continuity. Statistical analysis of the dominant anomaly type within the region reveals the main abnormal patterns of photosynthetic function, providing crucial clues for subsequent disease type identification. This is because physiological disorders caused by different pathogens often preferentially manifest in specific fluorescence parameters. Calculating the average rate of change of feature intensity quantifies the severity of anomaly development, reflecting the activity and progression speed of disease infection, while the average rate of area expansion reflects the speed of anomaly range expansion. These two rates characterize the dynamic evolution of the disease from the dimensions of intensity and breadth, respectively. Combining these two into a two-dimensional feature vector preserves the core dynamic information of anomaly development and achieves a structured representation of the data. This provides a unified input format for subsequent matching with standard vectors in the disease response pattern library, enabling the invention to identify disease types based on dynamic patterns of disease development rather than static features at a single moment, thus more accurately capturing the unique infection process characteristics of different diseases.

[0039] In another preferred embodiment of the present invention, the specific construction process of the disease response pattern library in step S7 is as follows: Multiple groups of healthy plants were selected and inoculated with a single type of known disease. After inoculation, each plant was placed in a dark room and steps S1 to S6 were performed to obtain the dynamic evolution feature vector of each abnormal region on each inoculated plant. At the same time, the disease type label corresponding to each abnormal region was recorded. Each dynamic evolution feature vector and its corresponding disease type label were combined into a sample, and all samples constituted the training dataset. Each dynamic evolution feature vector is used as network input, and the corresponding disease type label is converted into a one-hot encoding as the training target. The neural network parameters are initialized, and samples from the training dataset are input into the neural network in batches for forward propagation. The predicted probability distribution is obtained by normalizing the exponential function at the output layer. The cross-entropy loss between the predicted probability distribution and the true one-hot encoded label is calculated. The gradient of the loss with respect to the neural network parameters is calculated using the backpropagation algorithm, and the neural network parameters are updated using the gradient descent algorithm. The above process is repeated until the classification accuracy of the neural network on the validation set tends to stabilize. The neural network parameters at this time are saved as a disease response pattern library.

[0040] By collecting real data from a large number of known disease infection samples, the neural network autonomously learns the differences in fluorescence dynamics evolution characteristics of different disease types, establishing a mapping relationship between abstract two-dimensional vectors and specific disease types. Using one-hot encoding as the training objective clearly expresses the non-numerical relationships between categories, avoiding the incorrect assignment of order or magnitude to different types. The iterative training process of batch input and gradient descent allows the network to gradually adjust parameters to minimize prediction errors, ultimately converging to a state that accurately distinguishes various diseases. The network parameters saved after training are essentially a mathematical function mapping the input feature vector to the probability distribution of disease types. When feature vectors of subsequent unknown samples are input, the network can output the corresponding disease type based on the learned patterns. This data-driven pattern library construction method avoids the subjectivity and limitations of manually setting classification rules, automatically discovering subtle differences in fluorescence dynamics evolution characteristics of different diseases. This enables the invention to make accurate judgments based on a large amount of prior knowledge when facing actual unknown samples, ultimately achieving early identification based on dynamic disease development patterns rather than static symptoms.

[0041] In another preferred embodiment of the present invention, if multiple mutually separated abnormal regions appear on the same blade, a dynamic evolution feature vector of each abnormal region is extracted, wherein the dynamic evolution feature vector is composed of the average rate of change of the feature intensity and the average rate of area expansion of the region. Calculate the Euclidean distance between the dynamic evolution feature vectors of any two anomalous regions, and mark two anomalous regions whose Euclidean distance is less than a preset similarity threshold as a pair of similar anomalous regions; For each pair of similar anomalous regions, the earliest and latest labeled times of all fluorescence dynamics anomalous points in the two anomalous regions are extracted, the time interval overlap length of the labeled times of the two anomalous regions is calculated, and similar anomalous region pairs with an overlap length greater than a preset overlap threshold are marked as spatiotemporal synchronization anomalous region pairs. All anomalous regions connected by spatiotemporal synchronization anomalous regions are grouped into the same associated anomalous region group, which contains multiple spatially separated anomalous regions. Anomalies that do not form a spatiotemporally synchronized anomaly pair with any other anomalies are retained as independent anomalies.

