Tobacco disease automatic identification method and model

Abstract

The invention discloses a tobacco disease automatic identification method and model, and the method comprises the following steps: calculating the probability of infecting a disease K based on image features, obtaining non-image features, calculating a non-image feature adjustment factor, calculating a comprehensive adjustment factor, calculating the disease incidence probability, and carrying out the arrangement from large to small according to the disease incidence probability. And outputting probabilities and disease types of the first N diseases, and displaying corresponding image features and non-image features. The method has the advantages that the limitation of single image analysis is broken through, and a multi-source information fusion decision framework of image features, environmental climate, growth stages and historical illness conditions is constructed.

Description

Technical Field

[0001] This invention relates to the field of automatic disease identification technology, and in particular to an automatic identification method and model for tobacco diseases. Background Technology

[0002] Currently, automatic identification of tobacco diseases mainly relies on image processing and deep learning technologies, classifying diseases by analyzing visual features such as leaf color, texture, and shape. However, this method has significant limitations: on the one hand, many diseases exhibit highly similar visual symptoms (e.g., red spot disease and wildfire lesions), making misjudgment highly likely based solely on images; on the other hand, it ignores the agroecological nature of disease occurrence—disease outbreaks are the result of interactions between pathogens, hosts, and the environment, and their occurrence is closely related to specific climatic conditions, tobacco growth stages, and historical disease history in the field. Relying solely on image recognition carries a high risk of misjudgment and delayed early warnings when early disease symptoms are not obvious, symptoms are similar, or image quality is poor.

[0003] Therefore, to improve system reliability and early warning capabilities, it is necessary to break through the limitations of single image analysis and build a multi-source intelligent decision-making model that integrates real-time image features, environmental and climate data, crop growth period and historical disease information, so as to achieve more accurate and robust automated diagnosis. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an automatic identification method and model for tobacco diseases that integrates real-time image features, environmental climate data, crop growth period and historical disease information.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is an automatic identification method for tobacco diseases, comprising the following steps: S1. Calculate the probability of disease K infection based on image features: The collected crop images are transmitted to a trained deep learning diagnostic model, which directly outputs the preliminary probability distribution of the symptoms presented in the images belonging to various diseases, denoted as P. k (image); P k (image) represents the probability of crop infection with disease K determined based on image features; S2. Obtain non-image features: Non-image features include meteorological data from the past N days, the current growth stage, and the actual incidence rate of diseases in the past n years. Meteorological data includes daily temperature, humidity, light intensity, wind speed, and precipitation. The growth stage is the number of days after transplanting, with the day of transplanting being the first day. S3. Calculate the non-image feature adjustment factor: Calculate the non-image feature adjustment factor for disease K based on non-image features, including climate matching degree C. k Growth stage matching degree G k (t) and historical disease weight Hk ; S4. Calculate the comprehensive adjustment factor T k : T k = β 1k *C k + β 2k *G k (t) + β 3k *H k ; and β 1k +β 2k +β 3k =1; Where β 1k β 2k β 3k These represent the weights of climate, growth stage, and historical disease transfer for disease K; β 1k β 2k β 3k All are greater than or equal to 0; S5. Calculate the probability of onset P. K : ; S6. According to the disease incidence probability P K Sort the diseases from largest to smallest, output the probability and type of the top N diseases, and display the corresponding image features and non-image features.

[0006] Furthermore, in step S3, the climate matching degree C k The calculation process is as follows: S31. Calculate the temperature matching degree Mtemp, humidity matching degree Mrh, light matching degree Mlight, wind speed matching degree Mwind, and precipitation matching degree Mrain respectively. S32. Determine the relative weights of temperature, humidity, light intensity, wind speed, and precipitation, and represent them as αtemp, αrh, αlight, αwind, and αrain, respectively; αtemp + αrh + αlight + αwind + αrain = 1; S33, Calculate C k C k It is the weighted geometric mean of five climatic factors: temperature, humidity, light intensity, wind speed, and precipitation. C k = (Mtemp^αtemp) × (Mrh^αrh) × (Mlight^αlight) × (Mwind^αwind) ×(Mrain^αrain).

[0007] Furthermore, the method for calculating the matching degree Gk(t) during the growth stage in step S3 is as follows: ; in: Qk represents the basic matching degree of disease k, 0 ≤ Qk≤0.3; Rk represents the peak susceptibility of disease k, indicating the theoretical maximum matching degree at the optimal time, 0≤Rk≤ 1; t 0,k The center point representing the most susceptible period for disease k indicates the time when disease k is most susceptible after transplanting. 0,k The day is the most likely time for this to happen; the unit is the day. ρk represents the time width parameter of disease k, which controls the rate at which susceptibility decays over time. The unit is days. The larger the value, the longer the period of disease susceptibility. Qk, Rk, t for different diseases K 0,k The values ​​of ρk are set in advance.

[0008] Calculate C once a day k And according to C k Prompt, C k If the value is less than 0.4, a green indicator will appear, indicating safety; if the value is less than or equal to C... k If the value is less than 0.6, a yellow warning is issued, indicating that attention is needed; if the value is 0.6 ≤ C, the warning is positive. k If the value is less than 0.8, an orange indicator is displayed, suggesting a need for caution; C k If the value is ≥ 0.8, a red alert will be issued.

