A pesticide application effect evaluation method and system based on plant feature analysis

Abstract

The application discloses a pesticide application effect evaluation method and system based on plant feature analysis, and particularly relates to the field of image recognition and understanding, and is used for solving the problem that the continuous disease spots and the new pesticide damage spots cannot be accurately distinguished through image semantic segmentation and image classification because the appearance is homogeneous. The image before and after pesticide application is acquired, image semantic segmentation is performed to obtain the disease spot area before pesticide application and the necrotic spot area after pesticide application, the pixel traversal cost is calculated along the path of each pixel from the boundary to the necrotic spot to mark the expansion tendency pixels, the disease spot area is taken as the starting seed to perform constrained neighborhood propagation on the expansion tendency pixels to divide the continuous disease spot area and the new pesticide damage spot area, after the division result is modified through boundary neighborhood color attenuation verification, the area change rate and the area proportion of the modified area are used for image classification to output the pesticide application effect evaluation result.

Description

Technical Field

[0001] This invention relates to the field of image recognition and understanding technology, and more specifically, to a method and system for evaluating the effectiveness of pesticide application based on plant feature analysis. Background Technology

[0002] In the process of crop disease and pest control, objective evaluation of pesticide application effects is the foundation for adjusting pesticide application plans and achieving reduced dosage and increased efficiency. Currently, commonly used evaluation methods include two categories: manual field surveys and automated analysis based on machine vision. Manual surveys rely on plant protection personnel visually observing changes in plant characteristics such as the drying and shedding of lesions and the state of new leaves after pesticide application to determine efficacy; this method is labor-intensive and highly subjective. Automated evaluation methods using image processing technology typically collect images of the crop canopy or leaves after pesticide application, use image semantic segmentation to extract leaf areas and abnormal areas such as lesions and dead tissue from the background, and then use image classification to categorize these areas or the overall plant condition into effectiveness levels, outputting evaluation conclusions such as effective or ineffective.

[0003] However, under real field conditions, necrotic spots appearing on crops after pesticide application could be either the result of existing diseases continuing to develop without being suppressed, or phytotoxicity caused by the pesticide itself. Both types of spots appear as brown or yellowish-brown necrotic lesions, often surrounded by a chlorotic halo. They are highly similar in morphology, texture, and color, constituting a visual homogeneity. However, their agronomic meanings are diametrically opposed: the former indicates that the pesticide failed to control the disease, while the latter indicates that the pesticide was too active or improperly applied, causing plant damage. Existing semantic segmentation methods rely solely on the visual patterns of single-phase images for differentiation, failing to capture the different formation mechanisms and temporal priors behind the two types of spots. This easily leads to misclassifying pesticide-burned spots as target lesions or mistaking residual lesions for newly formed phytotoxic spots. This confusion at the segmentation and classification levels can cause subsequent efficacy evaluations to reach conclusions completely opposite to reality, such as poor efficacy or phytotoxicity, severely affecting the reliability of the evaluation results. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides a method and system for evaluating pesticide application effects based on plant characteristic analysis to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for evaluating pesticide application efficacy based on plant characteristic analysis, comprising: S1: Acquire images before and after application of the drug; S2: Perform image semantic segmentation on the image before application to obtain the lesion area before application; perform image semantic segmentation on the image after application to obtain the necrotic area after application; S3: Calculate the pixel crossing cost along the path from the boundary of the lesion area before drug application to each pixel in the necrotic area after drug application, and mark pixels with continuously low crossing costs and no sudden changes as pixels with expansion tendency. S4: Starting with the diseased area before application of pesticide, the pixels with the tendency to expand are spread in the neighborhood of the necrotic area after application of pesticide. The spread conditions are constrained by the pixel crossing cost and neighborhood similarity. The spread stops at the point where the crossing cost changes. The area covered by the spread is divided into the continuous diseased area and the rest is divided into the new pesticide-damaged area. S5: Based on the image before application, perform boundary neighborhood color attenuation verification on the continuous lesion area, remove pixels with abnormal boundary neighborhood color attenuation from the continuous lesion area and assign them to the newly formed drug-damaged lesion area, and obtain the corrected continuous lesion area and the corrected newly formed drug-damaged lesion area. S6: Using image classification, based on the area change rate of the corrected continuous lesion area and the area proportion of the corrected new pesticide-damaged area, output the pesticide application effect evaluation results.

[0006] Furthermore, images before and after application of the pesticide are acquired, including: Multispectral images of the crop canopy before pesticide application were acquired using multispectral imaging equipment and used as pre-application images. Collect multispectral images after pesticide application as post-application images; Both the pre-application and post-application multispectral images include red and near-infrared band images.

[0007] Furthermore, semantic segmentation is performed on the pre-application image to obtain the pre-application lesion region; semantic segmentation is performed on the post-application image to obtain the post-application necrotic lesion region, including: Band operations were performed on the red and near-infrared band images of the images before and after pesticide application to generate vegetation index images before and after pesticide application. Threshold segmentation was performed on the vegetation index images before and after pesticide application to extract areas with low vegetation vitality. Spatial matching was performed between the low vegetation vitality area in the image before pesticide application and the visible light texture in the image before pesticide application. Leaf vein texture and shadow areas were removed to obtain the diseased area before pesticide application. Spatial matching was performed between the low vegetation vitality areas in the post-application image and the visible light texture of the post-application image. Leaf vein texture and shadow areas were removed to obtain the necrotic spot areas after application.

[0008] Furthermore, along the path from the boundary of the lesion area before application to each pixel within the necrotic area after application, the pixel traversal cost is calculated. Pixels with continuously low traversal costs and no abrupt changes are marked as pixels with a tendency to expand, including: For each pixel in the necrotic patch region after drug application, determine the shortest path from the boundary of the lesion region before drug application to that pixel; Along the shortest path, the difference in vegetation index at the corresponding position between the vegetation index image before and after pesticide application is calculated pixel by pixel and used as the single-point crossing cost. The pixel crossing cost is obtained by summing the single-point crossing costs on the shortest path. If the cost of a pixel crossing is lower than a preset cost threshold and the change in the cost of a single point crossing of adjacent pixels on the shortest path is lower than a preset mutation threshold, then the pixel is marked as an expansion-prone pixel.

[0009] Further, determining the shortest path from the boundary of the lesion area before drug application to the pixel includes: calculating the distance transformation map of the lesion area before drug application to obtain the initial distance from each pixel in the necrotic lesion area after drug application to the boundary of the lesion area before drug application; using the single-point crossing cost as the pixel movement weight, starting from the pixel, moving gradually along the direction where the initial distance decreases and the cumulative movement weight is minimized to the boundary of the lesion area before drug application to obtain the shortest path.

[0010] Furthermore, using the lesion area before pesticide application as the starting seed, pixels with the tendency to expand are propagated in the necrotic lesion area after pesticide application. The propagation conditions are constrained by pixel crossing cost and neighborhood similarity. Propagation stops at the point where the crossing cost changes. The area covered by the propagation is divided into the continuous lesion area, and the rest is divided into the newly formed pesticide-damaged lesion area, including: The pixels contained in the diseased area before application, which serves as the starting seed, are marked as propagated pixels; The vegetation index difference between the propagated pixel and its neighboring pixels with the tendency to expand is calculated based on the vegetation index image after pesticide application as the neighborhood similarity. Pixels with a neighborhood similarity lower than a similarity threshold and whose pixel traversal cost is lower than a preset mutation threshold are added to the already propagated pixels. Repeat the process until no new pixels are added. The area covered by the already spread pixels is considered as the continued lesion area, and the area in the necrotic lesion area after application that is not covered by the already spread pixels is considered as the newly formed phytotoxic lesion area.

[0011] Furthermore, based on the pre-application image, boundary neighborhood color attenuation is verified in the continuous lesion area. Pixels with abnormal color attenuation in the boundary neighborhood are removed from the continuous lesion area and assigned to the newly formed pesticide-damaged lesion area, resulting in corrected continuous lesion areas and corrected newly formed pesticide-damaged lesion areas, including: Extract the boundary pixels of the continuous lesion region; Extend the neighborhood outwards along the boundary pixels to obtain the boundary neighborhood pixel band; Based on the vegetation index images before and after application of the pre-application image, the vegetation index difference between each pixel in the boundary neighborhood pixel band and the pre-application and post-application images is calculated as the color attenuation. Pixels whose color decay exceeds the color decay threshold are identified as pixels with abnormal color decay in the boundary neighborhood; Pixels with abnormal color decay in the boundary neighborhood are removed from the continuous lesion area and added to the newly formed drug-damaged lesion area to obtain the corrected continuous lesion area and the corrected newly formed drug-damaged lesion area.

