Crop disease severity dynamic estimation method and system based on multi-scale fusion

Abstract

The invention discloses a crop disease severity dynamic estimation method and system based on multi-scale fusion, provides a spatial stratified sampling strategy, realizes refined partitioning based on growth potential difference and disease risk clustering, ensures that sampling points cover field key areas, and meanwhile, improves the accuracy of crop disease severity estimation. The dynamic time compensation sampling strategy is combined with the field disease development situation and environmental factor self-adaptive adjustment sampling to provide high-quality and dynamic data support for estimation; multi-source features such as scab quantity and edge roughness are introduced into the leaf scale to describe scab morphology and development difference in a fine granularity manner, and the plant scale is fused with weights such as leaf space position to highlight contribution of functional leaves and key leaves to the health of the whole plant; the spatial distribution pattern of the field diseases is effectively reflected by combining the field scale with the growth potential partition weight, the risk partition weight, the health degree and the like of the plants, and finally, the constructed dynamic time integral model is fused with the propagation rate and the fertility sensitivity to obtain a dynamic comprehensive estimation result of the accumulated influence of the diseases in the monitoring period.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing and deep learning, and particularly relates to a crop disease severity dynamic estimation method and system based on multi-scale fusion. BACKGROUND

[0002] At each stage of crop growth, it is often easy to be infected with various diseases, which not only threatens the agricultural yield and quality, but also directly affects the economic benefit and safety of agricultural production. Disease severity is a key indicator for measuring the degree of crop impact, and its estimation result is of great significance for scientific prevention and yield prediction.

[0003] In recent years, some researches have formed a multi-scale disease severity estimation path from leaf to plant to field. At the leaf scale, the proportion of lesion area to total leaf area is taken as the core indicator, and the formula is used to estimate the leaf-level disease severity. At the plant scale, the leaf-level disease severity of multiple leaves sampled in a single plant is integrated, and the formula is used to estimate the plant-level disease severity. At the field scale, the plant-level disease severity of multiple plants sampled in the field is integrated, and the formula is used to estimate the field-level disease severity.

[0004] In terms of field data acquisition, the existing methods generally use a fixed rule field sampling strategy, that is, plant sampling is performed according to the row and column division of the field or the preset sampling points, disease observation is performed on a limited number of plants, the disease severity is estimated, and the disease severity level of the entire field is inferred from the result.

[0005] In terms of time series information utilization, the existing technology is mostly based on static data, and usually selects a certain specific growth stage or a fixed time point of the crop for observation to obtain the disease severity result at that time. Single monitoring is the main form, and the obtained result is generally used to reflect the disease occurrence and spatial distribution state of the crop at the current period.

[0006] In summary, the existing multi-scale crop disease severity estimation method cannot comprehensively and accurately reflect the disease severity level of the crop at the leaf, plant, field and other scales. At the leaf scale, only the lesion area ratio is relied on to calculate the leaf-level disease severity, without considering the remaining characteristics of the lesion, which may underestimate the severity estimation result of different disease types and different development stages. At the plant scale, only the severity information of the sampled leaves is simply averaged, without considering the contribution difference of the leaves due to their different properties, which may cover up the role of key leaves and deviate from the actual severity result. At the field scale, only the severity information of the sampled plants is simply averaged, without considering the spatial heterogeneity and representativeness of the sampling points, which may cover up the key disease area and make it difficult to provide reliable basis for precise prevention and control.

[0007] Estimating crop disease severity at the field scale often requires field sampling. However, existing sampling strategies rely on fixed rules and do not combine remote sensing images with disease risk clustering for zoning. They also fail to perform dynamic compensation sampling based on changes in field severity and environmental factors during critical disease periods. This results in insufficient representativeness and comprehensiveness of the sampling results, affecting the timeliness and accuracy of field severity estimation.

[0008] Furthermore, existing methods for estimating crop disease severity are mostly static estimates, failing to consider the dynamic cumulative effects of diseases over time. Disease severity changes continuously over time, while static analysis methods can only obtain the severity at a specific moment, lacking cross-temporal data accumulation and dynamic analysis, making it difficult to reflect the development trend and spread patterns of diseases. Summary of the Invention

[0009] In view of this, the present invention provides a method and system for dynamic estimation of crop disease severity based on multi-scale fusion, which realizes accurate and dynamic assessment of crop disease severity by integrating multi-source data of three spatial scales (leaf, plant, and field) and time scale.

[0010] The method for dynamically estimating crop disease severity based on multi-scale fusion provided by this invention specifically includes the following steps:

[0011] Step 1: Collect remote sensing images of the crop planting area, including RGB and multispectral images. Establish a spatial stratified sampling strategy. First, divide the crop planting area into multiple field areas. Then, based on the crop growth potential of the field areas calculated from the multispectral images, divide the field areas into different growth potential zones. Divide the RGB images into grid cells in each growth potential zone, extract crop areas and lesion areas, calculate lesion density and aggregation index, and cluster them. Based on the clustering results, determine high-risk, medium-risk, and low-risk areas within the growth potential zone. Use a dynamic time-compensated sampling strategy to adjust the sampling frequency based on the severity at the field scale and environmental factors.

[0012] Step 2: Set the basic sampling period, and set the target number of plants and the number of leaves selected from different parts of the plants for high-risk, medium-risk and low-risk areas respectively. Obtain RGB images of leaves and plants, and record plant and leaf information; for field areas, calculate environmental parameters based on temperature and rainfall.

