Low-altitude remote sensing monitoring system for phenotypes of field crops

Abstract

The invention relates to the technical field of remote sensing image edge enhancement, in particular to a low-altitude remote sensing monitoring system for field crop phenotypes. The system comprises a remote sensing image acquisition module used for acquiring hyperspectral remote sensing images of field crops; the lesion area enhancement module is used for dividing the hyperspectral remote sensing image of the visible light wave band into a healthy leaf area, a slight lesion area and a serious lesion area; further dividing the serious lesion area into a real serious lesion area and a soil area; the field crop monitoring module is used for fusing a Sobel operator and a Krisch operator in the process of calculating the gradient of each pixel point when carrying out edge detection on the hyperspectral remote sensing image of the visible light band by utilizing a Canny edge detection algorithm, and monitoring the growth state of field crops by utilizing a neural network model in combination with the hyperspectral remote sensing image of each time period and an edge detection result; the accuracy of field crop growth state monitoring is improved.

Description

Technical Field

[0001] The present application relates to the technical field of remote sensing image edge enhancement, and in particular to a low-altitude remote sensing monitoring system for field crop phenotypes. Background Art

[0002] In modern agricultural production, accurate and efficient monitoring of crop growth status is of great significance for increasing yields, optimizing resource utilization, and achieving sustainable agricultural development. Crop phenotypic information is a key indicator of its health and environmental adaptability. Traditional ground-based monitoring technologies are difficult to meet the needs of modern farmland, which restricts the implementation of precision agriculture. With the development of remote sensing technology and intelligent agricultural equipment, low-altitude remote sensing monitoring based on drones can be used as an effective strategy for modern farmland crop monitoring and accurate grasp of farmland information. Hyperspectral remote sensing technology captures high-resolution reflectance spectra of crop canopies in the visible to near-infrared bands, which can invert key biophysical and chemical parameters such as chlorophyll content, water stress, and nitrogen distribution, revealing crop growth trends and environmental response mechanisms, and realizing intelligent and precise monitoring of crops in modern large fields.

[0003] When monitoring crop growth using hyperspectral remote sensing imagery, edge detection can help extract the outlines and edges of objects in the image and enhance image edge features, which is important for analyzing crop morphology and structural changes. In hyperspectral imagery, edge information can be closely related to the spatial distribution of chlorophyll content, especially in different leaf regions or at different growth stages. Chlorophyll distribution changes are correlated with the edge characteristics of crop morphology.

[0004] However, the commonly used Canny edge detection algorithm primarily focuses on horizontal and vertical edges, often used to capture the coarse structure of image edges. It lacks multi-directional and refined edge information, and performs poorly in complex or noisy image regions, especially when noise levels are high or background textures are complex. This low edge detection accuracy can reduce the accuracy of feature extraction when using hyperspectral remote sensing imagery to monitor crop growth, resulting in poor image edge enhancement and errors in crop growth monitoring results. Summary of the Invention

[0005] In order to solve the above technical problems, the purpose of this application is to provide a low-altitude remote sensing monitoring system for field crop phenotypes. The technical solutions adopted are as follows: This application proposes a low-altitude remote sensing monitoring system for field crop phenotypes, the system comprising: Remote sensing image acquisition module, which collects hyperspectral remote sensing images of field crops in the same area at different time periods; The lesion area enhancement module divides the hyperspectral remote sensing image in the visible light band into healthy leaf areas, slightly lesion areas, and severely lesion areas based on the grayscale difference of each pixel in the hyperspectral remote sensing image in the visible light band for each time period; Determining the intensity of change in the grayscale of pixels in the neighborhood of each pixel in the severe lesion area in each time period over multiple historical time periods; performing skeleton extraction on the severe lesion area in each time period to obtain skeleton images of the severe lesion area, and determining the average skeleton width of each skeleton image; Analyze the change trend of the average skeleton width of each skeleton graph in the severe lesion area in each time period and the corresponding matched skeleton graph in the hyperspectral remote sensing images of multiple historical time periods, and combine the change intensity of the pixel points in the adjacent areas of each skeleton graph to divide the severe lesion area in each time period into a real severe lesion area and a soil area; The field crop monitoring module determines the grayscale thresholds that can distinguish each region based on the regions that have been divided in the hyperspectral remote sensing image. When using the Canny edge detection algorithm to perform edge detection on the hyperspectral remote sensing image in the visible light band, the Sobel operator and the Krisch operator are integrated in the process of calculating the gradient of each pixel to obtain the optimized gradient value of each pixel. The grayscale thresholds are used to determine the distribution weight of the Sobel operator and the Krisch operator when calculating the gradient of each pixel in each region of the hyperspectral remote sensing image. The hyperspectral remote sensing images of each time period are combined with the edge detection results and a neural network model is used to monitor the growth status of field crops.