[0012] Furthermore, the neighborhood is expanded outward along the boundary pixels to obtain a boundary neighborhood pixel band, including: for each boundary pixel in the continuous lesion area, the gradient direction of the boundary pixel in the vegetation index image after pesticide application is calculated as the expansion direction; pixels of a preset width are continuously taken outward from the boundary pixels along the expansion direction, and the expanded pixels corresponding to all boundary pixels are merged to form a boundary neighborhood pixel band.

[0013] Furthermore, using image classification, based on the area change rate of corrected persistent lesion areas and the area proportion of corrected newly formed pesticide-damaged areas, the pesticide application effect evaluation results are output, including: The pixel area of ​​the corrected continuation lesion region is calculated as the continuation lesion area, and the pixel area of ​​the lesion region before application of the drug is calculated as the lesion area before application of the drug. The ratio of the continuation lesion area to the lesion area before application of the drug is taken as the change rate of the continuation lesion area. The pixel area of ​​the newly formed pesticide-damaged area is calculated as the area of ​​the newly formed pesticide-damaged area, and the pixel area of ​​the necrotic area after pesticide application is calculated as the total area of ​​the necrotic area. The ratio of the area of ​​the newly formed pesticide-damaged area to the total area of ​​the necrotic area is taken as the proportion of the area of ​​the newly formed pesticide-damaged area. Using image classification, the pesticide application effect evaluation results are output based on the rate of change of continuous lesion area and the proportion of newly formed pesticide-damaged area.

[0014] On the other hand, the present invention provides a pesticide application effect evaluation system based on plant characteristic analysis, comprising: The image acquisition module is used to acquire images before and after application of the pesticide. The semantic segmentation module is used to perform image semantic segmentation on the image before drug application to obtain the lesion region before drug application; and to perform image semantic segmentation on the image after drug application to obtain the necrotic region after drug application. The cost labeling module is used to calculate the pixel crossing cost along the path from the boundary of the lesion area before drug application to each pixel in the necrotic area after drug application, and to label pixels with continuously low crossing costs and no abrupt changes as pixels with expansion tendency. The region division module is used to propagate pixels with the tendency to expand in the neighborhood of the necrotic spots after the application of pesticides, starting with the diseased spots before pesticide application. The propagation conditions are constrained by the pixel crossing cost and neighborhood similarity. Propagation stops at the point where the crossing cost changes. The area covered by the propagation is divided into the continuous diseased spot area and the rest is divided into the new pesticide-damaged spot area. The color verification module is used to verify the color decay of the boundary neighborhood of the continuous lesion area based on the pre-application image. Pixels with abnormal color decay at the boundary neighborhood are removed from the continuous lesion area and assigned to the newly damaged lesion area, thus obtaining the corrected continuous lesion area and the corrected newly damaged lesion area. The image classification module is used to evaluate the effectiveness of pesticide application by classifying images and outputting results based on the area change rate of the corrected continuous lesion area and the area ratio of the corrected new pesticide-damaged area.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By combining pixel traversal cost and neighborhood propagation, post-application necrotic spots are reconstructed into two parts with different causes: a continuation lesion area and a newly formed drug-damaged area. The pixel traversal cost is calculated along the path of each pixel from the boundary of the lesion area before application to the necrotic area after application, capturing the gradual decline in vegetation vitality during the progressive spread of the disease. The location of the traversal cost mutation is used as the boundary between disease expansion and drug damage truncation. Neighborhood propagation starts from the lesion area before application and grows outward under the dual constraints of traversal cost and neighborhood similarity. This ensures that the delineation of the continuation lesion area has spatial continuity, spectral consistency, and temporal evolution smoothness, avoiding confusion between necrotic spots that are similar in appearance but different in cause.

[0016] 2. Based on the continuation of the lesion area, boundary neighborhood color attenuation verification is performed. The pixel band outside the boundary is extracted along the vegetation vitality gradient direction and the color attenuation degree is calculated. Pixels with chlorotic halo features are corrected and classified into the new pesticide damage area, further eliminating misclassification caused by edge blurring. In the image classification stage, two indicators are combined as classification basis: the area change rate of the corrected continuation lesion area and the area proportion of the corrected new pesticide damage area. This can simultaneously evaluate the degree of pesticide inhibition on diseases and the risk of damage to crops in the two-dimensional feature space, and output classification results including significant efficacy, moderate efficacy, ineffective efficacy, and pesticide damage. This ensures that the evaluation chain composed of semantic segmentation and image classification is consistent with the real agronomic causal relationship. Attached Figure Description

[0017] Figure 1 This is a flowchart of a pesticide application effect evaluation method based on plant characteristic analysis according to the present invention; Figure 2 This is a schematic diagram of the structure of a pesticide application effect evaluation system based on plant characteristic analysis according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 This invention presents a method for evaluating pesticide application efficacy based on plant characteristic analysis, comprising: S1: Acquire images before and after application of the drug; S2: Perform image semantic segmentation on the image before application to obtain the lesion area before application; perform image semantic segmentation on the image after application to obtain the necrotic area after application; S3: Calculate the pixel crossing cost along the path from the boundary of the lesion area before drug application to each pixel in the necrotic area after drug application, and mark pixels with continuously low crossing costs and no sudden changes as pixels with expansion tendency. S4: Starting with the diseased area before application of pesticide, the pixels with the tendency to expand are spread in the neighborhood of the necrotic area after application of pesticide. The spread conditions are constrained by the pixel crossing cost and neighborhood similarity. The spread stops at the point where the crossing cost changes. The area covered by the spread is divided into the continuous diseased area and the rest is divided into the new pesticide-damaged area. S5: Based on the image before application, perform boundary neighborhood color attenuation verification on the continuous lesion area, remove pixels with abnormal boundary neighborhood color attenuation from the continuous lesion area and assign them to the newly formed drug-damaged lesion area, and obtain the corrected continuous lesion area and the corrected newly formed drug-damaged lesion area. S6: Using image classification, based on the area change rate of the corrected continuous lesion area and the area proportion of the corrected new pesticide-damaged area, output the pesticide application effect evaluation results.

[0020] In the specific implementation of S1, a multispectral imaging device is used to acquire a pre-application multispectral image of the crop canopy as the pre-application image, and a post-application multispectral image of the crop canopy is also acquired using the same device. The multispectral imaging device is a drone platform equipped with a multispectral camera. The multispectral camera has a red light imaging channel and a near-infrared imaging channel. The center wavelength of the red light imaging channel is in the range of 620 nm to 680 nm (e.g., 660 nm), and the center wavelength of the near-infrared imaging channel is in the range of 780 nm to 1100 nm (e.g., 850 nm). The red light imaging channel is used to generate red light images, and the near-infrared imaging channel is used to generate near-infrared images. Both the pre-application and post-application multispectral images include both red light and near-infrared images.

[0021] Multispectral images were acquired within 24 hours prior to pesticide application, under clear, cloudless conditions during the morning when the crop canopy was free of dew and the solar altitude angle was greater than 30 degrees. A solar altitude angle greater than 30 degrees ensured sufficient sunlight to illuminate the crop canopy, minimizing the impact of canopy shadows on the radiation intensity values ​​in the red and near-infrared bands of the multispectral images. The red and near-infrared band images from the pre-application multispectral images reflect the physiological state of the crop before application.

[0022] Post-application multispectral images were acquired between the 3rd and 7th day after pesticide spraying, for example, on the 5th day after spraying, under clear, cloudless conditions. The acquisition time was the same as that for the pre-application multispectral images. During acquisition, the same flight altitude, multispectral camera attitude angle, and multispectral camera exposure parameters were maintained as for the pre-application multispectral images. The ground sampling distance at the flight altitude was better than 5 cm per pixel. A ground sampling distance of 5 cm per pixel ensures that a single pixel corresponds to a square area no larger than 5 cm by 5 cm on the crop canopy, enabling the differentiation of disease spots and pesticide damage spots at the individual leaf level on the crop canopy.