[0013] Step 3: Analyze the RGB image of the leaf to obtain the pixel area of ​​the lesion region, the pixel area of ​​the leaf, the number of lesions, the roughness of the lesion edge, and the variance of the lesion color. Calculate the first severity based on the constructed enhanced severity estimation model. Calculate the second severity based on the first severity, leaf spatial location weight, leaf age weight, and area weight. Calculate the field-scale enhanced disease severity based on the second severity, plant growth vigor zoning weight, plant risk zoning weight, local spatial heterogeneity, and plant health status.

[0014] Step 4: Construct a dynamic integral model of disease impact, using field-scale enhanced disease severity to calculate the dynamic severity of crop diseases during the monitoring period.

[0015] Furthermore, the remote sensing images are acquired by using a drone equipped with a high-precision camera and a multispectral camera to acquire RGB images and multispectral images respectively, and recording the GPS coordinates of each image. The RGB images and multispectral images are spatiotemporally aligned.

[0016] Furthermore, the method for calculating the crop growth potential (NDVI) of a field area from multispectral imagery is as follows:

[0017]

[0018] Wherein, NDVI is the Normalized Difference Vegetation Index, with a value range of -1 to 1; NIR represents the reflectance in the near-infrared band, and R represents the reflectance in the red band.

[0019] Furthermore, the lesion density of the grid cell is calculated as follows:

[0020]

[0021] Among them, D lesion A represents the lesion density within a grid cell. spot A represents the area of ​​a lesion pixel within the grid. grid This represents the area of ​​crop pixels within the grid.

[0022] Furthermore, the aggregation index of the grid cell is calculated as follows:

[0023]

[0024] Among them, I cluster D represents the clustering index within the grid. lesion D represents the lesion density within the grid. neighbor This represents the average lesion density of all grids within the corresponding actual ground setting range centered on this grid.

[0025] Furthermore, the constructed enhanced severity estimation model is as follows:

[0026]

[0027] Among them, LSD enh A represents the severity of enhanced disease at the leaf scale. lesion A represents the pixel area of ​​the lesion region. leaf The pixel area represents the entire leaf, C represents the number of lesions, E represents the edge roughness of the lesions, and V represents the pixel area of ​​the entire leaf. color The variance of lesion color is represented by α1, α2, and α3, which are all adjustment parameters.

[0028] Furthermore, the second severity is calculated based on the first severity, leaf spatial location weight, leaf age weight, and area weight, specifically as follows:

[0029]

[0030] Among them, PSD enh LSD is the severity of enhanced disease at the plant scale, also known as the second severity level. enh (i) represents the leaf-scale severity of the enhanced disease on the i-th leaf, N represents the total number of leaves sampled from the plant, and w h (i) represents the spatial position weight of the i-th leaf, w a (i) represents the leaf age weight of the i-th leaf, w s (i) represents the leaf area weight of the i-th leaf.

[0031] Furthermore, the enhanced disease severity at the field scale is calculated based on the second severity, plant growth vigor zoning weight, plant risk zoning weight, local spatial heterogeneity, and plant health status, specifically as follows:

[0032]

[0033] Among them, FSD enh PSD indicates the severity of enhanced disease at the field level. enh (j) represents the plant-level severity of enhanced disease in the j-th crop, where β1 and β2 are adjustment parameters, and M is the total number of plants sampled in the field; w v (j) represents the partition weight of plant growth potential, w r (j) represents the plant risk zoning weight, D j NDVI represents local spatial heterogeneity. j This represents the NDVI value of the j-th crop.

[0034] Furthermore, the dynamic integral model for the impact of the disease is as follows:

[0035]

[0036] The integral is calculated using the trapezoidal numerical integration method, where T1 and T2 represent the start and end points of time in days, and FSD is... enh S(t) represents the severity of the enhanced disease at the field scale on day t; S(t) represents the growth period sensitivity function; and G(t) represents the propagation rate influencing factor, calculated as follows:

[0037]

[0038] Where δ represents the propagation rate weight, This indicates the rate of disease transmission.

[0039] This invention proposes a dynamic estimation system for crop disease severity based on multi-scale fusion, characterized by including a UAV remote sensing data acquisition and sampling strategy generation module, a field data sampling module, a multi-scale enhanced disease severity calculation module, and a disease dynamic impact assessment module.

[0040] The UAV remote sensing data acquisition and sampling strategy generation module acquires RGB images and multispectral images of the crop planting area, divides the field area into plots according to spatial range, calculates the crop growth potential of the plot area based on the vegetation index of the multispectral image, and divides it into different levels of growth potential zones according to thresholds; divides grid cells within each growth potential zone, calculates lesion density and aggregation index to determine different risk areas within each growth potential zone, and outputs a complete sampling space; supports dynamic adjustment of sampling frequency based on historical field-scale disease severity data and real-time environmental factors.

[0041] The field data sampling module determines sampling points based on the sampling space and sampling frequency output by the UAV remote sensing data acquisition and sampling strategy generation module; sets the target number of plants according to risk level and the number of leaves according to plant part; collects RGB images of leaves and plants and records information; collects environmental parameters, records temperature and rainfall, and calculates the average value as field environmental parameters.