[0042] When multiple spatially separated anomalous regions appear on the same leaf, it is necessary to determine whether there is an inherent relationship between them. This is because infection by the same pathogen may form multiple independent infection points on different parts of the leaf. Although these infection points are spatially separated, they often have similar dynamic evolution characteristics and synchronous development processes. By calculating the Euclidean distance between feature vectors, the similarity of different anomalous regions in terms of feature intensity change rate and area expansion rate can be quantified, allowing for the selection of regions with similar features. Further comparison of the overlap length of time intervals can confirm whether these similar regions are actively developing within the same time period, thus eliminating cases where, although the features are similar, their development sequences are completely staggered. Regions that pass the spatiotemporal synchronization test are grouped into the same associated abnormal region group. This means that these spatially separated but similar and synchronously developing regions are regarded as manifestations of the same disease event in different parts of the leaf and should be judged as a whole in subsequent disease type identification. Independent abnormal regions may represent infection points of another disease or local anomalies caused by non-disease factors and need to be treated separately. This spatiotemporal correlation analysis enables the present invention to correctly handle multi-region relationships in complex situations, avoid misjudging multiple infection points of the same disease as different diseases, and avoid erroneously merging infection points of different diseases. This provides a more accurate basis for regional division that conforms to the actual pathological process for subsequent accurate output of disease type.

[0043] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for early identification of crop diseases based on computer vision, characterized in that, Includes the following steps: S1, Acquire delayed fluorescence time-series images of the target crop leaves under multiple consecutive excitation cycles, each excitation cycle including a pulse excitation and a subsequent fluorescence decay process; S2, divide the delayed fluorescence time-series image into leaf region grids, extract the fluorescence intensity value of each grid unit in each excitation cycle, and construct a fluorescence intensity response curve with the excitation cycle number as the independent variable; S3, numerically fit each fluorescence intensity response curve, extract the average slope of the rising phase, the fluctuation amplitude of the peak region and the fluctuation frequency of the decay phase, and combine them into the fluorescence dynamics feature vector of the corresponding grid cell. S4. Arrange the fluorescence dynamics characteristic vectors of all grid units within the same excitation cycle according to their spatial positions to form the excitation response rate distribution map, peak stability distribution map, and decay oscillation characteristic distribution map, respectively. S5, track the trend of feature value changes of each spatial location in continuous excitation cycles, and mark spatial locations where the feature value changes monotonically and the change amplitude exceeds the preset threshold as fluorescence dynamics abnormal points. S6, perform spatial clustering on the fluorescence dynamics anomalies, merge adjacent anomalies into anomaly regions, calculate the average rate of change of feature values ​​and the rate of expansion of region area for each anomaly region, and combine them into a dynamic evolution feature vector of the anomaly region. S7, match the dynamic evolution feature vector with the standard vector in the disease response pattern library, and output the corresponding disease type when the matching degree exceeds the preset threshold.

2. The method for early identification of crop diseases based on computer vision according to claim 1, characterized in that, In S1, the specific acquisition process of the delayed fluorescence time-series image is as follows: In a darkroom environment, a high-sensitivity camera and a pulsed excitation light source are fixed on a movable bracket, with the camera's optical axis aligned with the canopy of the crop to be tested, and the illumination range of the pulsed excitation light source covering the entire leaf area within the camera's field of view. The pulse excitation light source is set to repeatedly emit pulse light at a preset fixed time interval. The duration and intensity of each pulse light remain constant, and the time interval between adjacent pulses is greater than the time required for the fluorescence signal generated by the previous pulse excitation to completely decay below the camera's light sensitivity threshold. At the moment each pulse light emission ends, a high-sensitivity camera is activated to continuously acquire images of the blade surface within the current field of view. The camera captures images at a preset fixed frame rate until the intensity of the fluorescence signal generated by the current pulse excitation decays below the camera's photosensitive threshold, at which point acquisition stops. All images acquired in this acquisition are arranged in chronological order of capture time as a delayed fluorescence image sequence for that excitation cycle. After completing the acquisition of the current field of view, the movable support is moved to the next crop canopy location to be tested, and the pulse excitation and image acquisition process is repeated until the data acquisition of all preset acquisition points is completed. The delayed fluorescence image sequences obtained in all excitation cycles are arranged in the order of acquisition point location and excitation cycle to form a delayed fluorescence time series image set covering multiple acquisition points and multiple consecutive excitation cycles. Each frame in this delayed fluorescence time series image set is marked with its acquisition point identifier, excitation cycle number and shooting time number within that cycle.