[0023] The same flight altitude, the same multispectral camera attitude angle, and the same multispectral camera exposure parameters enable spatial registration between the post-application multispectral images and the pre-application multispectral images. Spatial registration means that pixels at the same spatial location in the red band images of the post-application and pre-application multispectral images correspond to the same location in the crop canopy, and pixels at the same spatial location in the near-infrared band images of the post-application and pre-application multispectral images correspond to the same location in the crop canopy. Spatial registration provides a pixel-level spatial correspondence basis for subsequent pixel-based cost calculation.

[0024] When collecting multispectral images before and after pesticide application, the field of view of the multispectral camera includes the entire crop canopy and a portion of the bare soil background. The portion of the bare soil background is used to help determine the distinction between vegetation and non-vegetation when generating subsequent vegetation index images.

[0025] Both pre- and post-application multispectral images were stored in the original radiometric calibration data format. This format preserves the radiant intensity values ​​of each pixel in the red band image and the near-infrared band image. The radiant intensity values ​​are digitally quantified values ​​of the light energy reflected from the ground objects received by the multispectral camera after photoelectric conversion, reflecting the reflectivity of the crop canopy in the red and near-infrared bands.

[0026] The reflectivity of the red light band in the reflectivity characteristics is affected by the chlorophyll content of the leaf and the internal tissue structure of the leaf, while the reflectivity of the near-infrared band in the reflectivity characteristics is affected by the reflection at the interface between the spongy tissue and the palisade tissue inside the leaf.

[0027] When crops are infected by diseases, the decrease in chlorophyll content in leaves leads to an increase in red light reflectance, while the damage to the internal tissue structure of leaves leads to a decrease in near-infrared reflectance. Similarly, when crops are burned by pesticides, the rapid dehydration and necrosis of leaf cells also lead to an increase in red light reflectance and a decrease in near-infrared reflectance. The direction of change in red light and near-infrared reflectance is the same in both disease infection and pesticide burn cases, but the magnitude of change differs. This difference in magnitude stems from the fact that disease infection is a gradual process of tissue necrosis, while pesticide burns are a sudden process. The rate of decrease in near-infrared reflectance during gradual tissue necrosis is slower than that during sudden tissue necrosis. This difference in magnitude is used in subsequent band calculations to distinguish between tissue necrosis caused by diseases and tissue necrosis caused by pesticides.

[0028] In the specific implementation of S2, band operations are performed on the red band image and near-infrared band image of the pre-application image to generate a pre-application vegetation index image. Similarly, band operations are performed on the red band image and near-infrared band image of the post-application image to generate a post-application vegetation index image. The band operation uses the normalized difference vegetation index (NDVIpre) calculation. The vegetation index value of each pixel in the pre-application vegetation index image is calculated as: NDVIpre = (NIRpre - Redpre) / (NIRpre + Redpre); where NDVIpre represents the vegetation index value of a pixel in the pre-application vegetation index image, NIRpre represents the radiant intensity value of the corresponding pixel in the near-infrared band image of the pre-application image, and Redpre represents the radiant intensity value of the corresponding pixel in the red band image of the pre-application image. The vegetation index value of each pixel in the post-application vegetation index image is calculated as: NDVIpost = (NIRpost - Redpost) / (NIRpost + Redpost); where NDVIpost represents the vegetation index value of a pixel in the post-application vegetation index image, NIRpost represents the radiant intensity value of the corresponding pixel in the near-infrared band image of the post-application image, and Redpost represents the radiant intensity value of the corresponding pixel in the red band image of the post-application image. The vegetation index value reflects the vegetation vitality of the crop canopy at the corresponding pixel location. A higher vegetation index value indicates stronger vegetation vitality, while a lower vegetation index value indicates weaker vegetation vitality or that the pixel corresponds to a non-vegetated surface.

[0029] Threshold segmentation was performed on the pre-application vegetation index image to extract regions with low vegetation vitality. The vegetation vitality segmentation threshold used was determined based on the mean vegetation index of the bare soil background area and the mean vegetation index of the healthy vegetation area in the pre-application vegetation index image. A bare soil background area was selected in the pre-application vegetation index image, corresponding to a portion of the bare soil background covered in the pre-application multispectral image. The mean vegetation index of all pixels within this bare soil background area was calculated as the soil reference value. A healthy crop area was selected in the pre-application vegetation index image, corresponding to a location in the pre-application multispectral image completely covered by healthy crop canopy without disease spots or pesticide damage spots. The mean vegetation index of all pixels within this healthy crop area was calculated as the health reference value. The vegetation vitality segmentation threshold was set to the average of the soil reference value and the health reference value. Regions in the pre-application vegetation index image where the vegetation index value was lower than the vegetation vitality segmentation threshold were extracted as regions with low vegetation vitality before application. Threshold segmentation was performed on the vegetation index image after pesticide application to extract the low vegetation vitality region after pesticide application. The same vegetation vitality segmentation threshold as the vegetation index image before pesticide application was used for threshold segmentation. The region composed of pixels whose vegetation index value was lower than the vegetation vitality segmentation threshold in the vegetation index image after pesticide application was extracted as the low vegetation vitality region after pesticide application.

[0030] The red band image of the pre-application image is used as a visible light texture image for leaf vein texture and shadow recognition. In the red band image of the pre-application image, the local texture intensity of each pixel is calculated. The local texture intensity of a pixel is the standard deviation of the radiant intensity values ​​of the red band image of the pre-application image within its surrounding neighborhood window. The size of the neighborhood window is set based on the ground sampling distance of the pre-application image and the typical spacing of leaf veins on the crop leaf. The ground coverage area corresponding to the size of the neighborhood window is less than 1 / 2 of the typical spacing between two adjacent main veins on the crop leaf; for example, the size of the neighborhood window is 5 pixels by 5 pixels. A leaf vein sample area is selected in the red band image of the pre-application image. The leaf vein sample area contains typical leaf vein texture but does not contain lesions or necrotic spots. The cumulative distribution of the local texture intensity of all pixels in the leaf vein sample area is calculated, and the local texture intensity value corresponding to the 90th percentile of the cumulative distribution is used as the leaf vein texture threshold. In the red band image before application of pesticides, a shadow sample area is selected. The shadow sample area corresponds to the position in the multispectral image before application of pesticides that is in shadow due to canopy shading. The mean value of the radiation intensity of all pixels in the shadow sample area is calculated as the shadow brightness threshold.

[0031] Spatial matching was performed on the low vegetation vitality areas before pesticide application. Each pixel in the low vegetation vitality areas before pesticide application was compared with leaf vein texture thresholds and shadow brightness thresholds. If the local texture intensity of a pixel in the low vegetation vitality area before pesticide application is higher than the leaf vein texture threshold in the red light band image of the image before pesticide application, the pixel is identified as a leaf vein texture pixel and removed from the low vegetation vitality area before pesticide application. If the radiation intensity value of a pixel in the low vegetation vitality area before pesticide application is lower than the shadow brightness threshold in the red light band image of the image before pesticide application, the pixel is identified as a shadow area pixel and removed from the low vegetation vitality area before pesticide application. The low vegetation vitality areas before pesticide application after removing leaf vein texture pixels and shadow area pixels are used as the lesion areas before pesticide application.

[0032] Spatial matching is performed on low vegetation vitality areas after pesticide application. Each pixel in the low vegetation vitality area is compared with a leaf vein texture threshold and a shadow brightness threshold, using the same thresholds as those used before pesticide application. If the local texture intensity of a pixel in the low vegetation vitality area after pesticide application is higher than the leaf vein texture threshold in the red light band image of the post-application image, the pixel is identified as a leaf vein texture pixel and removed from the low vegetation vitality area. If the radiation intensity value of a pixel in the low vegetation vitality area after pesticide application is lower than the shadow brightness threshold in the red light band image of the post-application image, the pixel is identified as a shadow area pixel and removed from the low vegetation vitality area. The low vegetation vitality area after pesticide application, after removing leaf vein texture pixels and shadow area pixels, is designated as the necrotic spot area after pesticide application.

[0033] The process involves extracting low vegetation vitality areas from pre-application vegetation index images and removing leaf vein textures and shadow areas to obtain pre-application lesion areas, and extracting low vegetation vitality areas from post-application vegetation index images and removing leaf vein textures and shadow areas to obtain post-application necrotic spot areas. This method utilizes the common characteristic that both disease infection and pesticide burns lead to a decrease in chlorophyll content in crop leaves, thereby reducing the vegetation index value. This allows both lesion and necrotic spot areas to be segmented and captured by the vegetation index threshold. At the same time, visible light texture matching is used to exclude interference from leaf vein textures on healthy leaves and shadow areas inside the canopy, which also show low vegetation index values ​​in the vegetation index images but are not lesion or necrotic spot areas.