[0009] Furthermore, the method for calculating the historical disease weight Hk in step S3 is as follows: ; in: n represents the number of historical years considered; i represents the historical year number, i=1 indicates the most recent year, and i=n indicates the oldest year; λ represents the decay factor, which controls the decay rate of historical influence; 0 <λ≤1, the larger λ is, the more persistent the historical influence is; I k,i The standardized incidence rate of disease k in year i, 0≤I k,i ≤ 1; I k,i =P k,i / P k,max ; P k,i This represents the actual incidence rate of disease k in year i. P k,max This represents the theoretical maximum incidence rate of disease k.

[0010] Furthermore, in step S4, β1k β 2k β 3k The method for determining the weights is to manually establish weight matrices for different disease types.

[0011] Furthermore, the specific method of step S1 is as follows: S11. Image Acquisition: Acquire raw images of flue-cured tobacco in the field, including overall images of the plant, images of the stem, and images of several leaves from the same plant. S12. Image preprocessing: Preprocess the original images of flue-cured tobacco collected in the field; S13, Calculate P leaf-n(k) For each leaf image, leaf features are extracted, and the probability of disease on each leaf for different diseases K is calculated and denoted as P. leaf-n(k) K represents the disease, and leaf-n represents the nth leaf of the plant. S14, Leaf Evidence Aggregation: Based on the P of several leaves leaf-n(k) The weighted average disease probability P of the leaves was calculated. leaf(k) ; S15, Calculate P stem(k) Based on the plant stem images, stem features are extracted, and the probability P of stem disease in response to different diseases K is calculated. stem(k) ; S16, Calculate P whole(k) Based on the overall image of the plant, extract overall features and calculate the probability P of the plant developing different diseases K. whole(k) .

[0012] S17. Based on the calculated P leaf(k) P stem(k) P whole(k) The probability P of plant developing disease K was calculated. k (image)

[0013] Furthermore, in step S14, P leaf-(k) The calculation method includes the following steps: S141. Calculate the weights Wn of several blades:

[0014] Alesion,n is the area of ​​the nth leaf; Aleaf,n represents the total area of ​​several lesions on the nth leaf; It is a constant; S142, Calculate P leaf(k) : ; in: N is the total number of leaves analyzed; P leaf-n(k) Let n be the probability of disease K occurring on the nth leaf. Wn is the weight of the nth leaf.

[0015] Furthermore, in step S17, P k The calculation method for (image) includes the following steps: S171. Determine the weights of leaves, stems, and the whole plant, denoted as w respectively. leaf(k) ,w stem(k) ,w whole(k) ; S172, Calculate P k (image): .

[0016] Furthermore, the method in step S171 includes the following steps: S1711. Based on experience, experts pre-determine an initial weight matrix for leaves, stems, and the whole plant for different diseases k; denoted as: w leaf(k)1 ,w stem(k)1 ,w whole(k)1 Satisfying w leaf(k)1 +w stem(k)1 +w whole(k)1 =1; S1712. Preset a threshold range for the probability of disease on leaves, stems, and the whole plant for each disease k. S1713. In actual diagnosis, based on the calculated P... leaf(k) P stem(k) P whole(k) Each component is compared with its respective incidence probability threshold range. If the incidence probability threshold range is exceeded, the initial weight matrix is ​​adjusted to obtain the final weight matrix. The adjustment rule is: if the probability value is higher than the upper limit, it means that the symptoms of that part are very typical, so its weight is increased; if the probability value is lower than the lower limit, it means that the symptoms of that part are not obvious, so its weight is decreased. S1714. Normalize the adjusted weights to obtain the final weight matrix w for leaves, stems, and the whole plant. leaf(k) ,w stem(k) ,w whole(k) .

[0017] Furthermore, The range is 2.5-3.5.

[0018] Furthermore, The value is 3.

[0019] Furthermore, the adjustment range in step S1713 is an adjustment constant. If it exceeds the upper limit, the weight increases the adjustment constant; if it exceeds the lower limit, the weight decreases the adjustment constant.

[0020] Another adjustment range can be set to adjust proportionally based on the degree to which the threshold is exceeded.

[0021] Furthermore, in step S13, P leaf-n(k) The calculation method includes the following steps: S131. Calculate the feature value fi of the leaf features extracted for each leaf; S132, Eigenvalue Standardization: Standardize the eigenvalue fi for each feature i; finorm = (fi - μi) / σi; Where fi is the original feature value; μi is the mean of the feature in the training set; σi is the standard deviation of the feature in the training set; and finorm is the standardized feature value. S133. Determine the weight matrix W(k,i) of the image features of different diseases K; S134. Calculate the base score S for each disease K. k (image): ; Where: b k This represents the prior probability of disease k when there is no feature information. S135. Calculate the preliminary probability of image features: Calculate the base score S k (image) converted to the probability of disease incidence P leaf-n(k) : .

[0022] Furthermore, the image feature weight matrix W(k,i) in step S133 is determined by the following method: W(k,i)=α·Wexpert+(1-α)·Wlearned; Where Wexpert is a weight matrix determined by experts based on experience; Wlearned is the weight matrix obtained through data-driven learning; α is the expert trust coefficient, α∈[0,1]. α is set manually and can be adjusted in a timely manner according to the operation.

[0023] Furthermore, μi and σi in step S132 are updated periodically: μi(new)=θ*μi(old)+(1-θ)*μi(newsample); ; Where θ is the forgetting factor, and the value of θ can be defined manually.

[0024] Furthermore, in step S134, b k It is calculated based on historical incidence rates: ; Pkprior represents the historical incidence probability of disease k over a past period of time.

