A Deep Learning-Based Intelligent Detection System for Rice Seedling Blast Disease

Abstract

This invention discloses a deep learning-based intelligent detection system for rice seedling blast disease, belonging to the field of crop detection technology. It solves the technical problem that early or atypical symptoms of the disease may not be accurately identified in a timely manner, leading to further disease development and greater losses. Based on feature extraction results, the system classifies the disease risk of leaves and panicles, with different control strategies corresponding to different risk levels. The system continuously monitors image data from the growing and harvesting areas, enabling timely detection of subtle changes in leaf and panicle characteristics, effectively reducing yield losses and quality declines caused by the disease. The disease risk classification of leaves and panicles is combined with meteorological data from various management grids and hydrological data from irrigation areas. The system analyzes the risk of external disease conditions and integrates this analysis with the leaf and panicle disease risk classification for comprehensive risk warning.

Description

Technical Field

[0001] This invention belongs to the field of crop detection, specifically a deep learning-based intelligent detection system for rice seedling blast disease. Background Technology

[0002] Rice blast is a major disease of rice, causing significant yield reductions, sometimes as high as 40%-50%, or even complete crop failure. It occurs in various regions, most commonly affecting the leaves and nodes, and can cause varying degrees of yield loss. Early and severe outbreaks of neck blast or node blast can lead to whiteheads and even total crop failure. Traditional disease diagnosis methods often rely on manual experience, which is easily influenced by subjective factors, leading to inaccurate disease identification. Farmers with different levels of experience may have varying symptom assessments of rice seedling blast, and some subtle disease characteristics may be overlooked, thus affecting the diagnostic results.

[0003] Existing rice disease detection systems lack precise extraction and analysis of leaf and panicle characteristics, making it difficult to accurately distinguish between different types and severity of diseases. They also lack multidimensional analysis that combines meteorological and hydrological factors, which may prevent timely and accurate identification of diseases in their early stages or with atypical symptoms, leading to further disease development and greater losses. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a deep learning-based intelligent detection system for rice seedling blast disease, which is used to solve the technical problem that the disease may not be identified in a timely and accurate manner in the early stage or with atypical symptoms, leading to further development of the disease and greater losses.

[0005] To address the above problems, a first aspect of the present invention provides an intelligent detection system for rice seedling blast disease based on deep learning, comprising:

[0006] Data acquisition module: Acquires map data of the rice planting area, divides the rice planting area into harvest area, growth area, irrigation area and auxiliary area, and sets up management grids of different densities. Each management grid is equipped with a UAV data acquisition module and a land data acquisition module.

[0007] Regional image processing nodes: Based on the location of the management grid, regional image processing nodes are set up in the areas where several adjacent management grids are located. The regional image processing nodes receive data collected by the data acquisition modules in the surrounding management grids and obtain meteorological data of each management grid from the authorized meteorological center.

[0008] A primary image screening model is established at the regional image processing node to initially classify images into risk images and safe images. The risk images and corresponding grid data are then sent to the cloud image processing platform.

[0009] Cloud-based image processing platform: Extracts leaf morphology features and lesion features from image data of the growing area, and classifies leaf disease risk.

[0010] Based on the image data of the harvest area, the morphological features and abnormal features of the ear are extracted, and the risk of ear disease is classified.

[0011] Based on meteorological data from each management grid and hydrological data from the irrigation area, the risk of external disease conditions in each grid is analyzed. Combined with the risk classification of leaf disease and ear disease, risk warnings are issued for the regional grid.

[0012] Optionally, in one example of the above aspects, management grids of different densities are set up, with each management grid equipped with a UAV data acquisition module and a land data acquisition module, including the following steps:

[0013] A 100m×100m density grid was set up in the harvest area, a 200m×200m density grid was set up in the growth area, and a hydrological monitoring module was set up in the irrigation area.

[0014] Each management grid is equipped with a drone data acquisition module and a land data acquisition module. The drone data acquisition module collects high-definition image data and infrared image data of rice leaves or panicles; the land data acquisition module collects soil temperature and humidity, as well as high-definition image data and infrared image data of plant leaves.

[0015] Optionally, in one example of the above aspects, a preliminary image screening model is established at the regional image processing node to initially classify images into risk images and safe images. The risk images and their corresponding grid data are then sent to the cloud image processing platform, including the following steps:

[0016] Historical data of multi-angle high-definition images and infrared images of rice leaves or panicles were acquired. Image data of rice leaves or panicles at different times in areas where seedling blast occurred were labeled as risk images, and image data of rice leaves or panicles at different times in areas where seedling blast did not occur were labeled as safe images.

[0017] By training a deep learning model with labeled images at regional image processing nodes, a primary image screening model is established. The trained primary image screening model will initially classify images into risk images and safe images.

[0018] Safe images are stored locally, while risk images and meteorological and soil data from the grid are sent to a cloud-based image processing platform.