[0034] The leaf vein texture threshold is determined based on the statistical distribution of local texture intensity in the actual leaf vein sample area in the pre-application image. The shadow brightness threshold is determined based on the radiation intensity value of the actual shadow sample area in the pre-application image. The vegetation vitality segmentation threshold is determined based on the mean vegetation index of the actual bare soil background area and the actual healthy crop area in the pre-application vegetation index image. The three thresholds are all directly derived from the sample statistics under the same acquisition conditions as the pre-application image, ensuring that the thresholds match the actual image features. Under different light conditions, different crop varieties, and different growth stages, the thresholds can be updated by reselecting the corresponding sample area, which has practical operability and adaptability.

[0035] Compared to existing methods that rely solely on vegetation index threshold segmentation without removing leaf vein texture and shadows, the interference removal process can more accurately separate the real pre-application lesion areas and post-application necrotic areas from the complex field canopy. This provides an accurate and clean input area for subsequent pixel-crossing cost-based differentiation of continuing lesions and new pesticide-damaged spots, allowing subsequent pixel-crossing cost calculations and neighborhood propagation analysis to be based on the real lesion and necrotic spot boundaries after excluding leaf vein texture and shadow interference.

[0036] In the specific implementation of S3, for each pixel in the post-application necrotic spot region, the shortest path from the pre-application lesion region boundary to that pixel in the post-application necrotic spot region is determined. The process of determining the shortest path is as follows: A distance transformation map of the pre-application lesion region is calculated. This distance transformation map is a two-dimensional image with the same size as the pre-application vegetation index image. The value of each pixel in the distance transformation map represents the shortest spatial distance from that pixel to the pre-application lesion region boundary. The distance transformation map is calculated as follows: all pixels inside the pre-application lesion region are assigned a value of 0. Starting from the pixels with 0 values, the map expands outward layer by layer. For each layer expanded outward, the distance transformation value of the expanded pixels is incremented by 1 until the distance transformation map covers all pixels in the post-application necrotic spot region. This yields the initial distance from each pixel in the post-application necrotic spot region to the pre-application lesion region boundary. This initial distance reflects the spatial proximity of that pixel in the post-application necrotic spot region to the pre-application lesion region. In the distance transformation map, the unit of distance transformation value is the number of pixels. Since the images before and after application of pesticides are spatially registerable, the number of pixels can be converted into the actual ground distance.

[0037] After obtaining the initial distance, the single-point crossing cost is used as the pixel movement weight. The single-point crossing cost is defined as the absolute value of the difference between the vegetation index value in the pre-application vegetation index image and the vegetation index value at the corresponding position in the post-application vegetation index image for the i-th pixel along the path from the boundary of the lesion area before application to a pixel in the necrotic area after application. The single-point crossing cost is calculated as: Ci = |NDVIpre(xi,yi) - NDVIpost(xi,yi)|; where Ci represents the single-point crossing cost of the i-th pixel on the path, NDVIpre(xi,yi) represents the vegetation index value of the i-th pixel at coordinates (xi,yi) in the pre-application vegetation index image, and NDVIpost(xi,yi) represents the vegetation index value of the i-th pixel at coordinates (xi,yi) in the post-application vegetation index image. The cost of single-point crossing reflects the change in vegetation vitality at that pixel location before and after pesticide application: the smaller the change, the closer the vegetation state at that location is before and after pesticide application, and the more likely it is to be an area where plant tissue has not undergone drastic changes during the natural spread of the disease; the larger the change, the more drastic the change in vegetation vitality at that location is after pesticide application, indicating that the location may have suffered tissue truncation due to pesticide burns.

[0038] Based on pixel movement weights, starting from a target pixel in the post-application necrotic spot area, the path moves progressively towards the pre-application lesion boundary along a decreasing initial distance. The path with the smallest cumulative movement weight is selected from all possible paths. During each step, neighboring pixels are checked, and candidate pixels with an initial distance value one less than the current pixel's initial distance value are selected. The cumulative movement weight from the target pixel through each candidate pixel to the pre-application lesion boundary is calculated, and the candidate pixel with the smallest cumulative movement weight is selected as the next pixel on the path. This backtracking process continues until the pre-application lesion boundary is reached, thus obtaining the shortest path for that target pixel. This shortest path is not a spatial Euclidean shortest path, but rather an optimal path that minimizes the cumulative change in vegetation index while progressively increasing spatial distance. This ensures the path traverses areas of relatively stable vegetation vitality before and after application, avoiding areas of sudden increase in vegetation vitality, thus better aligning with the biological process of gradual disease spread in leaf tissue.

[0039] After obtaining the shortest path, the single-point traversal costs of all pixels on the shortest path are summed to obtain the pixel traversal cost of this target pixel. The pixel traversal cost represents the cumulative change in overall vegetation vitality experienced as the lesion area extends from the pre-application boundary along the shortest path to this target pixel. The smaller the cumulative amount, the closer the vegetation tissue state along the path from the lesion boundary to this target pixel is to the pre-application lesion state, and the stronger the tendency for this target pixel to belong to the natural extension of the disease.

[0040] Subsequently, a dual-condition judgment is applied to the pixel crossing cost. This dual-condition judgment involves a preset cost threshold and a preset mutation threshold. The preset cost threshold is determined statistically based on the pixel crossing costs of confirmed lesion expansion pixels and pesticide-damaged pixels in the necrotic area after pesticide application. Specifically, it is set as follows: through field investigation and marking, a group of pixels in the necrotic area after pesticide application are identified as continuing lesions caused by natural disease expansion, and a group of pixels are identified as newly formed pesticide-damaged pixels caused by pesticide burns. The average pixel crossing cost of the continuing lesion pixel group and the average pixel crossing cost of the newly formed pesticide-damaged pixel group are calculated separately, and the average of the two average pixel crossing costs is taken as the preset cost threshold. If the pixel crossing cost of the target pixel is lower than the preset cost threshold, it is preliminarily considered that the target pixel has a tendency to belong to the natural expansion of disease.

[0041] The preset mutation threshold is determined based on the statistical distribution of the single-point crossing cost difference between adjacent pixels on the typical lesion expansion path. Specifically, the method is as follows: extract multiple shortest paths from the boundary of the lesion area before drug application to the pixels continuing the lesion as typical lesion expansion paths, calculate the absolute value of the single-point crossing cost difference between all adjacent pixels on the typical lesion expansion path, statistically analyze the cumulative distribution of the absolute value distribution, and take the absolute value corresponding to the 95th percentile as the preset mutation threshold.

[0042] On the shortest path to the target pixel, the change in single-point crossing cost of adjacent pixels is checked pair by pair. The change in single-point crossing cost is defined as the absolute value of the difference between the single-point crossing costs of two adjacent pixels. If the change in single-point crossing cost of all adjacent pixels on the entire shortest path is lower than the preset mutation threshold, it indicates that the change in vegetation vitality along the path is continuous and gradual, consistent with the progressive necrosis process of lesion infection. If the change in single-point crossing cost of any pair of adjacent pixels on the shortest path is equal to or exceeds the preset mutation threshold, it indicates that a sudden and drastic change in vegetation vitality has occurred at that location, suggesting that pesticide damage may have occurred at that location. Even if the pixel crossing cost is lower than the preset cost threshold, the target pixel should not be marked as a pixel with a tendency to expand.

[0043] When a target pixel simultaneously meets two conditions—that its pixel crossing cost is lower than a preset cost threshold and the change in single-point crossing cost of all adjacent pixels on the shortest path is lower than a preset mutation threshold—this target pixel is marked as an expansion-prone pixel. The expansion-prone pixel marking integrates two dimensions: overall cumulative crossing cost and local mutation detection. It can exclude pixels with low overall cumulative cost but sudden changes in vegetation vitality at some point along the path, accurately identifying the true expansion area along the lesion expansion direction where vegetation vitality declines gradually and is not interrupted by pesticide damage. Transforming the gradual progression of disease and the suddenness of pesticide damage into a quantifiable threshold judgment method, compared to existing technologies that rely solely on spatial continuity or comparison of single-phase vegetation indices, can more accurately distinguish the causes of necrotic spots after pesticide application, providing a more reliable prior for subsequent neighborhood spread.