[0042] The multi-scale enhanced disease severity calculation module analyzes leaf RGB images collected by the field data sampling module, extracts the pixel area of ​​the lesion region and the pixel area of ​​the leaf, and counts the number of lesions, the roughness of the lesion edge, and the variance of the lesion color. The enhanced severity estimation model is used to calculate the leaf-scale enhanced disease severity of a single leaf. Based on the severity of all leaves of a single plant, the plant-scale enhanced disease severity is calculated by combining the weights of spatial location, leaf age, and area. Based on the severity of all plants in the field, the field-scale enhanced disease severity is obtained by combining the weights of plant growth vigor zoning, plant risk zoning, local spatial heterogeneity, and plant health status.

[0043] The Disease Dynamic Impact Assessment module calculates the dynamic severity of crop diseases during the monitoring period based on the field-scale enhanced disease severity output by the multi-scale enhanced disease severity calculation module.

[0044] Beneficial effects:

[0045] 1. The spatial stratified sampling strategy proposed in this invention achieves refined zoning based on differences in growth potential and disease risk clustering, ensuring that sampling points cover key areas in the field. At the same time, the dynamic time-compensated sampling strategy can adaptively increase sampling density at critical periods by combining the development trend of diseases in the field and environmental factors. This avoids the blindness of fixed-period sampling and ensures sufficient data in key areas and critical periods, providing high-quality and dynamic data support for multi-scale disease severity estimation.

[0046] 2. This invention provides a multi-scale disease severity estimation system with multi-feature fusion and precise weighting, which significantly improves the accuracy of disease severity estimation. At the leaf scale, it introduces multi-source features such as lesion number, edge roughness, and color variance to finely characterize lesion morphology and developmental differences. At the plant scale, it integrates the weights of leaf spatial location, leaf age, and leaf area to highlight the contribution of functional leaves and key leaves to the overall health of the plant. At the field scale, it combines the weights of plant growth vigor zoning, risk zoning, health status, and local spatial heterogeneity to effectively reflect the spatial distribution pattern of diseases in the field.

[0047] 3. This invention integrates the sensitivity of the growth period with the rate of disease transmission, realizing the dynamic accumulation and temporal characterization of disease severity. By constructing a dynamic time integral model, the cumulative impact of the disease during the monitoring period is quantified, overcoming the limitations of static estimation that only reflects the state at a single moment, and providing a scientific basis for the timing selection and effectiveness evaluation of control measures. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the dynamic estimation method for crop disease severity based on multi-scale fusion provided by the present invention.

[0049] Figure 2 This is a schematic diagram of the system structure of the crop disease severity dynamic estimation system based on multi-scale fusion provided by the present invention. Detailed Implementation

[0050] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0051] The present invention provides a method for dynamic estimation of crop disease severity based on multi-scale fusion, the process of which is as follows: Figure 1 As shown, the specific steps include:

[0052] Step 1: Collect UAV remote sensing images of the crop planting area, including RGB images and multispectral images, and establish a sampling strategy combining spatial stratified sampling and dynamic time-compensated sampling. Divide the sampling space according to the spatial stratified sampling strategy. First, divide the crop planting area into multiple field areas. Then, calculate the crop growth potential of the field areas based on the multispectral images. Based on the high, medium, and low growth potential thresholds, divide the field areas into high growth potential zones, medium growth potential zones, and low growth potential zones. Within each growth potential zone, divide the corresponding RGB image into grid cells according to the set spatial range. Extract the crop area and lesion area of ​​the grid cell, calculate the lesion density and aggregation index of the grid cell, and then cluster the lesion density and aggregation index. Based on the clustering results, determine different risk areas within the growth potential zone, including high-risk areas, medium-risk areas, and low-risk areas. Simultaneously, based on the calculated field-scale severity and environmental factors, dynamically adjust the sampling frequency using a dynamic time-compensated sampling strategy.

[0053] Specifically, the acquisition method of UAV remote sensing images can be as follows: the UAV is equipped with a high-precision camera and a multispectral camera to acquire RGB images and multispectral images respectively, and the GPS coordinates of each image are recorded. The RGB images and multispectral images are spatiotemporally aligned.

[0054] The method for calculating the NDVI of a field area based on multispectral imagery in this invention is as follows:

[0055]

[0056] Among them, NDVI is the normalized vegetation index, with a value range of -1 to 1. Negative values ​​usually correspond to non-vegetated areas, while higher positive values ​​indicate stronger photosynthetic activity and more vigorous growth of vegetation, thus directly reflecting the crop growth potential level; NIR represents the reflectance in the near-infrared band, and R represents the reflectance in the red band.

[0057] Furthermore, since UAV imaging may be affected by uneven lighting, sensor noise, etc., leading to abnormal local NDVI values, this invention can use median filtering to smooth the NDVI image, effectively eliminating single-point noise while preserving the spatial distribution trend of NDVI. Subsequently, regions with NDVI ≤ 0 are removed by threshold cropping, finally generating an NDVI grayscale image. Partitioning based on growth potential can avoid repeated sampling in areas with the same growth potential, improving sampling efficiency.

[0058] The method for calculating the lesion density of the grid cell in this invention is as follows:

[0059]

[0060] Among them, D lesion A represents the lesion density within a grid cell.spot A represents the area of ​​a lesion pixel within the grid. grid This represents the area of ​​crop pixels within the grid.