3. The method for early identification of crop diseases based on computer vision according to claim 1, characterized in that, In S3, the specific process for generating the fluorescence dynamics feature vector is as follows: Numerical fitting was performed on each fluorescence intensity response curve. The fitting parameters were calculated using a nonlinear regression model that included an exponential decay term. Based on the fitting parameters, the average rate of change of the curve from the starting point of the rising phase to the first peak point was calculated. This average rate of change was defined as the excitation response rate value of the grid cell. In the peak region of the fluorescence intensity response curve, the fluorescence intensity values ​​corresponding to the peak point and several consecutive moments before and after it are selected to form a local sequence. The standard deviation of all fluorescence intensity values ​​in the local sequence is calculated, and the standard deviation is defined as the peak fluctuation amplitude of the grid cell. Spectral analysis is performed on the decay phase of the fluorescence intensity response curve. The decay segment data from the peak point to the end point of the curve is converted into a frequency domain signal through Fourier transform. The frequency component corresponding to the maximum amplitude in the frequency domain signal is extracted and the frequency component is defined as the decay oscillation frequency value of the grid cell. The excitation response rate value, peak fluctuation amplitude value, and decay oscillation frequency value are combined in a preset order to generate a three-dimensional numerical vector as the fluorescence dynamics feature vector of the grid cell.

4. The method for early identification of crop diseases based on computer vision according to claim 1, characterized in that, In step S4, the specific process of forming the excitation response rate distribution map, peak stability distribution map, and damped oscillation characteristic distribution map is as follows: Extract the fluorescence dynamics feature vectors of all grid cells within the same excitation cycle, extract the excitation response rate value from each feature vector, and fill it into the same row and column positions of a two-dimensional matrix according to the row and column numbers of the grid cell in the blade region. The number of rows of the two-dimensional matrix is ​​equal to the number of grid cells in the vertical direction of the blade region, and the number of columns is equal to the number of grid cells in the horizontal direction. The filled two-dimensional matrix is ​​defined as the excitation response rate distribution map of the excitation cycle. The peak fluctuation amplitude values ​​of the fluorescence dynamics eigenvectors of all grid cells within the same excitation cycle are extracted and filled into the corresponding positions of another two-dimensional matrix of the same size according to the same row and column numbers of each grid cell. The filled two-dimensional matrix is ​​defined as the peak stability distribution map of the excitation cycle. Take the decay oscillation frequency values ​​from the fluorescence dynamics eigenvectors of all grid cells within the same excitation period, and fill them into the corresponding positions of a third two-dimensional matrix of the same size according to the same row and column numbers of each grid cell. Define the filled two-dimensional matrix as the decay oscillation characteristic distribution map of the excitation period.

5. The method for early identification of crop diseases based on computer vision according to claim 1, characterized in that, In S5, the specific identification process for fluorescence dynamics anomalies is as follows: For the excitation response rate distribution map, peak stability distribution map, and damped oscillation characteristic distribution map, a sequence of the same type of distribution map under multiple consecutive excitation cycles is obtained. Each distribution map in the sequence is aligned according to the same spatial coordinate system so that the same row number and column number position in the distribution map of different cycles corresponds to the same spatial position of the blade region. For each spatial location, feature values ​​are extracted from the excitation response rate distribution sequence, peak stability distribution sequence, and decay oscillation feature distribution sequence to form a time series of the location on the three features. The difference between adjacent periodic feature values ​​in each time series is calculated. When at least two consecutive periods of positive difference and absolute value exceed a preset threshold appear in the time series of any feature, the feature is determined to have a continuous upward trend. When at least two consecutive periods of negative difference and absolute value exceed a preset threshold appear, the feature is determined to have a continuous downward trend. Spatial locations where any feature shows a continuous upward or downward trend are marked as fluorescence kinetic anomalies, and the feature type corresponding to the anomaly is recorded.