[0044] In the specific implementation of S4, all pixels in the pre-application lesion area, which serves as the starting seed, are marked as propagated pixels, and a binary labeling map is established. This binary labeling map has the same size as the post-application vegetation index image. Propagated pixels are assigned a value of 1 at their corresponding positions in the binary labeling map, while the remaining pixels are assigned a value of 0 at their corresponding positions. At the initial labeling time, only the pixels corresponding to the pre-application lesion area are set to 1 in the binary labeling map, while the pixels in the post-application necrotic area that do not belong to the pre-application lesion area are set to 0.

[0045] In each iteration of propagation, the 8-neighborhood of each propagated pixel in the current set of propagated pixels is traversed. The 8-neighborhood refers to the 8 directly adjacent pixels centered on the propagated pixel. For candidate pixels in the 8-neighborhood that belong to the expansion tendency pixel but have not yet been marked as propagated pixels, the neighborhood similarity between the propagated pixel and the candidate pixel and the pixel traversal cost difference are calculated.

[0046] Neighborhood similarity is defined as the absolute value of the difference between the vegetation index value of the propagated pixel in the post-application vegetation index image and the vegetation index value of the candidate pixel in the post-application vegetation index image. The neighborhood similarity is calculated as: S = |NDVIpost(xs,ys) - NDVIpost(xc,yc)|; where S represents the neighborhood similarity, NDVIpost(xs,ys) represents the vegetation index value of the propagated pixel at coordinates (xs,ys) in the post-application vegetation index image, and NDVIpost(xc,yc) represents the vegetation index value of the candidate pixel at coordinates (xc,yc) in the post-application vegetation index image. Neighborhood similarity reflects the degree of similarity between the propagated pixel and the candidate pixel in terms of vegetation vitality after application. A smaller neighborhood similarity value indicates that the vegetation vitality of the propagated pixel and the candidate pixel are more similar, and they are more likely to belong to the same continuous necrosis type.

[0047] The similarity threshold is determined based on the consistency of vegetation index values ​​within the continuous lesion area. Specifically, it is set as follows: Multiple pixel pairs are randomly selected from the pixels within the lesion area before application (serving as the starting seed). The absolute value of the difference between the vegetation index values ​​of the two pixels in each pair in the vegetation index image after application is calculated. The cumulative distribution of these absolute values ​​is then calculated, and the absolute value corresponding to the highest percentile of the cumulative distribution is taken as the similarity threshold. For example, the absolute value corresponding to the 90th percentile of the cumulative distribution is used as the similarity threshold. The similarity threshold reflects the normal fluctuation range of vegetation vitality in the vegetation index image after application within the lesion area. A difference in vegetation index below the similarity threshold is considered a normal fluctuation within the same type of necrotic tissue.

[0048] The pixel traversal cost difference is defined as the absolute value of the difference between the pixel traversal cost of a candidate pixel and the pixel traversal cost of a propagated pixel. The pixel traversal cost of both the candidate and propagated pixels has been obtained during the pixel traversal cost calculation process. If the pixel traversal cost difference is lower than a preset mutation threshold, it indicates that the increase in pixel traversal cost between the candidate pixel and the propagated pixel has not exceeded the allowable upper limit for gradual lesion expansion, and no mutation truncation has occurred on the expansion path. The preset mutation threshold is the same as the preset mutation threshold used when marking pixels with expansion tendency.

[0049] For candidate pixels that meet both conditions—neighborhood similarity below a similarity threshold and pixel traversal cost difference below a preset mutation threshold—the candidate pixels are marked as propagated pixels. This involves assigning a value of 1 to the corresponding position in the binary labeling image and adding the candidate pixel to the propagated pixel set. After completing the 8-neighborhood processing for all propagated pixels in this iteration, the next iteration continues, repeating the traversal, calculation, and marking process until no new pixels are added to the propagated pixel set in a single iteration, at which point the iteration stops.

[0050] When the iteration stops, the spatial area covered by all pixels with a value of 1 in the binary labeled image is the continuous lesion region. The continuous lesion region represents the necrotic part that spreads outward from the boundary of the lesion region before application, maintains consistency with the lesion region before application in terms of vegetation vitality characteristics after application, and has a gradual change in crossing cost. The remaining pixels in the necrotic lesion region after application, excluding the continuous lesion region, are classified as the new pesticide-damaged lesion region. The new pesticide-damaged lesion region is located outside the abrupt change in crossing cost. The vegetation vitality characteristics of the new pesticide-damaged lesion region are significantly different from those of the lesion region before application, or the crossing cost changes abruptly. The new pesticide-damaged lesion region corresponds to tissue necrosis directly caused by pesticide burns.

[0051] The neighborhood propagation process organically combines spatial continuity constraints, post-application vegetation vitality similarity constraints, and pixel traversal cost smoothness constraints. Existing techniques that only utilize spatial connectivity partitioning methods classify all spatially connected necrotic spots as continuing lesions, ignoring situations where pesticide burns may be adjacent to lesions but not formed by lesion expansion. By adding neighborhood similarity constraints, burn areas with significantly different vegetation vitality from the pre-application lesion areas can be removed. Further, by adding pixel traversal cost difference constraints, peripheral burn areas that cross the traversal cost mutation boundary can be excluded. The causes of post-application necrotic spot areas are analyzed from spatial, spectral, and temporal evolution dimensions, distinguishing between continuing lesion areas and newly formed pesticide-damaged areas, providing a more accurate regional partitioning basis for subsequent boundary neighborhood color attenuation verification and image classification evaluation.

[0052] In the specific implementation of S5, the boundary pixels of the continuous lesion region are extracted. A boundary pixel is a pixel in the continuous lesion region that has at least one adjacent pixel that does not belong to the continuous lesion region. Each pixel in the continuous lesion region is traversed, and the 8-neighborhood of each pixel is checked. If at least one pixel in the 8-neighborhood of a certain pixel does not belong to the continuous lesion region, then this pixel is recorded as a boundary pixel, and all boundary pixels constitute the boundary pixel set.

[0053] For each boundary pixel in the boundary pixel set, an outward neighborhood expansion is performed to obtain the boundary neighborhood pixel band. The specific method for outward neighborhood expansion is as follows: the gradient direction of the boundary pixel in the vegetation index image after pesticide application is calculated as the expansion direction. The gradient direction is calculated based on the gradient of the vegetation index image after pesticide application, which includes a horizontal gradient component and a vertical gradient component. The horizontal gradient component is obtained by convolving the vegetation index image after pesticide application with a horizontal Sobel kernel, and the vertical gradient component is obtained by convolving the vegetation index image after pesticide application with a vertical Sobel kernel. The gradient direction is calculated based on the horizontal and vertical gradient components as: θ = atan2(Gy, Gx); where θ represents the gradient direction of the boundary pixel, Gy represents the vertical gradient component, Gx represents the horizontal gradient component, and atan2 represents the four-quadrant arctangent function. The gradient direction reflects the method by which the vegetation index value changes most drastically after pesticide application at the boundary pixel. At the transition point between normal and necrotic tissue, the vegetation index value changes most dramatically at the boundary. The gradient direction points from the area with high vegetation index values ​​to the area with low vegetation index values, that is, from the inside of the continuous lesion area to the outside of normal tissue. Expanding outward along this direction can accurately capture the plant tissue area located outside the boundary of the continuous lesion area.

[0054] After obtaining the gradient direction, pixels of a preset width are continuously sampled outward from the boundary pixel along the gradient direction. The preset width is set based on the ground sampling distance of the post-application image and the typical width of the chlorotic halo surrounding the pesticide burn spot. The chlorotic halo refers to the light yellow transition zone formed around the leaf cell necrosis spot caused by pesticide burn due to uneven degradation of chlorophyll. The typical width of the chlorotic halo is related to the type of pesticide, application concentration, and crop leaf structure. The actual ground distance corresponding to the preset width matches the typical width of the chlorotic halo; for example, the preset width is set to 3 pixels. When continuously sampling pixels along the gradient direction, starting from the boundary pixel, the process moves pixel by pixel along the gradient direction. Each move records the pixel corresponding to the new position. After moving the preset width a number of times, the process stops, and all pixels passed during the movement are taken as the extended pixels corresponding to this boundary pixel. The above operation is repeated for all boundary pixels in the boundary pixel set. The extended pixels corresponding to all boundary pixels are merged, and duplicate pixels are removed to form the boundary neighborhood pixel band.