[0061] The aggregation index of the grid cells in this invention is calculated as follows:

[0062]

[0063] Among them, I cluster D represents the clustering index within the grid. lesion D represents the lesion density within the grid. neighbor This represents the average lesion density of all grids within the corresponding actual ground setting range centered on this grid.

[0064] In this invention, the method for determining different risk regions within the growth potential zone based on the clustering results of lesion density and aggregation index is as follows:

[0065] Lesion density and clustering index are used as two-dimensional input features of the K-means clustering algorithm. k is set to 3, corresponding to high, medium and low risk areas. The rationality of clustering is verified by the silhouette coefficient to ensure that the similarity of grid features within the same category is greater than the first threshold. Iteration stops when the change of cluster center is less than the second threshold. Finally, risk areas are defined based on the clustering results: high-risk areas are those with lesion density greater than the third threshold and clustering index greater than the fourth threshold; medium-risk areas are those with lesion density in the fifth threshold range and clustering index in the sixth threshold range; and low-risk areas are those with lesion density less than the seventh threshold and clustering index less than the eighth threshold.

[0066] Step 2: Based on the sampling space determined in Step 1, a spatial stratified sampling strategy is used for preliminary data sampling. A basic sampling period is set. For high-risk, medium-risk, and low-risk areas, the number of target plants to be selected in each area and the number of leaves to be selected from different parts of the plants are set. RGB images of the leaves and the plants are obtained, and the plant ID, plant GPS coordinates, and the position of the leaves in the plant are recorded. For field areas, the temperature and rainfall at multiple set locations are recorded, and the average value is taken as the environmental parameter.

[0067] The acquisition of RGB images of the leaves and the plant itself is achieved using a ground-based mobile platform equipped with a 6-DOF robotic arm and an imaging end effector. This end effector includes a high-definition RGB camera and a supplementary lighting unit. Upon reaching the target plant, a visual servo aligns the arm with the leaf position, selecting representative leaves from the top, middle, and bottom of each plant. High-definition RGB images of the leaves are acquired at a preset working distance and in a fixed posture, while simultaneously recording the plant's position. Subsequently, the platform takes a picture of the entire plant at an angle above it, automatically generating and recording a plant ID that corresponds to all the leaves photographed on that plant, and simultaneously recording the plant's GPS coordinates.

[0068] The method for obtaining temperature and rainfall data at multiple designated locations within a field area is to employ a five-point deployment method within each field area, placing one environmental sensor node at the center and four corners of the field to record the temperature and rainfall at each location.

[0069] Step 3: Analyze the leaf RGB images obtained in Step 2 to obtain the pixel area of ​​the lesion region and the pixel area of ​​the leaf. Determine the number of lesions, the roughness of the lesion edge, and the variance of the lesion color. Then, use the constructed enhanced severity estimation model to calculate the enhanced disease severity at the leaf scale. Based on the enhanced disease severity at the leaf scale of all leaves on the plant in the sample, spatial location weight, leaf age weight, and area weight, calculate the enhanced disease severity at the plant scale. Based on the enhanced disease severity at the plant scale of all plants in the field area, plant growth vigor zoning weight, plant risk zoning weight, local spatial heterogeneity, and plant health status, calculate the enhanced disease severity at the field scale.

[0070] The enhanced severity estimation model at the blade scale constructed in this invention is as follows:

[0071]

[0072] Among them, LSD enh Indicates the severity of enhanced diseases at the leaf scale; A lesion A represents the pixel area of ​​the lesion region. leaf The pixel area represents the entire leaf, C represents the number of lesions, E represents the edge roughness of the lesions, and V represents the pixel area of ​​the entire leaf. color The variance of lesion color is represented by α1, α2, and α3, which are all adjustment parameters.

[0073] Specifically, the pixel area A of the blade leaf The U-Net semantic segmentation model can be used to segment the RGB images of leaves. First, the leaf outline is extracted, and then the lesions are segmented from the leaf region to calculate the pixel area of ​​the lesion region and the pixel area of ​​the leaf.

[0074] The number of lesions, C, can be obtained by performing connected component analysis on the lesion region using the 8-neighborhood connectivity criterion, and counting the number of connected regions of the lesions. Since the number of lesions varies greatly, a logarithmic function can also be used to compress the number of lesions, preserving the numerical differences while avoiding the excessive influence of maxima.

[0075] The edge roughness E of the lesion is calculated by dividing the actual perimeter of the lesion outline by the perimeter of a circle with the same area as the lesion.

[0076] Lesion color variance V color It is used to identify the heterogeneity of lesion development stages. It extracts the HSV values ​​of all pixels from the segmented lesion area, where H is hue, S is saturation, and V is lightness. It calculates the variance of the three channels H, S, and V respectively, and then assigns weights to the variances of the three channels according to the specific disease type.

[0077] The parameters α1, α2, and α3 were initially set based on expert experience, and then fine-tuned through small-scale field trials.

[0078] This invention achieves precise aggregation of the overall impact of plant diseases by assigning weights to the spatial location, age, and area of ​​leaves. The calculation formula is as follows:

[0079]

[0080] Among them, PSD enh LSD indicates the severity of enhanced disease at the plant scale. enh (i) represents the leaf-scale severity of the enhanced disease on the i-th leaf, N represents the total number of leaves sampled from the plant, and w h (i) represents the spatial position weight of the i-th leaf, w a (i) represents the leaf age weight of the i-th leaf, w s (i) represents the leaf area weight of the i-th leaf.