6. The method for early identification of crop diseases based on computer vision according to claim 1, characterized in that, In step S6, the specific process for obtaining the dynamic evolution feature vector of the abnormal region is as follows: Spatial clustering based on Euclidean distance is performed on the spatial coordinates of all fluorescence dynamics anomalies. Anomalies whose distance to each other is less than the preset neighborhood radius are grouped into the same anomaly region. Each anomaly region contains several anomalies and their corresponding feature types. For each abnormal region, the number of abnormal points in the region belonging to the excitation response rate abnormality type, peak stability abnormality type, and decay oscillation characteristic abnormality type is counted, and the abnormality type with the largest number is determined as the dominant abnormality type of the abnormal region. Extract the feature values ​​of all anomalies in the anomaly region at their respective marked times, calculate the arithmetic mean of these feature values ​​as the mean feature intensity of the anomaly region, extract the time interval between the earliest and latest marked times in the anomaly region, and divide the mean feature intensity by the time interval to obtain the average rate of change of feature intensity of the anomaly region. Obtain the spatial boundary of the anomalous region under continuous excitation cycles, calculate the number of grid cells covered by the anomalous region in each cycle as the region area of ​​that cycle, divide the difference in region area between adjacent cycles by the region area of ​​the previous cycle to obtain the area change rate of that cycle, and take the arithmetic mean of the area change rates of all cycles to obtain the average rate of area expansion of the anomalous region. The average rate of change of the characteristic intensity and the average rate of area expansion are combined in a preset order to generate a two-dimensional numerical vector as the dynamic evolution characteristic vector of the abnormal region.

7. The method for early identification of crop diseases based on computer vision according to claim 1, characterized in that, In S7, the specific construction process of the disease response model library is as follows: Multiple groups of healthy plants were selected and inoculated with a single type of known disease. After inoculation, each plant was placed in a dark room and steps S1 to S6 were performed to obtain the dynamic evolution feature vector of each abnormal region on each inoculated plant. At the same time, the disease type label corresponding to each abnormal region was recorded. Each dynamic evolution feature vector and its corresponding disease type label were combined into a sample, and all samples constituted the training dataset. Each dynamic evolution feature vector is used as network input, and the corresponding disease type label is converted into a one-hot encoding as the training target. The neural network parameters are initialized, and samples from the training dataset are input into the neural network in batches for forward propagation. The predicted probability distribution is obtained by normalizing the exponential function at the output layer. The cross-entropy loss between the predicted probability distribution and the true one-hot encoded label is calculated. The gradient of the loss with respect to the neural network parameters is calculated using the backpropagation algorithm, and the neural network parameters are updated using the gradient descent algorithm. The above process is repeated until the classification accuracy of the neural network on the validation set tends to stabilize. The neural network parameters at this time are saved as a disease response pattern library.

8. The method for early identification of crop diseases based on computer vision according to claim 6, characterized in that, It also includes the case where multiple mutually separated anomalous regions appear on the same blade, in which the dynamic evolution feature vector of each anomalous region is extracted, and the dynamic evolution feature vector is composed of the average rate of change of the feature intensity and the average rate of area expansion of the region. Calculate the Euclidean distance between the dynamic evolution feature vectors of any two anomalous regions, and mark two anomalous regions whose Euclidean distance is less than a preset similarity threshold as a pair of similar anomalous regions; For each pair of similar anomalous regions, the earliest and latest labeled times of all fluorescence dynamics anomalous points in the two anomalous regions are extracted, the time interval overlap length of the labeled times of the two anomalous regions is calculated, and similar anomalous region pairs with an overlap length greater than a preset overlap threshold are marked as spatiotemporal synchronization anomalous region pairs. All anomalous regions connected by spatiotemporal synchronization anomalous regions are grouped into the same associated anomalous region group, which contains multiple spatially separated anomalous regions. Anomalies that do not form a spatiotemporally synchronized anomaly pair with any other anomalies are retained as independent anomalies.

Images