[0055] After obtaining the boundary neighborhood pixel band, the color decay rate of each pixel in the boundary neighborhood pixel band is calculated. The color decay rate is defined as the absolute value of the difference between the vegetation index value of a pixel at the corresponding position in the pre-application vegetation index image and the vegetation index value at the corresponding position in the post-application vegetation index image. The color decay rate is calculated as: D = |NDVIpre(xd,yd) - NDVIpost(xd,yd)|; where D represents the color decay rate, NDVIpre(xd,yd) represents the vegetation index value of a pixel at coordinates (xd,yd) in the pre-application vegetation index image, and NDVIpost(xd,yd) represents the vegetation index value of the same pixel at coordinates (xd,yd) in the post-application vegetation index image. The color decay rate reflects the decrease in vegetation vitality at a pixel position in the boundary neighborhood pixel band from before to after application; a larger decrease indicates a more severe decrease in vegetation vitality at that position.

[0056] The color attenuation threshold is used to distinguish between the normal decline in vegetation vitality at the edge of a continuing lesion area and the decline in vegetation vitality caused by the characteristic chlorotic halo at the edge of pesticide burns. The specific method for setting the color attenuation threshold is as follows: A reference edge area is selected within the lesion area before pesticide application, located inside the boundary of the lesion area before application. This reference edge area consists of pixels that belong to the lesion area before application and are still classified as part of the continuing lesion area after application. The absolute value of the difference between the vegetation index value at the corresponding position in the vegetation index image before application and the vegetation index value at the corresponding position in the vegetation index image after application for each pixel within the reference edge area is calculated as the reference attenuation value. The cumulative distribution of all reference attenuation values ​​is then calculated, and the reference attenuation value corresponding to a higher percentile of the cumulative distribution is used as the color attenuation threshold. For example, the reference attenuation value corresponding to the 95th percentile of the cumulative distribution is used as the color attenuation threshold. The reference edge area represents the boundary portion of the lesion area that still belongs to the continuing lesion area after application, and the reference attenuation value reflects the range of normal color attenuation in the boundary area during the natural expansion or regression of the lesion.

[0057] The color decay of each pixel in the boundary neighborhood pixel band is compared with the color decay threshold. If the color decay of a pixel in the boundary neighborhood pixel band exceeds the color decay threshold, this pixel is identified as a boundary neighborhood color decay aberration pixel. The vegetation vitality decline of boundary neighborhood color decay aberration pixels significantly exceeds the normal color decay level of the boundary of the continuous lesion area, exhibiting typical chlorotic halo characteristics. This indicates that the location of this pixel experienced drastic vegetation vitality changes beyond the natural succession range of lesions before and after pesticide application, consistent with edge tissue damage caused by pesticide burns. All pixels identified as boundary neighborhood color decay aberration pixels are removed from the continuous lesion area, and the removed pixels are added to the newly formed pesticide-damaged lesion area to obtain the corrected continuous lesion area and the corrected newly formed pesticide-damaged lesion area. The pixels with abnormal color decay in the boundary neighborhoods that were originally classified as part of the continuing lesion area were confirmed by color decay verification to have a vegetation vitality decline pattern that is inconsistent with the gradual change in the continuing lesion area. They are more consistent with the edge chlorosis caused by pesticide burns. Therefore, they were corrected and classified into the new pesticide burn area. This means that the corrected continuing lesion area only includes the necrotic parts that are truly part of the natural spread of the disease, and the corrected new pesticide burn area completely includes all the necrotic areas caused by pesticide burns.

[0058] The boundary neighborhood color attenuation verification process utilizes the unique chlorotic halo phenomenon surrounding pesticide burn spots. At the chlorotic halo, the vegetation index value gradually recovers to normal levels from the necrotic center outwards. However, the vegetation index value in the transition zone before recovery is significantly lower than that at the same location before pesticide application, forming a color attenuation characteristic specific to pesticide damage. Naturally expanding disease edges typically do not form uniformly wide and significantly color-attenuated chlorotic transition zones because the vegetation vitality decline pattern at the disease expansion front is progressive tissue necrosis, which is fundamentally different in spatial gradient from the chemical burn pattern caused by pesticide penetration from the tissue surface inwards. By detecting color attenuation along the expansion direction defined by the gradient direction, compared to analysis in a fixed window or random direction, the true color attenuation characteristics from the boundary of the continuing lesion area outwards along the direction of the most drastic change in vegetation vitality can be captured more accurately. This effectively identifies and corrects boundary pesticide-damaged pixels that are mistakenly classified as continuing lesion areas, providing more accurate corrections for the area of ​​continuing lesion areas and newly formed pesticide-damaged areas for subsequent image classification and evaluation.

[0059] In the specific implementation of S6, the pixel area of ​​the corrected continuing lesion region is calculated as the continuing lesion area, and the total number of pixels contained in the corrected continuing lesion region is counted. The continuing lesion area is expressed in pixels. The pixel area of ​​the lesion region before application is calculated as the pre-application lesion area, and the total number of pixels contained in the pre-application lesion region is counted. The pre-application lesion area is expressed in pixels. The rate of change of the continuing lesion area is calculated as the ratio of the continuing lesion area to the pre-application lesion area: Rcontinue = Acontinue / Apre; where Rcontinue represents the rate of change of the continuing lesion area, Acontinue represents the continuing lesion area, and Apre represents the pre-application lesion area. The rate of change of the continuing lesion area is a dimensionless ratio. A rate of change of 1 indicates that the lesion area remains basically unchanged before and after application. A rate of change of less than 1 indicates that the lesion area has decreased after application compared to the pre-application lesion area. A rate of change of more than 1 indicates that the lesion area continues to expand after application compared to the pre-application lesion area.

[0060] The pixel area of ​​the newly formed pesticide-damaged area is calculated as the area of ​​the newly formed pesticide-damaged area. The total number of pixels contained in the newly formed pesticide-damaged area is counted, and the area of ​​the newly formed pesticide-damaged area is expressed in pixels. The pixel area of ​​the necrotic area after pesticide application is calculated as the total area of ​​the necrotic area. The total area of ​​the necrotic area after pesticide application is counted, and the total area of ​​the necrotic area is expressed in pixels. The proportion of the newly formed pesticide-damaged area is calculated as the ratio of the area of ​​the newly formed pesticide-damaged area to the total area of ​​the necrotic area: Rnew = Anew / Atotal; where Rnew represents the proportion of the newly formed pesticide-damaged area, Anew represents the area of ​​the newly formed pesticide-damaged area, and Atotal represents the total area of ​​the necrotic area. The proportion of the newly formed pesticide-damaged area is a dimensionless ratio. The closer the proportion of the newly formed pesticide-damaged area is to 1, the higher the proportion of necrotic parts caused by pesticide burns in the necrotic area after pesticide application. The closer the proportion of the newly formed pesticide-damaged area is to 0, the higher the proportion of necrotic parts caused by disease expansion in the necrotic area after pesticide application.

[0061] Image classification is used to output pesticide application efficacy evaluation results based on the rate of change in persistent lesion area and the proportion of newly formed pesticide-damaged spots. The image classification employs a two-dimensional feature space, with the two dimensions being the rate of change in persistent lesion area and the proportion of newly formed pesticide-damaged spots. The image classification categories include four categories: significant efficacy, moderate efficacy, ineffective efficacy, and pesticide damage occurrence. The classification boundaries are determined through a training sample set. The training sample set is constructed by collecting multiple sets of field data with known pesticide application effects. Each set of field data includes the actual effect label confirmed by field investigation, the corresponding rate of change in the area of ​​persistent lesions, and the corresponding proportion of the area of ​​newly formed pesticide-damaged spots. The actual effect label is one of the following: a label indicating significant efficacy, a label indicating moderate efficacy, a label indicating ineffective efficacy, or a label indicating pesticide damage. The rate of change in the area of ​​persistent lesions and the proportion of the area of ​​newly formed pesticide-damaged spots corresponding to different actual effect labels are plotted as scatter plots in a two-dimensional feature space. The classification boundaries between the scatter plot distribution areas of different actual effect labels are then defined in the two-dimensional feature space. The classification boundaries divide the two-dimensional feature space into four classification regions, corresponding to the categories of significant efficacy, moderate efficacy, ineffective efficacy, and pesticide damage, respectively.