[0025] Furthermore, the stem disease probability P in step S15 stem(k) And the probability of disease on the whole plant P in step S16 whole(k) The calculation method and P leaf-n(k) The calculation method is the same, except that the corresponding leaf features are replaced with stem features and plant features, and the weight matrix of leaf features is replaced with the weight matrix of stem features and plant features.

[0026] Furthermore, the leaf characteristics include at least one of the following: lesion area ratio, distribution dispersion, weighted average roundness, weighted average elongation, weighted average fractal dimension, hue skewness, and chromaticity dispersion.

[0027] Furthermore, the plant characteristics include at least one of the following: canopy structure volume index, leaf yellowing ratio, and apical wilting index.

[0028] Furthermore, the stem characteristics include at least one of the following: stem base melanization ratio, stem longitudinal stripe index, and stem base diameter abnormality rate.

[0029] The present invention also provides an automatic tobacco disease identification model capable of implementing the above method.

[0030] Beneficial effects of this invention: The core advantage of this solution lies in overcoming the limitations of single-image analysis and constructing a multi-source information fusion decision-making framework integrating "image features + environmental climate + growth stage + historical disease status". Its advantages are significant: First, by introducing prior knowledge such as climate matching degree and reproductive period matching degree, the ability of early diagnosis and warning has been greatly improved, solving the problem of misjudgment when relying solely on visual features when symptoms are in their early stages or similar.

[0031] Secondly, this invention introduces a disease-specific dynamic weight fusion mechanism, β 1k β 2k β 3k It can be customized for different diseases (e.g., soil-borne diseases rely more on historical weights, while airborne diseases rely more on climate weights), realizing adaptive adjustment of diagnostic strategies instead of applying fixed fusion rules to all diseases, which significantly improves the refinement and accuracy of the model.

[0032] This invention deeply integrates computer vision with agricultural pathology, achieving a leap from "symptom identification" to "cause diagnosis," and providing more reliable and intelligent decision support for precision prevention and control in the field. Detailed Implementation

[0033] The specific embodiments of the present invention will be further described below. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0034] Example 1 An automatic identification method for tobacco diseases includes the following steps: S1. Calculate the probability of disease K infection based on image features: The collected crop images are transmitted to a trained deep learning diagnostic model, which directly outputs the preliminary probability distribution of the symptoms presented in the images belonging to various diseases, denoted as P. k (image); P k (image) represents the probability of crop infection with disease K determined based on image features; S2. Obtain non-image features: Non-image features include meteorological data from the past N days, the current growth stage, and the actual incidence rate of diseases in the past n years. Meteorological data includes daily temperature, humidity, light intensity, wind speed, and precipitation. The growth stage is the number of days after transplanting, with the day of transplanting being the first day. S3. Calculate the non-image feature adjustment factor: Calculate the non-image feature adjustment factor for disease K based on non-image features, including climate matching degree C. k Growth stage matching degree G k (t) and historical disease weight H k ; S4. Calculate the comprehensive adjustment factor T k : T k = β 1k *C k + β 2k *G k (t) + β 3k *H k ; and β 1k +β 2k +β 3k =1; Where β 1k β 2k β 3k These represent the weights of climate, growth stage, and historical disease transfer for disease K; β1k β 2k β 3k All are greater than or equal to 0; S5. Calculate the probability of onset P. K : ; S6. According to the disease incidence probability P K Sort the diseases from largest to smallest, output the probability and type of the top N diseases, and display the corresponding image features and non-image features.

[0035] The aforementioned automatic tobacco disease identification method overcomes the limitations of single image analysis by constructing a multi-source information fusion decision-making framework integrating "image features + environmental climate + growth stage + historical disease status". Its advantages are significant: First, by introducing prior knowledge such as climate matching degree and reproductive period matching degree, the ability of early diagnosis and warning has been greatly improved, solving the problem of misjudgment when relying solely on visual features when symptoms are in their early stages or similar.

[0036] Secondly, this invention introduces a disease-specific dynamic weight fusion mechanism, β 1k β 2k β 3k It can be customized for different diseases (e.g., soil-borne diseases rely more on historical weights, while airborne diseases rely more on climate weights), realizing adaptive adjustment of diagnostic strategies instead of applying fixed fusion rules to all diseases, which significantly improves the refinement and accuracy of the model.

[0037] This invention deeply integrates computer vision with agricultural pathology, achieving a leap from "symptom identification" to "cause diagnosis," and providing more reliable and intelligent decision support for precision prevention and control in the field.

[0038] Specifically, in step S3, the climate matching degree C k The calculation process is as follows: S31. Calculate the temperature matching degree Mtemp, humidity matching degree Mrh, light matching degree Mlight, wind speed matching degree Mwind, and precipitation matching degree Mrain respectively. S32. Determine the relative weights of temperature, humidity, light intensity, wind speed, and precipitation, and represent them as αtemp, αrh, αlight, αwind, and αrain, respectively; αtemp + αrh + αlight + αwind + αrain = 1; S33, Calculate C k C k It is the weighted geometric mean of five climatic factors: temperature, humidity, light intensity, wind speed, and precipitation. C k= (Mtemp^αtemp) × (Mrh^αrh) × (Mlight^αlight) × (Mwind^αwind) ×(Mrain^αrain).

[0039] This invention uses a weighted geometric mean to integrate multiple meteorological factors, cleverly simulating the "barrel effect" (that any one climatic factor being severely unsuitable can significantly reduce the matching degree), which is more in line with the sensitivity of disease occurrence to extreme conditions.