[0081] Specifically, leaves at different spatial locations contribute significantly to plant growth and yield. Leaves closer to the top (e.g., flag leaves) receive more photosynthetically active radiation, resulting in stronger photosynthesis and higher priority for nutrient transport, thus having a greater impact on yield. Therefore, the spatial weight of leaves at the top, middle, and bottom decreases sequentially. Leaf age is estimated based on leaf color, texture, and location in the RGB image of the leaf, with the weight of new leaves, mature leaves, and old leaves decreasing sequentially. Furthermore, larger leaves contribute more to the overall plant health when disease occurs. Based on the proportion of leaf area to the total area of ​​all sampled leaves, all sampled leaves are categorized into large leaves, medium leaves, and small leaves, each assigned a decreasing weight.

[0082] Enhanced disease severity estimation at the field scale. Traditional field-level severity estimation often directly averages the severity of sampled plants. This step introduces weights based on plant growth vigor, risk zone, local spatial heterogeneity, and health status to achieve precise quantification of the overall disease level of the field. The calculation formula is as follows:

[0083]

[0084] Among them, FSD enh PSD indicates the severity of enhanced disease at the field level. enh (j) represents the plant-level severity of enhanced disease in the j-th crop, β1 and β2 are both adjustment parameters, with initial values ​​set based on expert experience, and then fine-tuned through small-scale field trials, where M is the total number of plants sampled in the field; w v (j) represents the partition weight of plant growth potential, w r (j) represents the plant risk zoning weight, D j To represent local spatial heterogeneity, based on GPS coordinates, the PSD of the j-th plant and its four neighboring plants is taken. enh The difference between the means is D j NDVI j The NDVI value represents the j-th crop, which reflects the health of the plant. The lower the value, the more likely it is to be affected by disease.

[0085] Based on the calculated severity and environmental factors at the field scale, the dynamic time-compensated sampling strategy established in step 1 is used to dynamically adjust the sampling frequency. Its core is to calculate ω... t The sampling frequency can be adjusted using the following formula:

[0086] ωt=γ1×|ΔFSD t |+γ2×R t +γ3×TDD t

[0087] Where, ω t This represents the empirical threshold for day t. When the value is greater than 0.5, the original basic sampling period is shortened. When two consecutive ω values ​​are obtained... t When the value is less than or equal to 0.5, compensation is stopped, and the original basic sampling period is returned; ΔFSD t R represents the change in field-scale severity on day t compared to the previous sampling, where γ1, γ2, and γ3 are adjustment parameters; t TDD represents the normalized rainfall on day t. t Let represent the normalized temperature accumulation on day t (Growing Degree Days), which is uniformly normalized using the following formula:

[0088]

[0089] Where X is the original value, X min X is the historical minimum value for that day. max This is the highest historical value for that day.

[0090] Step 4: Construct a dynamic integral model of disease impact by integrating the disease transmission rate over time with the crop growth period sensitivity weights. Use the enhanced disease severity at the field scale obtained in Step 3 to calculate the dynamic severity of crop diseases during the monitoring period, thus fully depicting the cumulative impact of diseases during the monitoring period.

[0091] The dynamic integral model of disease impact constructed in this invention is shown below:

[0092]

[0093] The integral is calculated using the trapezoidal numerical integration method, where T1 and T2 represent the start and end points of time in days, and FSD is... enh S(t) represents the field-level severity of enhanced disease on day t; S(t) represents the growth stage sensitivity function, which determines the crop's growth stage on day t by combining the drone's shooting time and plant images; G(t) represents the propagation rate influencing factor, and its calculation formula is as follows:

[0094]

[0095] Where δ represents the propagation rate weight, This represents the rate of disease spread, i.e., the daily variation rate of severity. G(t) is only weighted when its value is positive.

[0096] Using the multi-scale fusion-based dynamic estimation method for crop disease severity provided in this invention, a multi-scale fusion-based dynamic estimation system for crop disease severity was constructed. The system architecture is as follows: Figure 2 As shown, it enables precise and dynamic monitoring of crop diseases, including a UAV remote sensing data acquisition and sampling strategy generation module, a field data sampling module, a multi-scale enhanced disease severity calculation module, and a disease dynamic impact assessment module.

[0097] The UAV remote sensing data acquisition and sampling strategy generation module is responsible for acquiring remote sensing data and formulating sampling strategies, providing a basis for subsequent sampling. It controls the UAV to acquire RGB images and multispectral images of the crop planting area, and preprocesses the images to ensure data accuracy. Based on the plot boundaries of the RGB images, it divides the field areas according to spatial range. Based on the vegetation index of the multispectral images, it calculates the crop growth potential of the field areas and divides them into different levels of growth potential zones according to thresholds. Within each growth potential zone, it divides the area into grid cells, extracts the crop area and lesion area within the grid, calculates the lesion density and aggregation index, determines the high, medium, and low risk areas within each growth potential zone, and outputs the complete sampling space. Furthermore, it supports dynamically adjusting the sampling frequency based on historical field-scale disease severity data and real-time environmental factors.

[0098] The field data sampling module is responsible for collecting field data according to the sampling strategy, providing basic data for severity calculation. Based on the sampling space and sampling frequency output by the UAV remote sensing data collection and sampling strategy generation module, sampling points are determined; the number of target plants is set according to risk level, and the number of leaves is set according to plant location; RGB images of leaves and plants are collected, recording core information: plant ID, plant GPS coordinates, and leaf position within the plant; environmental parameters are collected, including temperature and rainfall, and the average values ​​are calculated as field environmental parameters.