[0062] The classification category boundaries are set as follows: calculate the mean change rate of the area of ​​the continuous lesion and the mean proportion of the area of ​​the newly formed drug-damaged lesion corresponding to each actual effect label. Based on the mean points corresponding to each actual effect label, determine the direction and cutoff position of the classification category boundaries in the two-dimensional feature space according to the distance between adjacent mean points and the dispersion of the scatter points, so that the scatter point density inside each classification category region is the highest and the scatter point overlap between different classification category regions is the lowest. For example, in field surveys confirming plots with significant efficacy, the average rate of change in the area of ​​persistent lesions is concentrated in the range of less than 0.5, and the average proportion of newly formed pesticide-damaged spots is concentrated in the range of less than 0.1, indicating that the lesions have significantly subsided after application and almost no pesticide-damaged spots have occurred. In field surveys confirming plots with moderate efficacy, the average rate of change in the area of ​​persistent lesions is concentrated in the range of 0.5 to 0.8, and the average proportion of newly formed pesticide-damaged spots is concentrated in the range of less than 0.1, indicating that the lesions have subsided somewhat after application, but the subsidence is not sufficient. In field surveys confirming plots with ineffective efficacy, the average rate of change in the area of ​​persistent lesions is concentrated in the range of greater than 0.8, and the average proportion of newly formed pesticide-damaged spots is concentrated in the range of less than 0.1, indicating that the lesions have not been suppressed and continue to expand after application. In field surveys confirming plots with pesticide damage, the average proportion of newly formed pesticide-damaged spots is concentrated in the range of greater than 0.1, indicating that obvious pesticide burn and necrosis spots appeared after application, and regardless of the value of the rate of change in the area of ​​persistent lesions, it is determined that pesticide damage has occurred.

[0063] The rate of change of the area of ​​persistent lesions and the proportion of the area of ​​newly formed pesticide-damaged spots in the plot to be evaluated are combined into a two-dimensional feature point. The two-dimensional feature point is mapped into a two-dimensional feature space to determine the classification category region where the two-dimensional feature point is located. The classification category corresponding to the classification category region where the two-dimensional feature point is located is used as the pesticide application effect evaluation result output. The pesticide application effect evaluation result is in the form of a label with obvious efficacy, a label with moderate efficacy, a label with ineffective efficacy, or a label with pesticide damage.

[0064] The image classification process combines the rate of change in the area of ​​persistent lesions with the proportion of newly formed pesticide-damaged spots as classification criteria. It simultaneously assesses the inhibitory effect of pesticides on diseases and the potential damage risk to crops. Compared to existing methods that classify based solely on either the rate of change in the area of ​​persistent lesions or the proportion of newly formed pesticide-damaged spots, this approach can identify different agronomic scenarios, such as lesion regression accompanied by pesticide damage and lesions that are uncontrolled without obvious pesticide damage. This results in a more objective and comprehensive evaluation of pesticide application effectiveness. The areas of persistent lesions and newly formed pesticide-damaged spots are derived from corrected persistent lesion regions and corrected newly formed pesticide-damaged spot regions, respectively. These regions are the final region division results after multiple corrections, including pixel crossing cost analysis, neighborhood propagation, and boundary neighborhood color attenuation verification. The rate of change in the area of ​​persistent lesions and the proportion of newly formed pesticide-damaged spots input into the image classification have eliminated interference from leaf vein texture and shadows, excluded pesticide-damaged truncated pixels at abrupt changes in crossing cost, and corrected misclassified pixels with chlorotic halos at the boundaries. This ensures that the classification evaluation is based on accurate plant feature analysis.

[0065] Example 2: Figure 2 A schematic diagram of a pesticide application efficacy evaluation system based on plant characteristic analysis is provided. The system includes: The image acquisition module is used to acquire images before and after application of the pesticide. The semantic segmentation module is used to perform image semantic segmentation on the image before drug application to obtain the lesion region before drug application; and to perform image semantic segmentation on the image after drug application to obtain the necrotic region after drug application. The cost labeling module is used to calculate the pixel crossing cost along the path from the boundary of the lesion area before drug application to each pixel in the necrotic area after drug application, and to label pixels with continuously low crossing costs and no abrupt changes as pixels with expansion tendency. The region division module is used to propagate pixels with the tendency to expand in the neighborhood of the necrotic spots after the application of pesticides, starting with the diseased spots before pesticide application. The propagation conditions are constrained by the pixel crossing cost and neighborhood similarity. Propagation stops at the point where the crossing cost changes. The area covered by the propagation is divided into the continuous diseased spot area and the rest is divided into the new pesticide-damaged spot area. The color verification module is used to verify the color decay of the boundary neighborhood of the continuous lesion area based on the pre-application image. Pixels with abnormal color decay at the boundary neighborhood are removed from the continuous lesion area and assigned to the newly damaged lesion area, thus obtaining the corrected continuous lesion area and the corrected newly damaged lesion area. The image classification module is used to evaluate the effectiveness of pesticide application by classifying images and outputting results based on the area change rate of the corrected continuous lesion area and the area ratio of the corrected new pesticide-damaged area.

[0066] All calculations involved in the embodiments are performed using dimensionless numerical values, and the preset parameters and thresholds in the calculations can be set by those skilled in the art according to actual conditions.

[0067] This technical solution can be flexibly deployed, for example, as embedded software running on device hardware, or installed on personal computers or other smart terminals with user interfaces, thus adapting to various hardware environments and usage requirements.

[0068] The above solutions can be implemented in software, hardware, firmware, or a combination thereof. When implemented in software, they are presented, in whole or in part, as a computer program product, containing one or more computer instructions or programs. When these instructions or programs are loaded and executed on a computer, results are produced corresponding to the processes or functions of the embodiments of this application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one medium to another via wired or wireless means, such as from a website, server, or data center via wired means such as fiber optic cables, twisted-pair cables, or coaxial cables, or wireless means such as infrared or microwaves to another site. A computer-readable storage medium refers to any usable medium that a computer can access or a data storage device such as a server or data center that contains one or more usable media, including magnetic media such as floppy disks, hard disks, and magnetic tapes, optical media such as DVDs, and semiconductor media such as solid-state drives.

[0069] The specific working process of the system, device and module can be found in the method embodiment, and will not be repeated here.

[0070] The disclosed systems, devices, and methods can be implemented in other ways. The device embodiments are for illustrative purposes only, and the module division is only a logical division. In practice, different divisions can be implemented, such as merging or integrating multiple modules or components, or omitting some features. Coupling, direct coupling, or communication connections between the components can be achieved through interfaces, while indirect coupling or communication connections can take electrical, mechanical, or other forms.

[0071] The modules described as separate components may or may not be physically separated. The components shown as modules can be physical hardware or software, and can be deployed centrally or distributed across multiple network nodes. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0072] The functional modules in each embodiment can be integrated into one processing module, or they can exist independently, or two or more modules can be integrated into one.

[0073] If the functionality is implemented as a software module and used as an independent product, it can be stored in a computer-readable storage medium. Under this understanding, the substantial contribution of the technical solution of this application can be embodied in the form of a software product. This computer software product is stored in a storage medium and contains instructions to cause a computer device, such as a personal computer, server, or network device, to execute all or part of the steps of the methods in the embodiments of this application. The storage medium includes any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory, random access memory, magnetic disk, or optical disk.

[0074] The above are merely specific embodiments of this application, and the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should fall within the scope of protection of this application.

Claims

1. A method for evaluating pesticide application efficacy based on plant characteristic analysis, characterized in that, include: S1: Acquire images before and after application of the pesticide; S2: Perform image semantic segmentation on the image before application to obtain the lesion region before application; perform image semantic segmentation on the image after application to obtain the necrotic region after application; S3: Calculate the pixel crossing cost along the path from the boundary of the lesion area before drug application to each pixel in the necrotic area after drug application, and mark pixels with continuously low crossing costs and no sudden changes as pixels with expansion tendency. S4: Starting with the diseased area before application of pesticide, the pixels with the tendency to expand are spread in the neighborhood of the necrotic area after application of pesticide. The spread conditions are constrained by the pixel crossing cost and neighborhood similarity. The spread stops at the point where the crossing cost changes. The area covered by the spread is divided into the continuous diseased area and the rest is divided into the new pesticide-damaged area. S5: Based on the image before application, perform boundary neighborhood color attenuation verification on the continuous lesion area, remove pixels with abnormal boundary neighborhood color attenuation from the continuous lesion area and assign them to the newly formed drug-damaged lesion area, and obtain the corrected continuous lesion area and the corrected newly formed drug-damaged lesion area. S6: Using image classification, based on the area change rate of the corrected continuous lesion area and the area proportion of the corrected new pesticide-damaged area, output the pesticide application effect evaluation results.