[0040] Specifically, the method for calculating the matching degree Gk(t) in step S3 during the growth stage is as follows: ; in: Qk represents the basic matching degree of disease k, 0 ≤ Qk≤0.3; Rk represents the peak susceptibility of disease k, indicating the theoretical maximum matching degree at the optimal time, 0≤Rk≤ 1; t 0,k The center point representing the most susceptible period for disease k indicates the time when disease k is most susceptible after transplanting. 0,k The day is the most likely time for this to happen; the unit is the day. ρk represents the time width parameter of disease k, which controls the rate at which susceptibility decays over time. The unit is days. The larger the value, the longer the period of disease susceptibility. Qk, Rk, t for different diseases K 0,k The values ​​of ρk are set in advance.

[0041] This invention uses Gaussian functions to characterize the susceptibility changes of diseases during specific growth stages, transforming the qualitative "susceptibility period" into a continuous and quantifiable matching curve.

[0042] Specifically, the calculation method for the historical disease weight Hk in step S3 is as follows: ; in: n represents the number of historical years considered; i represents the historical year number, i=1 indicates the most recent year, and i=n indicates the oldest year; λ represents the decay factor, which controls the decay rate of historical influence; 0 <λ≤1, the larger λ is, the more persistent the historical influence is; I k,i The standardized incidence rate of disease k in year i, 0≤I k,i ≤ 1; I k,i =P k,i / P k,max ; P k,iThis represents the actual incidence rate of disease k in year i. P k,max This represents the theoretical maximum incidence rate of disease k.

[0043] Specifically, in step S4, β 1k β 2k β 3k The method for determining this is to manually establish a weight matrix for different disease types. In another embodiment, β 1k β 2k β 3k It can also be obtained through data-driven learning.

[0044] This invention quantifies the cumulative effect of pathogens based on an exponential decay model, reflecting the "memory" of historical disease occurrences and its decay over time.

[0045] Specifically, in step S1, P k The calculation method for (image) includes the following steps: S11. Image Acquisition: Acquire raw images of flue-cured tobacco in the field, including overall images of the plant, images of the stem, and images of several leaves from the same plant. S12. Image preprocessing: Preprocess the original images of flue-cured tobacco collected in the field; S13, Calculate P leaf-n(k) For each leaf image, leaf features are extracted, and the probability of disease on each leaf for different diseases K is calculated and denoted as P. leaf-n(k) K represents the disease, and leaf-n represents the nth leaf of the plant. S14, Leaf Evidence Aggregation: Based on the P of several leaves leaf-n(k) The weighted average disease probability P of the leaves was calculated. leaf(k) ; S15, Calculate P stem(k) Based on the plant stem images, stem features are extracted, and the probability P of stem disease in response to different diseases K is calculated. stem(k) ; S16, Calculate P whole(k) Based on the overall image of the plant, extract overall features and calculate the probability P of the plant developing different diseases K. whole(k) .

[0046] S17. Based on the calculated P leaf(k) P stem(k) P whole(k) The probability P of plant developing disease K was calculated. k (image)

[0047] The above-mentioned image-based disease diagnosis process can integrate the image features of leaves, stems and whole plants to achieve multi-site collaborative diagnosis, improve the comprehensiveness and accuracy of disease identification, realize the fusion of multi-source information, and effectively improve the accuracy of disease diagnosis.

[0048] Specifically, in step S14, P leaf-(k) The calculation method includes the following steps: S141. Calculate the weights Wn of several blades:

[0049] Alesion,n is the area of ​​the nth leaf; Aleaf,n represents the total area of ​​several lesions on the nth leaf; It is a constant; S142, Calculate P leaf(k) : ; in: N is the total number of leaves analyzed; P leaf-n(k) Let n be the probability of disease K occurring on the nth leaf. Wn is the weight of the nth leaf.

[0050] By dynamically adjusting the weights of leaves, stems, and the whole plant based on typical disease characteristics, the model's adaptability to atypical symptoms is enhanced, achieving adaptive weight adjustment. In another embodiment, the selected typical disease characteristics can replace the lesion area ratio (i.e., Alesion,n / Aleaf,n) with other disease characteristics.

[0051] Specifically, in step S17, P k The calculation method for (image) includes the following steps: S171. Determine the weights of leaves, stems, and the whole plant, denoted as w respectively. leaf(k) ,w stem(k) ,w whole(k) ; S172, Calculate P k (image): .

[0052] Specifically, the method in step S171 includes the following steps: S1711. Based on experience, experts pre-determine an initial weight matrix for leaves, stems, and the whole plant for different diseases k; denoted as: w leaf(k)1 ,w stem(k)1 ,w whole(k)1 Satisfying w leaf(k)1 +wstem(k)1 +w whole(k)1 =1; S1712. Preset a threshold range for the probability of disease on leaves, stems, and the whole plant for each disease k. S1713. In actual diagnosis, based on the calculated P... leaf(k) P stem(k) P whole(k) Each component is compared with its respective incidence probability threshold range. If the incidence probability threshold range is exceeded, the initial weight matrix is ​​adjusted to obtain the final weight matrix. The adjustment rule is: if the probability value is higher than the upper limit, it means that the symptoms of that part are very typical, so its weight is increased; if the probability value is lower than the lower limit, it means that the symptoms of that part are not obvious, so its weight is decreased. S1714. Normalize the adjusted weights to obtain the final weight matrix w for leaves, stems, and the whole plant. leaf(k) ,w stem(k) ,w whole(k) .