[0099] The multi-scale enhanced disease severity calculation module is responsible for calculating the three levels of severity from leaf to field, achieving accurate quantification of disease severity. It analyzes leaf RGB images collected by the field data sampling module, extracts the pixel area of ​​lesion regions and leaf pixel area, and statistically analyzes the number of lesions, lesion edge roughness, and lesion color variance. An enhanced severity estimation model is used to calculate the leaf-scale enhanced disease severity of a single leaf. Based on the severity of all leaves on a single plant, combined with spatial location weights, leaf age weights, and area weights, the plant-scale enhanced disease severity is calculated. Based on the severity of all plants within a field, combined with plant growth vigor zoning weights, plant risk zoning weights, local spatial heterogeneity, and plant health status, the field-scale enhanced disease severity is obtained.

[0100] The disease dynamic impact assessment module is responsible for characterizing the cumulative impact of diseases and providing a basis for prevention and control decisions. Using a constructed dynamic integral model, the dynamic severity of crop diseases during the monitoring period is calculated based on the field-scale enhanced disease severity output by the multi-scale enhanced disease severity calculation module.

[0101] Example:

[0102] This embodiment takes a crop planting area as an example and uses the crop disease severity dynamic estimation method and system based on multi-scale fusion provided by this invention to realize dynamic estimation of crop disease severity based on UAV remote sensing images. The specific process includes:

[0103] S1. Collect UAV remote sensing images of crop planting areas and formulate spatial stratified sampling and dynamic time-compensated sampling strategies.

[0104] To ensure that the data collected subsequently can fully cover key areas of the field and accurately capture the dynamic changes during critical periods of disease, this step proposes a sampling strategy that combines spatial stratified sampling with dynamic time-compensated sampling. Spatial stratified sampling is the basis for data collection in step two, while dynamic time-compensated sampling can dynamically adjust the sampling frequency based on the environmental factors collected in step two and the field-level disease severity results in step three. The combination of the two can increase the representativeness and real-time nature of the sampled data.

[0105] Spatial stratified sampling achieves accurate sampling primarily through three levels of stratification. Prior to this, UAV image data needs to be acquired. During the stable lighting period from 10:00 AM to 12:00 PM, the UAV, equipped with both a high-precision camera and a multispectral camera, flies along a preset route to acquire RGB images and multispectral images. The RGB and multispectral images are required to be spatiotemporally aligned. The UAV is set to a flight altitude of 15m, with a forward overlap ≥80% and a lateral overlap ≥70%. The GPS coordinates of each image are recorded simultaneously. Spatial stratified sampling is then performed. Detailed stratification operations are as follows:

[0106] (1) The first layer is to divide the field into plots. The crop planting area is divided into multiple plots of 50m×50m to provide a basic spatial framework for subsequent stratification and to ensure that no sampling range is missed.

[0107] (2) The second layer is based on growth potential. Based on the field zoning, NDVI is calculated based on UAV multispectral images, and NDVI images are smoothed by 3×3 median filtering. This effectively eliminates single-point noise while preserving the spatial distribution trend of NDVI. Then, regions with NDVI≤0 are removed by threshold cropping, and finally, NDVI grayscale images are generated. Based on the division of each field area, 0.6≤NDVI≤1 is a high growth potential area, 0.4<NDVI<0.6 is a medium growth potential area, and 0<NDVI≤0.4 is a low growth potential area.

[0108] (3) The third layer is based on disease risk clustering. First, based on the UAV RGB image, in the high, medium and low growth potential areas, grid units are divided on the RGB image according to the actual ground space of 1m×1m, and the lesion density and aggregation index of the area are calculated. The lesion density and aggregation index are used as the two-dimensional input features of the K-means clustering algorithm, and k is set to 3, corresponding to high, medium and low risk areas. The rationality of clustering is verified by the silhouette coefficient to ensure that the similarity of grid features within the same category is greater than 0.7. The iteration stops when the change of the cluster center is less than 0.01. Finally, the risk areas are defined based on the clustering results: the high risk area is the area with lesion density greater than 15% and aggregation index greater than 1.2, the medium risk area is the area with lesion density between 5% and 15% and aggregation index between 0.8 and 1.2, and the low risk area is the area with lesion density less than 5% and aggregation index less than 0.8.

[0109] S2. Based on the sampling strategy, preliminary data on leaves, plants, fields, and the environment were collected.

[0110] Preliminary data sampling was conducted based on the spatial stratified sampling strategy proposed in Step 1, with a basic sampling period of 5 days. Within each field of the crop planting area, in the high, medium, and low growth potential zones, 5 to 6 target plants were selected in the high-risk sub-zone, 3 to 4 target plants in the medium-risk zone, and 1 to 2 target plants in the low-risk zone. A ground-based mobile platform equipped with a 6-DOF robotic arm and an imaging end effector, including a high-definition RGB camera and a supplementary lighting unit, was used. Upon reaching the target plant, a visual servo aimed at the leaf position, selecting 1 to 2 representative leaves from the top, middle, and bottom of each plant. High-definition RGB images of the leaves were acquired at a preset working distance (20 to 30 cm, preferably 25 cm) and a fixed posture, simultaneously recording the plant's position. Subsequently, the platform photographed the entire plant at an upward angle, automatically generating and recording a plant ID corresponding to all the photographed leaves, and simultaneously recording the plant's GPS coordinates.