[0019] Optionally, in one example of the above aspects, leaf morphology features and lesion features are extracted from the image data of the growing area, and leaf disease risk is classified, including the following steps:

[0020] For image data of the growth area, pixel-level masks of lesions are obtained by DeepLabV3+ segmentation model. Combined with the area of ​​leaf detection box, the area ratio of lesions is calculated. Noise areas in the detection box are excluded by IoU threshold filtering. The aspect ratio, roundness and compactness feature data of lesions are extracted for joint morphological feature analysis. Leaf disease risk is classified by combining leaf morphological features.

[0021] Optionally, in one example of the above aspects, the pixel-level mask of the lesion is obtained through the DeepLabV3+ segmentation model, and the lesion area ratio is calculated by combining it with the area of ​​the leaf detection box, including the following steps:

[0022] High-resolution images of rice leaves were collected, including images of healthy leaves and images of leaves with different degrees of seedling blast disease. The images were annotated to mark leaf areas and lesion areas, and corresponding segmentation masks and detection boxes were generated.

[0023] DeepLabV3+ is trained as a segmentation model using processed image data. The trained DeepLabV3+ model is then applied to image data collected in the growth region to generate pixel-level masks for lesions.

[0024] Calculate the lesion area percentage based on segmentation mask: Where Ar is the percentage of lesion area, I(pi) is the indicator function, which is 1 if the predicted mask pixel value is greater than 0.5, otherwise it is 0, i∈(1,2,…,Nleaf); Nleaf is the total number of pixels in the leaf area.

[0025] Optionally, in one example of the above aspects, the aspect ratio, roundness, and compactness of lesions are extracted for joint morphological feature analysis, and leaf disease risk is classified in combination with leaf morphological features, including the following steps:

[0026] Extracting aspect ratio feature data of lesions: Extract the side length of the minimum bounding rectangle of the lesion region. The aspect ratio feature value = length of the short side of the rectangle / length of the long side of the rectangle;

[0027] Extract the roundness feature data of the lesion: Roundness feature value = [4*π*(lesion area)] / (lesion area perimeter²);

[0028] Extract the compactness feature data of the lesion: compactness feature value = lesion area / area of ​​the smallest bounding rectangle of the lesion area;

[0029] In the historical data screening, after detecting lesions with an area smaller than the threshold in the grid, and without the lesions spreading to larger areas, the aspect ratio, roundness, and compactness feature values ​​of the corresponding lesion data are calculated, and the mean of the corresponding feature values ​​is taken as the standard value.

[0030] Calculate the degradation ratio of aspect ratio, roundness, and compactness features of lesion data in the image data collected from the growth area compared with the standard value;

[0031] Obtain the percentage of lesion area on leaves corresponding to the grid area, and combine this data with the deterioration ratio of aspect ratio, roundness, and compactness characteristics compared to the standard value to classify the leaf disease risk: ;

[0032] Where Grade is the leaf disease risk grading coefficient, Sf1 is the proportion of the aspect ratio characteristic value deteriorated compared with the standard value, Sf2 is the proportion of the roundness characteristic value deteriorated compared with the standard value, Sf3 is the proportion of the compactness characteristic value deteriorated compared with the standard value, and w1, w2, w3 and w4 are weighting coefficients.

[0033] Leaf disease risk is classified according to the leaf disease risk grading coefficient.

[0034] Optionally, in one example of the above aspects, for image data of the harvest area, the following steps are performed to extract ear morphological features and ear abnormality features, and to classify ear disease risk:

[0035] For image data from the harvest area, a loss function for ear detection is constructed. The ear region is detected by the YOLOv8 target detection model, and ear morphological features are extracted. The ear structure is segmented by Mask R-CNN, and ear length, ear density, and abnormal ear region features are extracted. A joint evaluation function of ear-lesion is constructed to classify ear disease risk.

[0036] Optionally, in one example of the above aspects, a loss function for ear detection is constructed, and the ear region is detected using a YOLOv8 object detection model to extract ear morphological features, including the following steps:

[0037] For abnormal areas in the ear, an ear detection loss function is constructed, and the ear region image is extracted using a YOLOv8 object detection model: Where L represents the boundary regression accuracy of the extracted ear region image, and IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box. Let be the Euclidean distance between the center point of the predicted box and the center point of the ground truth box, c be the length of the diagonal of the minimum bounding rectangle between the predicted box and the ground truth box, δ be the weight coefficient, v be the aspect ratio consistency coefficient, v=σ1(1−r)+σ2(1−C)+σ3(1−K), where r is the aspect ratio feature value, C is the aspect ratio feature value, K is the compactness feature value, and σ1, σ2 and σ3 are the corresponding weights;

[0038] The morphological features of the ear are extracted, including ear length feature value and ear curvature feature value. The morphological features of abnormal areas of the ear are extracted, including abnormal area length feature value and texture analysis feature value. The texture analysis value is the feature value obtained by weighted averaging of four types of texture feature values: contrast, energy, homogeneity and correlation.