[0053] Specifically, The range is 2.5-3.5.

[0054] Specifically, The value is 3.

[0055] Specifically, the adjustment range in step S1713 is an adjustment constant. If it exceeds the upper limit, the weight increases the adjustment constant; if it exceeds the lower limit, the weight decreases the adjustment constant.

[0056] Another adjustment range can be set to adjust proportionally based on the degree to which the threshold is exceeded.

[0057] The proposed initial weight matrix is ​​based on expert settings and is dynamically adjusted according to the actual incidence probability. It can dynamically adjust the importance of different evidence sources (leaves, stems, whole plants) according to the strength of real-time diagnostic evidence, thereby more precisely simulating the attention allocation of experts during diagnosis.

[0058] Specifically, in step S13, P leaf-n(k) The calculation method includes the following steps: S131. Calculate the feature value fi of the leaf features extracted for each leaf; S132, Eigenvalue Standardization: Standardize the eigenvalue fi for each feature i; finorm = (fi - μi) / σi; Where fi is the original feature value; μi is the mean of the feature in the training set; σi is the standard deviation of the feature in the training set; and finorm is the standardized feature value. S133. Determine the weight matrix W(k,i) of the image features of different diseases K; S134. Calculate the base score S for each disease K. k (image): ; Where: b k This represents the prior probability of disease k when there is no feature information. bk can reflect the prevalence of diseases: some diseases are very common in certain regions or seasons (such as red spot disease in rainy areas), and even if the symptoms are not particularly obvious, they are likely to occur; while some diseases are very rare and require strong symptom evidence for diagnosis; by setting bk, common diseases can be prioritized, which is in line with the epidemiological law.

[0059] S135. Calculate the preliminary probability of image features: Calculate the base score S k (image) converted to the probability of disease incidence P leaf-n(k) : .

[0060] Specifically, the image feature weight matrix W(k,i) in step S133 is determined by the following method: W(k,i)=α·Wexpert+(1-α)·Wlearned; Where Wexpert is a weight matrix determined by experts based on experience; Wlearned is the weight matrix obtained through data-driven learning; α is the expert trust coefficient, α∈[0,1]. α is set manually and can be adjusted in a timely manner according to the operation.

[0061] Traditional model weights are entirely data-driven, which can easily lead to model bias or overfitting in agricultural fields when there are insufficient samples or noisy annotations. This invention creatively introduces a controllable fusion mechanism by weighting a weight matrix Wexpert (determined by experts based on experience) and a weight matrix Wlearned (obtained through data-driven learning). By adjusting the expert trust coefficient α, the system can smoothly transition between expert-driven (when data is scarce) and data-driven (when data is abundant) modes, greatly improving the model's practicality and robustness.

[0062] Specifically, in this embodiment, Wlearned is calculated as follows: ; in: μk,i is the average value of the samples of disease K on feature i; μi is the average value of all samples on feature i; σi is the standard deviation of all samples on feature i; N represents the total number of all samples in the historical disease image database; Nk represents the total number of samples in the historical disease image database.

[0063] In another embodiment, Wlearned can also employ other calculation methods.

[0064] Specifically, μi and σi in step S132 are updated periodically: μi(new)=θ*μi(old)+(1-θ)*μi(newsample); ; Where θ is the forgetting factor, and the value of θ can be defined manually.

[0065] In existing technologies, the normalization parameters of training data are usually fixed after model deployment, which cannot adapt to feature distribution drift caused by changes in environment and variety. This invention updates μi and σi periodically, enabling the system to adaptively track changes in feature distribution over time and in the environment, continuously maintaining consistency in feature scale, thereby ensuring long-term recognition accuracy, which is difficult to achieve with static models.

[0066] Specifically, in step S134, b k It is calculated based on historical incidence rates: ; Pkprior represents the historical incidence probability of disease k over a past period of time.

[0067] Specifically, the probability of stem disease P in step S15 stem(k) And the probability of disease on the whole plant P in step S16 whole(k) The calculation method and P leaf-n(k) The calculation method is the same, except that the corresponding leaf features are replaced with stem features and plant features, and the weight matrix of leaf features is replaced with the weight matrix of stem features and plant features.

[0068] Specifically, leaf characteristics include at least one of the following: lesion area ratio, distribution dispersion, weighted average roundness, weighted average elongation, weighted average fractal dimension, hue skewness, and hue dispersion.

[0069] Specifically, plant characteristics include at least one of the following: canopy structure volume index, leaf yellowing rate, and apical wilting index.

[0070] Specifically, stem characteristics include at least one of the following: stem base melanization ratio, stem longitudinal stripe index, and stem base diameter abnormality rate.

[0071] In this embodiment, the calculation methods for area ratio, distribution dispersion, weighted average roundness, weighted average elongation, weighted average fractal dimension, hue skewness, hue dispersion, canopy structure volume index, leaf yellowing ratio, apical wilting index, stem base blackening ratio, stem longitudinal stripe index, and stem base diameter anomaly rate are as follows: (1) The characteristic value f1 of the area ratio of lesions is calculated as follows: 1. Image Preprocessing and Leaf Segmentation: The original RGB image is converted to a color space more suitable for segmentation (such as HSV or Lab). Utilizing the significant differences in color (such as green channel, saturation) and texture between healthy leaf tissue and the background, the entire leaf region is extracted using thresholding, edge detection, or machine learning models (such as U-Net) to obtain a leaf mask. The total number of pixels within this mask is calculated, which is leaf A.