2. The method for evaluating pesticide application efficacy based on plant characteristic analysis according to claim 1, characterized in that, Acquire images before and after application of the pesticide, including: Multispectral images of the crop canopy before pesticide application were acquired using multispectral imaging equipment and used as pre-application images. Collect multispectral images after pesticide application as post-application images; Both the pre-application and post-application multispectral images include red and near-infrared band images.

3. The method for evaluating pesticide application effectiveness based on plant characteristic analysis according to claim 1, characterized in that, Image semantic segmentation is performed on the pre-application image to obtain the pre-application lesion region; image semantic segmentation is performed on the post-application image to obtain the post-application necrotic lesion region, including: Band operations were performed on the red and near-infrared band images of the images before and after pesticide application to generate vegetation index images before and after pesticide application. Threshold segmentation was performed on the vegetation index images before and after pesticide application to extract areas with low vegetation vitality. Spatial matching was performed between the low vegetation vitality area in the image before pesticide application and the visible light texture in the image before pesticide application. Leaf vein texture and shadow areas were removed to obtain the diseased area before pesticide application. Spatial matching was performed between the low vegetation vitality areas in the post-application image and the visible light texture of the post-application image. Leaf vein texture and shadow areas were removed to obtain the necrotic spot areas after application.

4. The method for evaluating pesticide application efficacy based on plant characteristic analysis according to claim 1, characterized in that, Along the path from the boundary of the lesion area before drug application to each pixel within the necrotic area after drug application, the pixel traversal cost is calculated. Pixels with continuously low traversal costs and no abrupt changes are marked as pixels with a tendency to expand, including: For each pixel in the necrotic patch region after drug application, determine the shortest path from the boundary of the lesion region before drug application to that pixel; Along the shortest path, the difference in vegetation index at the corresponding position between the vegetation index image before and after pesticide application is calculated pixel by pixel and used as the single-point crossing cost. The pixel crossing cost is obtained by summing the single-point crossing costs on the shortest path. If the cost of a pixel crossing is lower than a preset cost threshold and the change in the cost of a single point crossing of adjacent pixels on the shortest path is lower than a preset mutation threshold, then the pixel is marked as an expansion-prone pixel.

5. The method for evaluating pesticide application effectiveness based on plant characteristic analysis according to claim 4, characterized in that, Determining the shortest path from the boundary of the lesion area before drug application to the pixel includes: calculating the distance transformation map of the lesion area before drug application to obtain the initial distance from each pixel in the necrotic lesion area after drug application to the boundary of the lesion area before drug application; using the single-point crossing cost as the pixel movement weight, starting from the pixel, moving step by step along the direction where the initial distance decreases and the cumulative movement weight is minimized to the boundary of the lesion area before drug application to obtain the shortest path.

6. The method for evaluating pesticide application efficacy based on plant characteristic analysis according to claim 1, characterized in that, Starting with the diseased area before pesticide application as the initial seed, pixels with expansion tendency are propagated in the necrotic area after pesticide application. The propagation conditions are constrained by pixel crossing cost and neighborhood similarity. Propagation stops at the point where the crossing cost changes abruptly. The area covered by the propagation is divided into the continuous diseased area, and the remaining area is divided into the newly formed pesticide-damaged area, including: The pixels contained in the diseased area before application, which serves as the starting seed, are marked as propagated pixels; The vegetation index difference between the propagated pixel and its neighboring pixels with the tendency to expand is calculated based on the vegetation index image after pesticide application as the neighborhood similarity. Pixels with a neighborhood similarity lower than a similarity threshold and whose pixel traversal cost is lower than a preset mutation threshold are added to the already propagated pixels. Repeat the process until no new pixels are added. The area covered by the already spread pixels is considered as the continued lesion area, and the area in the necrotic lesion area after application that is not covered by the already spread pixels is considered as the newly formed phytotoxic lesion area.

7. The method for evaluating pesticide application efficacy based on plant characteristic analysis according to claim 1, characterized in that, Based on the pre-application image, boundary neighborhood color attenuation is checked in the continuous lesion area. Pixels with abnormal color attenuation in the boundary neighborhood are removed from the continuous lesion area and assigned to the newly formed pesticide-damaged lesion area, resulting in corrected continuous lesion areas and corrected newly formed pesticide-damaged lesion areas, including: Extract the boundary pixels of the continuous lesion region; Extend the neighborhood outwards along the boundary pixels to obtain the boundary neighborhood pixel band; Based on the vegetation index images before and after application of the pre-application image, the vegetation index difference between each pixel in the boundary neighborhood pixel band and the pre-application and post-application images is calculated as the color attenuation. Pixels whose color decay exceeds the color decay threshold are identified as pixels with abnormal color decay in the boundary neighborhood; Pixels with abnormal color decay in the boundary neighborhood are removed from the continuous lesion area and added to the newly formed drug-damaged lesion area to obtain the corrected continuous lesion area and the corrected newly formed drug-damaged lesion area.

8. The method for evaluating pesticide application efficacy based on plant characteristic analysis according to claim 7, characterized in that, The neighborhood is expanded outward along the boundary pixel to obtain the boundary neighborhood pixel band, including: for each boundary pixel in the continuous lesion area, the gradient direction of the boundary pixel in the vegetation index image after the application of the pesticide is calculated as the expansion direction; pixels of a preset width are continuously taken outward from the boundary pixel along the expansion direction, and the expansion pixels corresponding to all boundary pixels are merged to form the boundary neighborhood pixel band.

9. The method for evaluating pesticide application efficacy based on plant characteristic analysis according to claim 1, characterized in that, Using image classification, based on the area change rate of corrected persistent lesions and the area proportion of corrected newly formed pesticide-damaged lesions, the pesticide application effect evaluation results are output, including: The pixel area of ​​the corrected continuation lesion region is calculated as the continuation lesion area, and the pixel area of ​​the lesion region before application of the drug is calculated as the lesion area before application of the drug. The ratio of the continuation lesion area to the lesion area before application of the drug is taken as the change rate of the continuation lesion area. The pixel area of ​​the newly formed pesticide-damaged area is calculated as the area of ​​the newly formed pesticide-damaged area, and the pixel area of ​​the necrotic area after pesticide application is calculated as the total area of ​​the necrotic area. The ratio of the area of ​​the newly formed pesticide-damaged area to the total area of ​​the necrotic area is taken as the proportion of the area of ​​the newly formed pesticide-damaged area. Using image classification, the pesticide application effect evaluation results are output based on the rate of change of continuous lesion area and the proportion of newly formed pesticide-damaged area.

10. A pesticide application effect evaluation system based on plant characteristic analysis, used to implement the pesticide application effect evaluation method based on plant characteristic analysis as described in any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire images before and after application of the pesticide. The semantic segmentation module is used to perform image semantic segmentation on the image before drug application to obtain the lesion region before drug application; and to perform image semantic segmentation on the image after drug application to obtain the necrotic region after drug application. The cost labeling module is used to calculate the pixel crossing cost along the path from the boundary of the lesion area before drug application to each pixel in the necrotic area after drug application, and to label pixels with continuously low crossing costs and no abrupt changes as pixels with expansion tendency. The region division module is used to propagate pixels with the tendency to expand in the neighborhood of the necrotic spots after the application of pesticides, starting with the diseased spots before pesticide application. The propagation conditions are constrained by the pixel crossing cost and neighborhood similarity. Propagation stops at the point where the crossing cost changes. The area covered by the propagation is divided into the continuous diseased spot area and the rest is divided into the new pesticide-damaged spot area. The color verification module is used to verify the color decay of the boundary neighborhood of the continuous lesion area based on the pre-application image. Pixels with abnormal color decay at the boundary neighborhood are removed from the continuous lesion area and assigned to the newly damaged lesion area, thus obtaining the corrected continuous lesion area and the corrected newly damaged lesion area. The image classification module is used to evaluate the effectiveness of pesticide application by classifying images and outputting results based on the area change rate of the corrected continuous lesion area and the area ratio of the corrected new pesticide-damaged area.

Images