[0039] Optionally, in one example of the above aspects, a combined ear-lesion assessment function is constructed to classify the disease risk of the ear, including the following steps:

[0040] Construct a morphological evaluation function for the ear of grain: Wherein, Lm is the ear morphology evaluation coefficient, Lsp is the ear length characteristic value, Lsp0 is the ear length standard value, Ksp is the ear curvature characteristic value, Ksp0 is the ear curvature standard value, α1 is the corresponding weight of ear length, and α2 is the corresponding weight of ear curvature.

[0041] Construct a lesion feature evaluation function: Wherein, Le is the evaluation coefficient of ear lesion characteristics, Lse is the characteristic value of abnormal region length, Lse0 is the standard value of abnormal region length, Kc is the characteristic value of lesion texture analysis, Kc0 is the standard value of lesion texture analysis, β1 is the corresponding weight of abnormal region length, and β2 is the corresponding weight of lesion texture analysis.

[0042] Based on the calculation formulas for the panicle morphology evaluation coefficient and the panicle lesion characteristic evaluation coefficient, a joint evaluation function for panicle and lesion is established: Wherein, Ljoi is the combined evaluation coefficient of ear-lesion, γ1 is the corresponding weight of the ear morphology evaluation coefficient, and γ2 is the corresponding weight of the ear lesion characteristic evaluation coefficient.

[0043] The disease risk of the ear is classified according to the combined assessment coefficient of ear and lesion.

[0044] Optionally, in one example of the above aspects, based on meteorological data of each management grid and hydrological data of the irrigation area, the risk of external disease conditions for each grid is analyzed. Combining leaf disease risk classification and ear disease risk classification, a risk warning is issued for the regional grid, including the following steps:

[0045] Acquire historical data, including meteorological data for each management grid and hydrological data for irrigation areas when seedling blight occurred and when the region did not experience seedling blight, and label the seedling blight risk data and normal data.

[0046] A spatiotemporal Transformer model is trained using labeled data. The trained model is then used to identify whether the meteorological data of each management grid and the hydrological data of the irrigation area are seedling blight risk data or normal data.

[0047] If the leaf disease risk classification or ear disease risk classification of the grid shows the most severe disease risk classification, or if the meteorological data of the management grid and the hydrological data of the irrigation area of ​​the management grid are identified as seedling blast disease risk data, then a risk warning will be issued for the corresponding grid.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] This invention classifies the disease risk of leaves and panicles based on feature extraction results, with different control strategies corresponding to different risk levels. The system continuously monitors image data from the growing and harvesting areas, enabling timely detection of subtle changes in leaf and panicle characteristics. Early warning prevents further spread and diffusion of diseases, effectively reducing yield losses and quality decline caused by diseases. Taking rice seedling blast as an example, if not controlled in time, it may lead to plant death and an increase in empty grains in the panicles, seriously affecting rice yield and quality. Through early warning and intervention, losses can be controlled to a minimum.

[0050] This invention combines leaf and panicle disease risk grading with meteorological data from various management grids and hydrological data from irrigation areas. The system analyzes the risk of external disease conditions by integrating meteorological data from each management grid and hydrological data from irrigation areas, and combines this analysis with the leaf and panicle disease risk grading for comprehensive risk early warning. Factors such as temperature, humidity, and light intensity in meteorological data, as well as soil moisture and irrigation water volume in hydrological data, are closely related to the occurrence and development of rice seedling blast disease. Through multi-source data fusion analysis, the risk of disease occurrence can be assessed more comprehensively and accurately, improving the reliability of early warning. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0053] Figure 2 This is a schematic diagram of the intelligent detection of rice seedling blast disease according to the present invention. Detailed Implementation

[0054] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0055] See also Figure 1-Figure 2 The first aspect of this invention provides a deep learning-based intelligent detection system for rice seedling blast disease, comprising:

[0056] Data acquisition module: Acquires map data of the rice planting area, divides the rice planting area into harvest area, growth area, irrigation area and auxiliary area, and sets up management grids of different densities. Each management grid is equipped with a UAV data acquisition module and a land data acquisition module.

[0057] Regional image processing nodes: Based on the location of the management grid, regional image processing nodes are set up in the areas where several adjacent management grids are located. The regional image processing nodes receive data collected by the data acquisition modules in the surrounding management grids and obtain meteorological data of each management grid from the authorized meteorological center.

[0058] A primary image screening model is established at the regional image processing node to initially classify images into risk images and safe images. The risk images and corresponding grid data are then sent to the cloud image processing platform.

[0059] Cloud-based image processing platform: Extracts leaf morphology features and lesion features from image data of the growing area, and classifies leaf disease risk.

[0060] Based on the image data of the harvest area, the morphological features and abnormal features of the ear are extracted, and the risk of ear disease is classified.

[0061] Based on meteorological data from each management grid and hydrological data from the irrigation area, the risk of external disease conditions in each grid is analyzed. Combined with the risk classification of leaf disease and ear disease, risk warnings are issued for the regional grid.

[0062] In this embodiment, the rice planting area is subdivided into a harvesting area, a growing area, an irrigation area, and an auxiliary area. This refined division of areas allows for more targeted management of different functional areas.