[0072] 2. Within the extracted leaf area, secondary segmentation is performed based on the color difference between lesions and healthy tissue: Specifically, a threshold segmentation method can be selected: In terms of saturation (S), brightness (V), or specific color components, the numerical range of lesions (usually appearing as yellow, brown, black, or white) differs significantly from that of healthy green areas. A binary mask of the lesion is obtained by setting a threshold or automatically determining the threshold using the Otsu method. The total number of pixels within the lesion mask is then calculated, which is the lesion.

[0073] Alternatively, machine learning segmentation can be used: pre-trained semantic segmentation models (such as U-Net and DeepLab) are used to directly classify the input image at the pixel level and output the lesion region. This method works better for complex lesions with blurred boundaries.

[0074] The total number of pixels within the lesion mask is the Alesion.

[0075] 3. Area Calculation and Ratio Determination: Count the total number of white pixels in the leaf mask and the lesion mask respectively, and use this count as the area of ​​each. Substitute this count into the following formula: f1 = Alesion / Aleaf.

[0076] (2) The method for calculating the characteristic value f2 of lesion distribution dispersion is as follows: 1. Locating the geometric center of lesions: First, identify all independent lesion regions on the leaf using image segmentation technology. For each lesion, calculate its geometric center coordinates (xi, yi); where i = 1, 2, 3, ..., r; r is the total number of lesions on the leaf. The geometric center coordinates are obtained by averaging the coordinates of all pixels within the lesion region. 2. Calculate the standard deviation of coordinates: Based on the centroid coordinates of all lesions, calculate the standard deviation σx of the x-coordinate sequence and the standard deviation σy of the y-coordinate sequence respectively; 3. Composite dispersion index f2: ; f2 represents the distribution dispersion. The larger the f2 value, the more dispersed the lesions are on the leaves; the smaller the value, the more concentrated the distribution.

[0077] Where σx and σy are the standard deviations of the centroid coordinates of all lesions in the x-axis and y-axis directions, respectively.

[0078] (3) The eigenvalue f3 of the weighted average circularity is calculated as follows: First, calculate f3,r separately for each lesion on the leaf; f3,r = 4π * Alesion / P 2 ; Where r represents the number of discrete lesions on the leaf; P is the perimeter of the lesion; Then, for the calculated r f3,r values, calculate the weighted average circularity f3 on the blade: f3 is calculated using the area-weighted average method: ; (4) The eigenvalue f4 of the weighted average elongation is calculated as follows: First, calculate f4,r separately for each lesion on the leaf; f4,r = Lmajor / Lminor; Where Lmajor is the length of the major axis of the lesion; Lminor is the length of the minor axis of the lesion; Then, for the r calculated f4,r values, calculate the weighted average circularity f4 of the blade: f4 is calculated using the area-weighted average method: ; (5) The formula for calculating the eigenvalue f5 of the weighted average fractal dimension is as follows: First, calculate the fractal dimension f5,r for each lesion on the leaf individually; f5,r=log(N(q)) / log(1 / q); Where N(q) is the number of boxes covering the lesions; q is the size of the box.

[0079] Then, for the calculated r f5,r values, calculate the weighted average f5 dimension of the blade: f5 is calculated using the area-weighted average method: ; (6) The eigenvalue f6 of the hue bias is calculated as follows: 1. Image segmentation and region extraction: Use segmentation techniques (thresholding or machine learning models) to accurately separate lesion regions from leaf images. Obtain a binary mask for this region.

[0080] 2. Color space conversion: Convert the original RGB image to the HSV (or HSL / HSI) color space, which is more in line with human vision for color perception.

[0081] 3. Extract the lesion tone array: Using the lesion mask obtained in step 1, extract the tone values ​​of all pixels located within the lesion area from the H channel image, forming an array {H1, H2, ..., H...} containing N tone values. N}

[0082] 4. Calculate the hue skewness: Calculate the mean μ of the hue value array. H and standard deviation σ H Then calculate the hue shift f6 using the following formula: ; Where H j Let be the hue value of the j-th pixel within the lesion area; μ H σ is the mean of the tonal values ​​of all pixels within the lesion area. H denoted as , where is the standard deviation of the hue values ​​of all pixels within the lesion area, and N is the total number of pixels within the lesion area.

[0083] f6 > 0 indicates that the data is right-skewed, with more pixels having high hue values ​​(such as yellowish or brownish); f6 < 0 indicates that the data is left-skewed, with more pixels having low hue values ​​(such as greenish); f6 close to 0 indicates that the distribution is symmetrical.

[0084] (7) The eigenvalue f7 of the chromaticity dispersion is calculated as follows: 1. Image segmentation and region extraction: Use segmentation techniques (thresholding or machine learning model) to accurately separate the lesion region from the leaf image and obtain the binary mask of the region.

[0085] 2. Color space conversion: Convert the original RGB image to the HSV (or HSL / HSI) color space; 3. Extract the saturation array of lesions: Using a lesion mask, extract the saturation values ​​of all pixels located within the lesion region from the S-channel image to form an array {S1, S2, ..., S...} containing N saturation values. N}; 4. Calculate the chromaticity dispersion: Calculate the mean μ of the saturation value array. S And calculate the chromaticity dispersion f7 according to the following formula: ; The larger the value of f7, the greater the difference between bright and dark areas within the lesion, and the more uneven the color; the smaller the value, the more uniform the color.