[0006] In one embodiment, dividing the hyperspectral remote sensing image in the visible light band into healthy leaf areas, slightly lesion areas, and severely lesion areas includes: The histogram analysis of the hyperspectral remote sensing image in the visible light band is performed, and the multi-threshold processing method is used to obtain two grayscale thresholds, which are recorded as T1 and T2, respectively, where T1 < T2; Based on the thresholds T1 and T2, the pixels in the hyperspectral remote sensing image in the visible light band are divided into healthy leaf areas, slightly lesion areas, and severely lesion areas.

[0007] In one embodiment, the pixel area with a grayscale value less than or equal to T1 in the hyperspectral remote sensing image in the visible light band is regarded as a healthy leaf area, the pixel area with a grayscale value greater than T1 and less than T2 is regarded as a mild lesion area, and the pixel area with a grayscale value greater than or equal to T2 is regarded as a severe lesion area.

[0008] In one embodiment, the expression of the variation intensity is: Where, is the change intensity of the p-th pixel in the severe lesion area in the hyperspectral remote sensing image of the visible light band in the t-th time period, is the mean grayscale value of the pixels in the neighborhood of the pixel at the same position as the p-th pixel in the hyperspectral remote sensing image of the visible light band in the i-th historical time period before the t-th time period; is the grayscale mean of the pixels in the neighborhood of the pixel at the same position as the p-th pixel in the hyperspectral remote sensing image of the visible light band in the i+1-th historical time period before the t-th time period, and n is the number of preset historical time periods in the t-th time period.

[0009] In one embodiment, dividing the severe lesion area in each time period into a real severe lesion area and a soil area includes: Performing a linear fit on the average skeleton width of the skeleton graphs corresponding to the matched skeleton graphs in the hyperspectral remote sensing images of multiple historical time periods in any skeleton graph in the severe lesion area in each time period to obtain the slope of the fitting line, obtaining the pixel points in the severe lesion area in each time period with a change intensity greater than a preset threshold, recording them as mutation pixels, and merging adjacent mutation pixels to obtain the mutation area; The real serious lesion area and the soil area are divided according to the positive and negative relationship of the slope and the mutation area.

[0010] In one embodiment, the dividing of the real serious lesion area and the soil area includes: If the slope is negative and the corresponding area of ​​any skeleton image is connected to or intersects with the mutation area, the corresponding area of ​​any skeleton image is taken as the real severe lesion area, and the area in the severe lesion area that is not the real severe lesion area is taken as the soil area.

[0011] In one embodiment, determining grayscale thresholds capable of distinguishing the regions includes: Based on the separated real serious lesion area and soil area, a grayscale threshold for distinguishing the real serious lesion area from the soil area can be obtained through histogram analysis, which is recorded as T3, where T3>T2>T1.

[0012] In one embodiment, determining the allocation weights of the Sobel operator and the Krisch operator when calculating the gradient for each pixel point in each region of the hyperspectral remote sensing image by each grayscale threshold includes: Based on the grayscale thresholds in the hyperspectral remote sensing image of the visible light band of each time period, the optimized threshold of each area in the hyperspectral remote sensing image of the visible light band of each time period is determined; The gradient calculation expression of each pixel point in each area of ​​the hyperspectral remote sensing image is: Where, is the optimized gradient amplitude of each pixel, is the gradient amplitude calculated by Sobel operator at each pixel, is the gradient amplitude calculated by Krisch operator at each pixel, is the weight assigned to the gradient magnitude, is the gradient direction calculated by Sobel operator for each pixel, is the gradient direction calculated by Krisch operator for each pixel, Assign weights to the gradient directions; The distribution weight of the gradient amplitude and the distribution weight of the gradient direction of each pixel are determined according to the optimized threshold of each area in the hyperspectral remote sensing image.

[0013] In one embodiment, determining the optimized threshold value for each region in the hyperspectral remote sensing image of the visible light band in each time period includes: For the healthy leaf area, the optimized threshold is T1; for the slightly lesion area, the optimized threshold is the average of T1 and T2; for the truly severe lesion area, the optimized threshold is the average of T2 and T3; for the soil area, the optimized threshold is T3.