[0063] The different densities of the management grid can be flexibly adjusted according to the actual conditions of the rice-growing area, such as topography, landforms, and planting density. In key monitoring areas or areas with complex terrain, a denser grid can be set up to obtain more detailed and accurate data.

[0064] Each management grid is equipped with both drone data acquisition modules and land-based data acquisition modules, forming a three-dimensional data acquisition network. Drones can quickly scan large rice-growing areas from the air, acquiring macroscopic images and data, enabling timely detection of large-scale disease outbreaks or abnormal growth conditions.

[0065] Based on the location of the management grid, regional image processing nodes are set up in the areas where several adjacent management grids are located. This distributed data processing architecture can effectively improve data processing efficiency. Each regional image processing node only needs to process the data collected by the surrounding management grids, avoiding network congestion and processing delays caused by the centralized transmission of large amounts of data to the central node for processing. At the same time, multiple regional image processing nodes can work in parallel, further accelerating the data processing speed and enabling the system to analyze and diagnose the collected data in a timely manner.

[0066] The distributed configuration of regional image processing nodes gives the system excellent scalability. When the rice planting area expands or data acquisition needs increase, new regional image processing nodes can be easily added without requiring large-scale modifications to the entire system. Simultaneously, this distributed architecture also improves the system's fault tolerance. If a regional image processing node fails, it will only affect the data processing of the surrounding management grids it is responsible for, preventing the entire system from crashing. Other nodes can still operate normally, ensuring the system's stability and reliability.

[0067] Based on the image data of the growing area, leaf morphology features and lesion features are extracted from the image data to classify the leaf disease risk.

[0068] Based on the image data of the harvest area, the morphological features and abnormal features of the ear are extracted, and the risk of ear disease is classified.

[0069] Based on meteorological data from each management grid and hydrological data from the irrigation area, the risk of external disease conditions in each grid is analyzed. Combined with the risk classification of leaf disease and ear disease, risk warnings are issued for the regional grid.

[0070] Extracting leaf morphology and lesion features from images of the growing area enables precise identification of leaf changes caused by different types of diseases. For example, rice blast causes spindle-shaped lesions on leaves; feature extraction can accurately capture this characteristic, distinguishing it from other similar diseases. Extracting panicle morphology and abnormal features from the harvest area can identify potential problems such as panicle discoloration and malformation, helping to comprehensively understand the health status of rice at each stage of growth.

[0071] Disease risk classification of leaves and ears based on feature extraction results provides a scientific basis for disease control. Different risk levels correspond to different control strategies. For example, in high-risk areas, priority can be given to allocating control personnel and resources, and using highly effective pesticides for targeted control to improve control effectiveness and reduce disease losses.

[0072] The system continuously monitors image data from the growing and harvesting areas, enabling timely detection of subtle changes in leaf and panicle characteristics. Early warning prevents further spread of diseases, effectively reducing yield losses and quality degradation caused by diseases. Taking rice seedling blast as an example, if not controlled in time, it can lead to plant death and an increase in empty grains in the panicle, severely impacting rice yield and quality. Through early warning and intervention, losses can be minimized.

[0073] Based on the risk classification of diseases on leaves and ears, it is easier to rationally allocate human, material, and financial resources in the later stages. For high-risk areas, resources are concentrated on key prevention and control; for low-risk areas, resource input is appropriately reduced. The risk early warning mechanism enables farmers to conduct targeted field management, avoiding the waste of time and energy caused by comprehensive inspections and indiscriminate prevention and control. Managers can quickly locate problem areas based on early warning information and take corresponding management measures, improving management efficiency and work effectiveness.

[0074] The disease risk grading for leaves and panicles combines meteorological data from each management grid and hydrological data from the irrigation area. The system analyzes the risk of external disease conditions by integrating this data with the leaf and panicle disease risk grading for comprehensive risk early warning. Factors such as temperature, humidity, and light intensity in meteorological data, as well as soil moisture and irrigation volume in hydrological data, are closely related to the occurrence and development of rice seedling blast disease. Through multi-source data fusion analysis, the risk of disease occurrence can be assessed more comprehensively and accurately, improving the reliability of early warnings.

[0075] In one embodiment of the present invention, management grids of different densities are set up, and each management grid is equipped with a UAV data acquisition module and a land data acquisition module, including the following steps:

[0076] A 100m×100m density grid was set up in the harvest area, a 200m×200m density grid was set up in the growth area, and a hydrological monitoring module was set up in the irrigation area.

[0077] Each management grid is equipped with a drone data acquisition module and a land data acquisition module. The drone data acquisition module collects high-definition image data and infrared image data of rice leaves or panicles; the land data acquisition module collects soil temperature and humidity, as well as high-definition image data and infrared image data of plant leaves.

[0078] In one embodiment of the present invention, a preliminary image screening model is established at a regional image processing node to initially classify images into risk images and safe images. The risk images and corresponding grid data are then sent to a cloud image processing platform, including the following steps:

[0079] Historical data of multi-angle high-definition images and infrared images of rice leaves or panicles were acquired. Image data of rice leaves or panicles at different times in areas where seedling blast occurred were labeled as risk images, and image data of rice leaves or panicles at different times in areas where seedling blast did not occur were labeled as safe images.