[0111] Within each field area, a five-point deployment method was used, with one environmental sensor node placed at the center and one at each of the four corners of the field. These nodes recorded the temperature and rainfall at each location, and the average value was used as the representative environmental parameter for the field. Temperature was recorded hourly, and the accumulated temperature for the day was the cumulative temperature over 24 hours.

[0112] The total number of plants sampled in each field area should be controlled between 27 and 36 to avoid excessive sampling and increased costs. The specific number can be adjusted according to the actual situation of the field. The sampling amount of high-risk sub-regions in each growth vigor zone should not be less than 40% of the total sample in the growth vigor zone to ensure key coverage of disease-prone areas.

[0113] S3. Based on the sampling data, define formulas to estimate the severity of enhanced leaf-level, plant-level, and field-level diseases, and combine dynamic time-compensated sampling strategies to iterate the sampling data.

[0114] This step integrates multiple features and weight adjustments at the leaf, plant, and field scales to achieve multi-scale enhanced disease severity estimation. At the same time, dynamic compensation sampling is triggered based on the estimation results to ensure that the data can capture disease changes in real time, providing high-quality multi-time point data support for the dynamic analysis in step four.

[0115] When calculating the severity of enhanced diseases at the plant scale, the weight of the top leaves was set to 1.2, the middle leaves to 1.0, the bottom leaves to 0.8, the new leaves to 1.2, the mature leaves to 1.0, and the old leaves to 0.8. Based on the proportion of the leaf area to all sampled leaves, all sampled leaves were divided into large leaves, medium leaves, and small leaves, and assigned weights of 1.2, 1.0, and 0.8, respectively.

[0116] When calculating the severity of enhanced diseases at the field scale, plant growth vigor zoning weights were set: 1.2 for plants from high growth vigor areas, 1.0 for plants from medium growth vigor areas, and 0.8 for plants from low growth vigor areas. Similarly, plant risk zoning weights were set: 1.2 for plants from high-risk areas, 1.0 for plants from medium-risk areas, and 0.8 for plants from low-risk areas. To determine local spatial heterogeneity, the PSD of the j-th plant and its four neighboring plants was taken based on GPS coordinates. enh The difference between the means is D j ;

[0117] Based on the field-level severity estimation results and environmental factors, combined with a dynamic time-compensated sampling strategy, the original sampling period may be shortened to 3 days, but the sampling method remains unchanged.

[0118] S4. Based on the field-level disease severity estimation results over multiple days, a dynamic time integral model weighted by the spread rate and fertility sensitivity is proposed to obtain the dynamic comprehensive estimation results of the cumulative impact of the disease during the monitoring period.

[0119] Since traditional static estimation only reflects the severity of the disease at a certain moment, this step constructs a dynamic integral model by integrating the disease transmission rate over time with the crop growth stage sensitivity weight. This model comprehensively depicts the cumulative impact of the disease during the monitoring period. The dynamic comprehensive estimation result of the cumulative impact of the disease during the monitoring period is represented by TSD. The growth stage sensitivity function S(t) is determined by combining the drone shooting time and plant images to determine the crop's growth stage on day t. The seedling stage is set to 0.5, the heading or flowering stage is set to 1.0, the grain filling stage is set to 1.2, and the maturity stage is set to 0.8. The transmission rate weight δ is set to 1.0.

[0120] Experiments show that the method and system provided by this invention can accurately estimate the severity and dynamically monitor the disease, guiding precise prevention and control in the field. For example, high-risk areas can be prioritized for treatment, avoiding indiscriminate application of pesticides. At the same time, by revealing the stage characteristics and rate differences of disease development, it can help assess the possible evolution risk of the disease, reduce yield losses caused by untimely prevention and control, and thus improve the economic benefits and sustainability of agricultural production.

[0121] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic estimation method for crop disease severity based on multi-scale fusion, characterized in that, Specifically, the following steps are included: Step 1: Collect remote sensing images of the crop planting area, including RGB images and multispectral images. Establish a spatial stratified sampling strategy. First, divide the crop planting area into multiple field areas. Then, based on the crop growth potential of the field areas calculated from the multispectral images, divide the field areas into different growth potential zones. Divide grid cells in each growth potential zone according to the RGB images. Extract crop areas and lesion areas, calculate lesion density and aggregation index, and cluster them. Based on the clustering results, determine high-risk areas, medium-risk areas, and low-risk areas within the growth potential zones. A dynamic time-compensated sampling strategy was adopted to adjust the sampling frequency based on the severity at the field scale and environmental factors; Step 2: Set the basic sampling period, set the target number of plants and the number of leaves selected from different parts of the plants for high-risk, medium-risk and low-risk areas respectively, obtain RGB images of leaves and plants, and record plant and leaf information. For the field area, environmental parameters are calculated based on temperature and rainfall. Step 3: Analyze the RGB image of the leaf to obtain the pixel area of ​​the lesion region, the pixel area of ​​the leaf, the number of lesions, the roughness of the lesion edge, and the variance of the lesion color. Calculate the first severity based on the constructed enhanced severity estimation model. Calculate the second severity based on the first severity, leaf spatial location weight, leaf age weight, and area weight. Calculate the field-scale enhanced disease severity based on the second severity, plant growth vigor zoning weight, plant risk zoning weight, local spatial heterogeneity, and plant health status. Step 4: Construct a dynamic integral model of disease impact, using field-scale enhanced disease severity to calculate the dynamic severity of crop diseases during the monitoring period.

2. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The remote sensing images are acquired by using a drone equipped with a high-precision camera and a multispectral camera to acquire RGB images and multispectral images respectively, and recording the GPS coordinates of each image. The RGB images and multispectral images are spatiotemporally aligned.

3. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The method for calculating the crop growth potential (NDVI) of a field area from multispectral imagery is as follows: Wherein, NDVI is the Normalized Difference Vegetation Index, with a value range of -1 to 1; NIR represents the reflectance in the near-infrared band, and R represents the reflectance in the red band.

4. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The method for calculating the lesion density of a grid cell is as follows: Among them, D lesion A represents the lesion density within a grid cell. spot A represents the area of ​​a lesion pixel within the grid. grid This represents the area of ​​crop pixels within the grid.

5. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The clustering index of a grid cell is calculated as follows: Among them, I cluster D represents the clustering index within the grid. lesion D represents the lesion density within the grid. neighbor This represents the average lesion density of all grids within the corresponding actual ground setting range centered on this grid.

6. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The constructed enhanced severity estimation model is as follows: Among them, LSD enh A represents the severity of enhanced disease at the leaf scale. lesion A represents the pixel area of ​​the lesion region. leaf The pixel area represents the entire leaf, C represents the number of lesions, E represents the edge roughness of the lesions, and V represents the pixel area of ​​the entire leaf. color The variance of lesion color is represented by α1, α2, and α3, which are all adjustment parameters.

7. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The second severity is calculated based on the first severity, leaf spatial location weight, leaf age weight, and area weight, specifically as follows: Among them, PSD enh LSD is the severity of enhanced disease at the plant scale, also known as the second severity level. enh (i) represents the leaf-scale severity of the enhanced disease on the i-th leaf, N represents the total number of leaves sampled from the plant, and w h (i) represents the spatial position weight of the i-th leaf, w a (i) represents the leaf age weight of the i-th leaf, w s (i) represents the leaf area weight of the i-th leaf.

8. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The enhanced disease severity at the field scale is calculated based on the second severity level, plant growth vigor zoning weight, plant risk zoning weight, local spatial heterogeneity, and plant health status. The specific method is as follows: Among them, FSD enh PSD indicates the severity of enhanced disease at the field level. enh (j) represents the plant-level severity of enhanced disease in the j-th crop, where β1 and β2 are adjustment parameters, and M is the total number of plants sampled in the field; w v (j) represents the partition weight of plant growth potential, w r (j) represents the plant risk zoning weight, D j NDVI represents local spatial heterogeneity. j This represents the NDVI value of the j-th crop.

9. The method for dynamically estimating the severity of crop diseases according to claim 1, characterized in that, The dynamic integral model for the impact of the disease is as follows: The integral is calculated using the trapezoidal numerical integration method, where T1 and T2 represent the start and end points of time in days, and FSD is... enh S(t) represents the severity of the enhanced disease at the field scale on day t; S(t) represents the growth period sensitivity function; and G(t) represents the propagation rate influencing factor, calculated as follows: Where δ represents the propagation rate weight, This indicates the rate of disease transmission.

10. A dynamic crop disease severity estimation system based on multi-scale fusion, constructed using the dynamic crop disease severity estimation method described in claim 1, characterized in that, It includes a module for UAV remote sensing data acquisition and sampling strategy generation, a field data sampling module, a multi-scale enhanced disease severity calculation module, and a disease dynamic impact assessment module; The UAV remote sensing data acquisition and sampling strategy generation module acquires RGB images and multispectral images of the crop planting area, divides the field area into plots according to spatial range, calculates the crop growth potential of the plot area based on the vegetation index of the multispectral image, and divides it into different levels of growth potential zones according to thresholds; divides grid cells within each growth potential zone, calculates lesion density and aggregation index to determine different risk areas within each growth potential zone, and outputs a complete sampling space; supports dynamic adjustment of sampling frequency based on historical field-scale disease severity data and real-time environmental factors. The field data sampling module determines sampling points based on the sampling space and sampling frequency output by the UAV remote sensing data acquisition and sampling strategy generation module; it sets the target number of plants according to risk level and the number of leaves according to plant part; it collects RGB images of leaves and plants and records information. Environmental parameters were collected, including temperature and rainfall, and the average values ​​were calculated as field environmental parameters. The multi-scale enhanced disease severity calculation module analyzes the leaf RGB images collected by the field data sampling module, extracts the pixel area of ​​the lesion region and the pixel area of ​​the leaf, and counts the number of lesions, the roughness of the lesion edge, and the variance of the lesion color. The enhanced severity estimation model is used to calculate the leaf-scale enhanced disease severity of a single leaf. Based on the severity of all leaves of a single plant, the plant-scale enhanced disease severity is calculated by combining the weights of spatial location, leaf age, and area. Based on the severity of all plants in the field, combined with the zoning weights of plant growth potential, plant risk, local spatial heterogeneity, and plant health, the field-scale enhanced disease severity is obtained. The Disease Dynamic Impact Assessment module calculates the dynamic severity of crop diseases during the monitoring period based on the field-scale enhanced disease severity output by the multi-scale enhanced disease severity calculation module.

Images