[0086] (8) The eigenvalue f8 of the canopy structure volume index is calculated as follows: 1. 3D data acquisition: Use multi-view stereo vision (SfM), lidar (LiDAR) or depth camera to acquire images or point cloud data of crop canopy from different angles.

[0087] 2. 3D Reconstruction and Alignment: Dense 3D point clouds are generated using algorithms such as SfM and then unified into a coordinate system based on the ground.

[0088] 3. Generate Canopy Height Model (CHM): Divide the horizontal plane into a regular grid. For each grid, calculate the maximum value (or a specific quantile) of the height of all point clouds within it as the canopy height of that grid. zij. Ground height can be obtained by fitting the lowest point or by using a filtering algorithm (such as cloth simulation filtering), where zij is the difference between the canopy top height and the ground elevation.

[0089] 4. Numerical Integration: Sum the heights of all grid cells in the CHM, then multiply by the area of ​​a single grid cell. Δx*Δy gives the canopy volume f8: ; The larger the f8 value, the more lush and healthier the plant.

[0090] (9) The characteristic value f9 of the yellowing ratio is calculated as follows: 1. Leaf Region Segmentation: Separate the leaf region from the image, excluding the background (soil, sky, etc.). This can be achieved using the supergreen index, threshold segmentation, or semantic segmentation model to obtain the leaf mask and calculate Aleaf.

[0091] 2. Color Space Conversion and Channel Extraction: Convert the RGB image of the leaf area to the HSV color space. Extract the H (hue) channel. Hue values ​​directly reflect the type of color (e.g., green, yellow, orange).

[0092] 3. Define the hue threshold for "yellowing" and calculate the area: Based on a large number of samples, define the hue range that falls within the yellowing range (e.g., Hlow≤H≤Hhigh). Typically, the Hue value of healthy green is around 35°~90° (depending on the spatial definition), while the typical yellowing hue range may be between 20°~60° or 25°~45°. Within the leaf mask, count the total number of pixels whose hue values ​​fall within this threshold range; this is the Yellow hue.

[0093] 4. Calculate the yellowing ratio f9 using the following formula: :f9= (Ayellow / Aleaf) × 100%; Ayellow: The pixel area of ​​the leaf that is judged to be "yellowed"; Aleaf: Total pixel area of ​​the leaf.

[0094] (10) The characteristic value f10 of the apical wilting index is calculated as follows: 1. Plant segmentation: Separate the plant outline from the background to obtain a binary mask (which can be achieved through color thresholding or semantic segmentation).

[0095] 2. Identify key points: Stem base point B: Take the midpoint of the lowest end of the mask (or the junction of the stem and the ground).

[0096] Actual apex point A: Take the uppermost point of the mask (or the apical growth point).

[0097] Expected highest point E: Determined by taking the typical height H plant of a healthy plant from the base point B and moving vertically upwards, or by taking the highest point of the mask in the vertical direction (if the plant is completely upright).

[0098] 3. Calculate key parameters: Plant height H: The vertical distance (pixel difference) between the base point B and the expected highest point E.

[0099] The vertical distance d is the actual apex point A and the expected highest point E (if A is lower than E, then d = yA - yE; in the image coordinate system, the y-axis is positive downwards).

[0100] 4. Calculate the apical wilting index f10: f10 = (d / Hplant) × 100%; The larger the f10 value, the more severe the drooping at the tip.

[0101] If the number of lesions r is large (r>50), you can set it to only calculate the features of the top 20 lesions with the largest area, i.e., set r=20. The specific number can be adjusted flexibly.

[0102] In another embodiment, image features can be flexibly added or removed, and the calculation method of the feature values ​​of each image feature can be flexibly adjusted according to the actual situation.

[0103] Example 2 An automatic tobacco disease identification model capable of implementing the above methods.

[0104] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. An automatic identification method for tobacco diseases, characterized in that: Includes the following steps: S1. Calculate the probability of disease K infection based on image features: The collected crop images are transmitted to a trained deep learning diagnostic model, which directly outputs the preliminary probability distribution of the symptoms presented in the images belonging to various diseases, denoted as P. k (image); P k (image) represents the probability of crop infection with disease K determined based on image features; S2. Obtain non-image features: Non-image features include meteorological data from the past N days, the current growth stage, and the actual incidence rate of diseases in the past n years. Meteorological data includes daily temperature, humidity, light intensity, wind speed, and precipitation. The growth stage is the number of days after transplanting, with the day of transplanting being the first day. S3. Calculate the non-image feature adjustment factor: Calculate the non-image feature adjustment factor for disease K based on non-image features, including climate matching degree C. k Growth stage matching degree G k (t) and historical disease weight H k ; S4. Calculate the comprehensive adjustment factor T k : T k = β 1k *C k + β 2k *G k (t) + β 3k *H k ; and β 1k +β 2k +β 3k = 1; Where β 1k β 2k β 3k These represent the weights of climate, growth stage, and historical disease transfer for disease K; β 1k β 2k β 3k All are greater than or equal to 0; S5. Calculate the probability of onset P. K : ； S6. According to the disease incidence probability P K Sort the diseases from largest to smallest, output the probability and type of the top N diseases, and display the corresponding image features and non-image features.

2. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: In step S3, the climate matching degree C k The calculation process is as follows: S31. Calculate the temperature matching degree Mtemp, humidity matching degree Mrh, light matching degree Mlight, wind speed matching degree Mwind, and precipitation matching degree Mrain respectively. S32. Determine the relative weights of temperature, humidity, light intensity, wind speed, and precipitation, and represent them as αtemp, αrh, αlight, αwind, and αrain, respectively. αtemp+αrh+αlight+αwind+αrain=1; S33, Calculate C k C k It is the weighted geometric mean of five climatic factors: temperature, humidity, light intensity, wind speed, and precipitation. C k = (Mtemp^αtemp) × (Mrh^αrh) × (Mlight^αlight) × (Mwind^αwind) ×(Mrain^αrain)。 3. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: The method for calculating the matching degree Gk(t) during the growth stage in step S3 is as follows: ； in: Qk represents the basic matching degree of disease k, 0 ≤ Qk≤0.3; Rk represents the peak susceptibility of disease k, indicating the theoretical maximum matching degree at the optimal time, 0≤Rk≤ 1; t 0,k The center point representing the most susceptible period for disease k indicates the time when disease k is most susceptible after transplanting. 0,k The day is the most likely time for this to happen; the unit is the day. ρk represents the time width parameter of disease k, which controls the rate at which susceptibility decays over time. The unit is days. The larger the value, the longer the period of disease susceptibility. Qk, Rk, t for different diseases K 0,k The values ​​of ρk are set in advance.

4. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: The method for calculating the historical disease weight Hk in step S3 is as follows: ； in: n represents the number of historical years considered; i represents the historical year number, i=1 indicates the most recent year, and i=n indicates the oldest year; λ represents the decay factor, which controls the decay rate of historical influence; 0 <λ≤1, the larger λ is, the more persistent the historical influence is; I k,i The standardized incidence rate of disease k in year i, 0≤I k,i ≤ 1; I k,i =P k,i / P k,max ; P k,i This represents the actual incidence rate of disease k in year i. P k,max This represents the theoretical maximum incidence rate of disease k.

5. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: In step S4, β 1k β 2k β 3k The method for determining the weights is to manually establish weight matrices for different disease types.

6. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: The specific method for step S1 is as follows: S11. Image Acquisition: Acquire raw images of flue-cured tobacco in the field, including overall images of the plant, images of the stem, and images of several leaves from the same plant. S12. Image preprocessing: Preprocess the original images of flue-cured tobacco collected in the field; S13, Calculate P leaf-n(k) For each leaf image, leaf features are extracted, and the probability of disease on each leaf for different diseases K is calculated and denoted as P. leaf-n(k) K represents the disease, and leaf-n represents the nth leaf of the plant. S14, Leaf Evidence Aggregation: Based on the P of several leaves leaf-n(k) The weighted average disease probability P of the leaves was calculated. leaf(k) ; S15, Calculate P stem(k) Based on the plant stem images, stem features are extracted, and the probability P of stem disease in response to different diseases K is calculated. stem(k) ; S16, Calculate P whole(k) Based on the overall image of the plant, extract overall features and calculate the probability P of the plant developing different diseases K. whole(k) ; S17. Based on the calculated P leaf(k) P stem(k) P whole(k) The probability P of plant developing disease K was calculated. k (image) 7. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: The leaf characteristics include at least one of the following: lesion area ratio, distribution dispersion, weighted average roundness, weighted average elongation, weighted average fractal dimension, hue skewness, and hue dispersion; the plant characteristics include at least one of the following: canopy structure volume index, leaf yellowing ratio, and apical wilting index; the stem characteristics include at least one of the following: stem base blackening ratio, stem longitudinal stripe index, and stem base diameter abnormality rate.

8. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: In step S14, P leaf-(k) The calculation method includes the following steps: S141. Calculate the weights Wn of several blades: ， Alesion,n is the area of ​​the nth leaf; Aleaf,n represents the total area of ​​several lesions on the nth leaf; It is a constant; S142, Calculate P leaf(k) : ； in: N represents the total number of leaves analyzed; P leaf-n(k) Let n be the probability of disease K occurring on the nth leaf. Wn is the weight of the nth leaf.

9. The method for automatic identification of tobacco diseases according to claim 1, characterized in that: In step S17, P k The calculation method for (image) includes the following steps: S171. Determine the weights of leaves, stems, and the whole plant, denoted as w respectively. leaf(k) ,w stem(k) ,w whole(k) The method is as follows: S1711. Based on experience, experts pre-determine an initial weight matrix for leaves, stems, and the whole plant for different diseases k; denoted as: w leaf(k)1 ,w stem(k)1 ,w whole(k)1 Satisfying w leaf(k)1 +w stem(k)1 +w whole(k)1 =1; S1712. Preset a threshold range for the probability of disease on leaves, stems, and the whole plant for each disease k. S1713. In actual diagnosis, based on the calculated P... leaf(k) P stem(k) P whole(k) Each component is compared with its respective incidence probability threshold range. If the incidence probability threshold range is exceeded, the initial weight matrix is ​​adjusted to obtain the final weight matrix. The adjustment rule is: if the probability value is higher than the upper limit, it means that the symptoms of that part are very typical, and its weight is increased. If the probability value is below the lower limit, it indicates that the symptoms in that area are not obvious, and its weight should be reduced. S1714. Normalize the adjusted weights to obtain the final weight matrix w for leaves, stems, and the whole plant. leaf(k) ,w stem(k) ,w whole(k) ; S172, Calculate P k (image): 。 10. An automatic identification model for tobacco diseases, characterized in that: The method for automatic identification of tobacco diseases as described in any one of claims 1-9 can be implemented.

Images