[0080] By training a deep learning model with labeled images at regional image processing nodes, a primary image screening model is established. The trained primary image screening model will initially classify images into risk images and safe images.

[0081] Safe images are stored locally, while risk images and meteorological and soil data from the grid are sent to a cloud-based image processing platform.

[0082] In one embodiment of the present invention, leaf morphology features and lesion features are extracted from image data of the growth area, and leaf disease risk is classified, including the following steps:

[0083] For image data of the growth area, pixel-level masks of lesions are obtained by DeepLabV3+ segmentation model. Combined with the area of ​​leaf detection box, the area ratio of lesions is calculated. By using IoU threshold filtering, noise areas (such as leaf edges and light reflection) in the detection box are excluded. The aspect ratio, roundness and compactness feature data of lesions are extracted for joint morphological feature analysis. Combined with leaf morphological features, leaf disease risk is classified.

[0084] In one embodiment of the present invention, the pixel-level mask of the lesion is obtained by using the DeepLabV3+ segmentation model, and the lesion area ratio is calculated by combining the area of ​​the leaf detection box, including the following steps:

[0085] High-resolution images of rice leaves were collected, including images of healthy leaves and images of leaves with different degrees of seedling blast disease. The images were annotated to mark leaf areas and lesion areas, and corresponding segmentation masks and detection boxes were generated.

[0086] DeepLabV3+ is trained as a segmentation model using processed image data. The trained DeepLabV3+ model is then applied to image data collected in the growth region to generate pixel-level masks for lesions.

[0087] Calculate the lesion area percentage based on segmentation mask: Where Ar is the percentage of lesion area, I(pi) is the indicator function, which is 1 if the predicted mask pixel value is greater than 0.5, otherwise it is 0, i∈(1,2,…,Nleaf); Nleaf is the total number of pixels in the leaf area.

[0088] In one embodiment of the present invention, the aspect ratio, roundness, and compactness of lesions are extracted for joint morphological feature analysis, and leaf disease risk is classified in combination with leaf morphological features, including the following steps:

[0089] Extracting aspect ratio feature data of lesions: Extract the side length of the minimum bounding rectangle of the lesion region. The aspect ratio feature value = length of the short side of the rectangle / length of the long side of the rectangle;

[0090] Extract the roundness feature data of the lesion: Roundness feature value = [4*π*(lesion area)] / (lesion area perimeter²). The closer the roundness feature value is to 1, the closer the shape of the lesion is to a circle.

[0091] Extract the compactness feature data of the lesion: compactness feature value = lesion area / area of ​​the smallest bounding rectangle of the lesion area. The closer the compactness feature value is to 1, the more compact the shape of the lesion.

[0092] In the historical data screening, after detecting lesions with an area smaller than the threshold of 5m×5m in the grid, and without the lesions spreading to larger areas, the aspect ratio, roundness, and compactness feature values ​​of the corresponding lesion data were calculated, and the mean of the corresponding feature values ​​was taken as the standard value.

[0093] Calculate the degradation ratio of aspect ratio, roundness, and compactness features of lesion data in the image data collected from the growth area compared with the standard value;

[0094] Based on a large amount of experimental data, the closer the lesion shape is to a circle, the more compact the lesion shape is, and the closer the length-to-width ratio is, the less likely the lesion is to spread. Therefore, when calculating the deterioration ratio, the deterioration ratio of the corresponding characteristic value to the standard value is = (standard value - corresponding characteristic value) / standard value.

[0095] Obtain the percentage of lesion area on leaves corresponding to the grid area, and combine this data with the deterioration ratio of aspect ratio, roundness, and compactness characteristics compared to the standard value to classify the leaf disease risk: ;

[0096] Among them, Grade is the leaf disease risk grading coefficient, Sf1 is the proportion of deterioration of the aspect ratio characteristic value compared with the standard value, Sf2 is the proportion of deterioration of the roundness characteristic value compared with the standard value, Sf3 is the proportion of deterioration of the compactness characteristic value compared with the standard value, and w1, w2, w3 and w4 are weighting coefficients, which were determined through experiments.

[0097] Leaf disease risk is classified according to the leaf disease risk grading coefficient.

[0098] In this embodiment, the standard value is calculated from the filtered data in the historical data, and the proportion of lesion area in the corresponding region is obtained. The leaf disease risk grading coefficient is calculated using the leaf disease risk grading coefficient calculation formula, and the obtained value is used as the first leaf disease risk threshold.

[0099] After screening historical data grids, lesions with an area larger than the threshold of 50m×50m were detected and the lesion data showed expansion and spread. The aspect ratio, roundness and compactness characteristic values ​​of the corresponding lesion data were calculated, and the mean of the corresponding characteristic values ​​was taken to calculate the leaf disease risk classification coefficient. The obtained value was used as the second leaf disease risk threshold.

[0100] When the leaf disease risk grading coefficient is greater than or equal to the second leaf disease risk threshold, it is determined to be a first-level leaf disease risk level.

[0101] When the leaf disease risk grading coefficient is greater than or equal to the first leaf disease risk threshold and less than the second leaf disease risk threshold, it is determined to be a level three leaf disease risk level.

[0102] When the leaf disease risk grading coefficient is less than the first leaf disease risk threshold, it is determined to be a third-level leaf disease risk level.

[0103] In one embodiment of the present invention, for image data of the harvest area, the following steps are performed to extract ear morphological features and ear abnormality features, and to classify ear disease risk:

[0104] For image data from the harvest area, a loss function for ear detection is constructed. The ear region is detected by the YOLOv8 target detection model, and ear morphological features are extracted. The ear structure is segmented by Mask R-CNN, and ear length, ear density, and abnormal ear region features are extracted. A joint evaluation function of ear-lesion is constructed to classify ear disease risk.

[0105] In one embodiment of the present invention, a loss function for ear detection is constructed, and the ear region is detected using a YOLOv8 object detection model to extract ear morphological features, including the following steps:

[0106] For abnormal areas in the ear, an ear detection loss function is constructed, and the ear region image is extracted using a YOLOv8 object detection model: Where L represents the boundary regression accuracy of the extracted ear region image, and IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box. Let c be the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, c be the length of the diagonal of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box, δ be the weight coefficient, δ=v / [(1−IoU)+v], v be the aspect ratio consistency coefficient, v=σ1(1−r)+σ2(1−C)+σ3(1−K), r be the aspect ratio feature value, C be the aspect ratio feature value, K be the compactness feature value, and σ1, σ2 and σ3 be the corresponding weights;

[0107] The morphological features of the ear are extracted, including ear length feature value and ear curvature feature value. The morphological features of abnormal areas of the ear are extracted, including abnormal area length feature value and texture analysis feature value. The texture analysis value is the feature value obtained by weighted averaging of four types of texture feature values: contrast, energy, homogeneity and correlation.

[0108] In this embodiment, morphological feature extraction:

[0109] The ear length feature value = the diagonal length of the ear detection frame × pixel resolution;

[0110] Where (x1,y1) are the base coordinates of the ear, (x3,y3) are the tip coordinates of the ear, and (x2,y2) are the midpoint coordinates of the ear.

[0111] In one embodiment of the present invention, a combined assessment function of ear and lesion is constructed to classify the disease risk of the ear, including the following steps:

[0112] Construct a morphological evaluation function for the ear of grain: Wherein, Lm is the ear morphology evaluation coefficient, Lsp is the ear length characteristic value, Lsp0 is the ear length standard value, Ksp is the ear curvature characteristic value, Ksp0 is the ear curvature standard value, α1 is the corresponding weight of ear length, and α2 is the corresponding weight of ear curvature.

[0113] Construct a lesion feature evaluation function: Wherein, Le is the evaluation coefficient of ear lesion characteristics, Lse is the characteristic value of abnormal region length, Lse0 is the standard value of abnormal region length, Kc is the characteristic value of lesion texture analysis, Kc0 is the standard value of lesion texture analysis, β1 is the corresponding weight of abnormal region length, and β2 is the corresponding weight of lesion texture analysis.

[0114] Based on the calculation formulas for the panicle morphology evaluation coefficient and the panicle lesion characteristic evaluation coefficient, a joint evaluation function for panicle and lesion is established: Wherein, Ljoi is the combined evaluation coefficient of ear-lesion, γ1 is the corresponding weight of the ear morphology evaluation coefficient, and γ2 is the corresponding weight of the ear lesion characteristic evaluation coefficient.

[0115] The disease risk of the ear is classified according to the combined assessment coefficient of ear and lesion.

[0116] In this embodiment, γ1 and γ2 are set to 0.4 and 0.6, β1 and β2 are set to 0.3 and 0.7, and α1 and α2 are set to 0.4 and 0.6.

[0117] When all feature values ​​are standard values ​​in historical data, the disease risk grading coefficient of the panicle is 0, and 0 is used as the first disease risk threshold of the panicle.

[0118] After screening historical data grids, lesion data that are detected in areas larger than the threshold of 50m×50m and show signs of expansion and spread are calculated. The corresponding feature values ​​are then taken, and the mean of the corresponding feature values ​​is used to calculate the disease risk grading coefficient of the ear. The obtained value is used as the second disease risk threshold of the ear.

[0119] When the disease risk grading coefficient of the ear is greater than or equal to the second ear disease risk threshold, it is determined to be a first-level ear disease risk level.

[0120] When the disease risk grading coefficient of the ear is greater than or equal to the first ear disease risk threshold and less than the second ear disease risk threshold, it is determined to be a level three ear disease risk level.

[0121] When the disease risk grading coefficient of the ear is less than the first ear disease risk threshold, it is determined to be a level three ear disease risk level.

[0122] In one embodiment of the present invention, based on meteorological data of each management grid and hydrological data of the irrigation area, the risk of external disease conditions for each grid is analyzed. Combining leaf disease risk classification and ear disease risk classification, a risk warning is issued for the regional grid, including the following steps:

[0123] Acquire historical data, including meteorological data for each management grid and hydrological data for irrigation areas when seedling blight occurred and when the region did not experience seedling blight, and label the seedling blight risk data and normal data.

[0124] A spatiotemporal Transformer model is trained using labeled data. The trained model is then used to identify whether the meteorological data of each management grid and the hydrological data of the irrigation area are seedling blight risk data or normal data.

[0125] If the leaf disease risk classification or ear disease risk classification of the grid shows the most severe disease risk classification, or if the meteorological data of the management grid and the hydrological data of the irrigation area of ​​the management grid are identified as seedling blast disease risk data, then a risk warning will be issued for the corresponding grid.

[0126] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A deep learning-based intelligent detection system for rice seedling blast disease, characterized in that, include: Data acquisition module: Acquires map data of the rice planting area, divides the rice planting area into harvest area, growth area, irrigation area and auxiliary area, and sets up management grids of different densities. Each management grid is equipped with a UAV data acquisition module and a land data acquisition module. Regional image processing nodes: Based on the location of the management grid, regional image processing nodes are set up in the areas where several adjacent management grids are located. The regional image processing nodes receive data collected by the data acquisition modules in the surrounding management grids and obtain meteorological data of each management grid from the authorized meteorological center. A primary image screening model is established at the regional image processing node to initially classify images into risk images and safe images. The risk images and corresponding grid data are then sent to the cloud image processing platform. Cloud-based image processing platform: Extracts leaf morphology features and lesion features from image data of the growing area, and classifies leaf disease risk. Based on the image data of the harvest area, the morphological features and abnormal features of the ear are extracted, and the risk of ear disease is classified. Based on meteorological data from each management grid and hydrological data from the irrigation area, the external disease risk of each grid is analyzed. Combining leaf disease risk grading and ear disease risk grading, risk warnings are issued for the regional grids, including: A loss function for ear detection is constructed, and the ear region is detected using the YOLOv8 object detection model to extract ear morphological features, including the following steps: For abnormal areas in the ear, an ear detection loss function is constructed, and the ear region image is extracted using a YOLOv8 object detection model: Where L represents the boundary regression accuracy of the extracted ear region image, and IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box. Let be the Euclidean distance between the center point of the predicted box and the center point of the ground truth box, c be the length of the diagonal of the minimum bounding rectangle between the predicted box and the ground truth box, δ be the weight coefficient, v be the aspect ratio consistency coefficient, v=σ1(1−r)+σ2(1−C)+σ3(1−K), where r is the aspect ratio feature value, C is the aspect ratio feature value, K is the compactness feature value, and σ1, σ2 and σ3 are the corresponding weights; Ear morphological features were extracted, including ear length feature values ​​and ear curvature feature values. Abnormal ear morphological features were extracted, including abnormal region length feature values ​​and texture analysis feature values. The texture analysis value was obtained by weighted averaging of four types of texture feature values: contrast, energy, homogeneity, and correlation. Constructing a combined assessment function for ear and lesion risk classification of ear disease includes the following steps: Construct a morphological evaluation function for the ear of grain: Wherein, Lm is the ear morphology evaluation coefficient, Lsp is the ear length characteristic value, Lsp0 is the ear length standard value, Ksp is the ear curvature characteristic value, Ksp0 is the ear curvature standard value, α1 is the corresponding weight of ear length, and α2 is the corresponding weight of ear curvature. Construct a lesion feature evaluation function: Wherein, Le is the evaluation coefficient of ear lesion characteristics, Lse is the characteristic value of abnormal region length, Lse0 is the standard value of abnormal region length, Kc is the characteristic value of lesion texture analysis, Kc0 is the standard value of lesion texture analysis, β1 is the corresponding weight of abnormal region length, and β2 is the corresponding weight of lesion texture analysis. Based on the calculation formulas for the panicle morphology evaluation coefficient and the panicle lesion characteristic evaluation coefficient, a joint evaluation function for panicle and lesion is established: Wherein, Ljoi is the combined evaluation coefficient of ear-lesion, γ1 is the corresponding weight of the ear morphology evaluation coefficient, and γ2 is the corresponding weight of the ear lesion characteristic evaluation coefficient. The disease risk of the ear is classified according to the combined assessment coefficient of ear and lesion.

2. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 1, characterized in that, Set up management grids of different densities, with each grid equipped with a UAV data acquisition module and a land data acquisition module, including the following steps: A 100m×100m density grid was set up in the harvest area, a 200m×200m density grid was set up in the growth area, and a hydrological monitoring module was set up in the irrigation area. Each management grid is equipped with a drone data acquisition module and a land data acquisition module. The drone data acquisition module collects high-definition image data and infrared image data of rice leaves or panicles; the land data acquisition module collects soil temperature and humidity, as well as high-definition image data and infrared image data of plant leaves.

3. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 1, characterized in that, A preliminary image screening model is established at the regional image processing node to initially classify images into risk images and safe images. The risk images and their corresponding grid data are then sent to the cloud image processing platform, including the following steps: Historical data of multi-angle high-definition images and infrared images of rice leaves or panicles were acquired. Image data of rice leaves or panicles at different times in areas where seedling blast occurred were labeled as risk images, and image data of rice leaves or panicles at different times in areas where seedling blast did not occur were labeled as safe images. By training a deep learning model with labeled images at regional image processing nodes, a primary image screening model is established. The trained primary image screening model will initially classify images into risk images and safe images. Safe images are stored locally, while risk images and meteorological and soil data from the grid are sent to a cloud-based image processing platform.

4. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 1, characterized in that, Based on the image data of the growing area, leaf morphology features and lesion features are extracted, and leaf disease risk is classified, including the following steps: For image data of the growth area, pixel-level masks of lesions are obtained by DeepLabV3+ segmentation model. Combined with the area of ​​leaf detection box, the area ratio of lesions is calculated. Noise areas in the detection box are excluded by IoU threshold filtering. The aspect ratio, roundness and compactness feature data of lesions are extracted for joint morphological feature analysis. Leaf disease risk is classified by combining leaf morphological features.

5. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 4, characterized in that, The pixel-level mask of the lesion is obtained using the DeepLabV3+ segmentation model. Combined with the area of ​​the leaf detection box, the lesion area ratio is calculated, including the following steps: High-resolution images of rice leaves were collected, including images of healthy leaves and images of leaves with different degrees of seedling blast disease. The images were annotated to mark leaf areas and lesion areas, and corresponding segmentation masks and detection boxes were generated. DeepLabV3+ is trained as a segmentation model using processed image data. The trained DeepLabV3+ model is then applied to image data collected in the growth region to generate pixel-level masks for lesions. Calculate the lesion area percentage based on segmentation mask: Where Ar is the percentage of lesion area, I(pi) is the indicator function, which is 1 if the predicted mask pixel value is greater than 0.5, otherwise it is 0, i∈(1,2,…,Nleaf); Nleaf is the total number of pixels in the leaf area.

6. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 4, characterized in that, The aspect ratio, roundness, and compactness of lesions were extracted for joint morphological feature analysis. Combined with leaf morphological features, leaf disease risk was classified, including the following steps: Extracting aspect ratio feature data of lesions: Extract the side length of the minimum bounding rectangle of the lesion region. The aspect ratio feature value = length of the short side of the rectangle / length of the long side of the rectangle; Extract the roundness feature data of the lesion: Roundness feature value = [4*π*(lesion area)] / (lesion area perimeter²); Extract the compactness feature data of the lesion: compactness feature value = lesion area / area of ​​the smallest bounding rectangle of the lesion area; In the historical data screening, after detecting lesions with an area smaller than the threshold in the grid, and without the lesions spreading to larger areas, the aspect ratio, roundness, and compactness feature values ​​of the corresponding lesion data are calculated, and the mean of the corresponding feature values ​​is taken as the standard value. Calculate the degradation ratio of aspect ratio, roundness, and compactness features of lesion data in the image data collected from the growth area compared with the standard value; Obtain the percentage of lesion area on leaves corresponding to the grid area, and combine this data with the deterioration ratio of aspect ratio, roundness, and compactness characteristics compared to the standard value to classify the leaf disease risk: Where Grade is the leaf disease risk grading coefficient, Sf1 is the proportion of the aspect ratio characteristic value deteriorated compared with the standard value, Sf2 is the proportion of the roundness characteristic value deteriorated compared with the standard value, Sf3 is the proportion of the compactness characteristic value deteriorated compared with the standard value, and w1, w2, w3 and w4 are weighting coefficients. Leaf disease risk is classified according to the leaf disease risk grading coefficient.

7. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 1, characterized in that, For image data from the harvest area, the following steps are performed to extract ear morphological features and ear abnormality features, and to classify ear disease risk: For image data from the harvest area, a loss function for ear detection is constructed. The ear region is detected by the YOLOv8 target detection model, and ear morphological features are extracted. The ear structure is segmented by Mask R-CNN, and ear length, ear density, and abnormal ear region features are extracted. A joint evaluation function of ear-lesion is constructed to classify ear disease risk.

8. The intelligent detection system for rice seedling blast disease based on deep learning according to claim 1, characterized in that, Based on meteorological data from each management grid and hydrological data from the irrigation area, the risk of external disease conditions for each grid is analyzed. Combining leaf disease risk classification and ear disease risk classification, risk warnings are issued for the regional grid, including the following steps: Acquire historical data, including meteorological data for each management grid and hydrological data for irrigation areas when seedling blight occurred and when the region did not experience seedling blight, and label the seedling blight risk data and normal data. A spatiotemporal Transformer model is trained using labeled data. The trained model is then used to identify whether the meteorological data of each management grid and the hydrological data of the irrigation area are seedling blight risk data or normal data. If the leaf disease risk classification or ear disease risk classification of the grid shows the most severe disease risk classification, or if the meteorological data of the management grid and the hydrological data of the irrigation area of ​​the management grid are identified as seedling blast disease risk data, then a risk warning will be issued for the corresponding grid.

Images