[0014] In one embodiment, determining the weight assigned to the gradient magnitude and the weight assigned to the gradient direction of each pixel includes: Calculate the ratio of a preset constant to the optimized threshold of each region in the hyperspectral remote sensing image, determine the minimum value between the ratio and the natural number 1, and the distribution weight of the gradient amplitude and the distribution weight of the gradient direction of each pixel point are both the minimum value corresponding to the region to which the pixel point belongs.

[0015] This application has the following beneficial effects: The present application divides the hyperspectral remote sensing image in the visible light band into healthy leaf areas, slightly lesion areas and seriously lesion areas based on the grayscale difference of each pixel in the hyperspectral remote sensing image in the visible light band; improves the precision of crop lesion identification, objectively reflects the diffusion trend of crop lesions, and strengthens the quantification and scientific nature of the lesion degree; performs skeleton extraction on the seriously lesion area in each time period, obtains each skeleton map of the seriously lesion area, and determines the average skeleton width of each skeleton map; avoids the interference of soil texture on lesion shape recognition, optimizes the pathological morphological representation, and enhances the accuracy of crop lesion area recognition; determines each grayscale threshold that can distinguish each area based on each area that has been divided in the hyperspectral remote sensing image; optimizes the proportion of Sobel operator and Krisch operator in the subsequent calculation process of each pixel gradient, and performs gradient calculation on pixel points in each area in the hyperspectral remote sensing image in a targeted manner; in the process of using the Canny edge detection algorithm to detect the hyperspectral remote sensing image in the visible light band When performing edge detection, the Sobel operator and the Krisch operator are fused in the process of calculating the gradient of each pixel point to obtain the optimized gradient value of each pixel point; the distribution weights of the Sobel operator and the Krisch operator for each pixel point in each area of ​​the hyperspectral remote sensing image when calculating the gradient are determined by each grayscale threshold, thereby improving the suitability of the distribution weight determination; by fusing the Sobel operator and the Krisch operator, the edges in the image can be detected more comprehensively, more edge features can be captured, and the accuracy and robustness of edge detection can be improved; the hyperspectral remote sensing images of each time period are combined with the results of the edge detection, and a neural network model is used to monitor the growth status of field crops; the present application can more clearly capture the edges, veins and different areas of the leaves by performing edge detection on the hyperspectral remote sensing images, providing more accurate spatial information, which helps the neural network to better focus on important areas in the subsequent chlorophyll content identification and avoid interference from irrelevant areas, thereby improving the accuracy of crop growth status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0017] Figure 1 A block diagram of a low-altitude remote sensing monitoring system for field crop phenotypes provided in one embodiment of the present application; Figure 2 Implementation block diagram for the lesion area enhancement module; Figure 3Implementation block diagram of the field crop monitoring module. DETAILED DESCRIPTION

[0018] To further illustrate the technical means and effectiveness of this application to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effectiveness of a low-altitude remote sensing monitoring system for field crop phenotypes proposed in this application. In the following description, references to different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0019] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0020] The specific scheme of the low-altitude remote sensing monitoring system for field crop phenotypes provided by this application is described in detail below with reference to the accompanying drawings.

[0021] See also Figure 1 , which shows a block diagram of a low-altitude remote sensing monitoring system for field crop phenotypes provided by an embodiment of the present application. The system includes: a remote sensing image acquisition module 101, a lesion area enhancement module 102, and a field crop monitoring module 103.

[0022] The remote sensing image acquisition module 101 acquires hyperspectral remote sensing images of field crops in the same area at different time periods.

[0023] Field crops refer to crops cultivated on large scales across large fields, primarily including grain crops (e.g., wheat, rice, corn), oilseed crops (e.g., rapeseed, peanuts), and fiber crops (e.g., cotton). This embodiment uses wheat crops as an example. An intelligent drone equipped with a hyperspectral imager conducts low-altitude scanning of a wheat field along a preset flight path, simultaneously capturing hundreds of continuous, narrow-band hyperspectral remote sensing images of wheat ranging from visible light to shortwave infrared. The hyperspectral remote sensing images of wheat collected in this embodiment are of a fixed, uniform area of ​​the wheat field, and are collected at fixed time intervals. In this embodiment, the fixed time interval is two days, meaning a hyperspectral remote sensing image of wheat is collected from the same area every two days. Implementers can set the length of the fixed time interval based on actual circumstances, and this embodiment does not impose any restrictions on this.

[0024] The hyperspectral remote sensing images of wheat collected in this example capture a large number of complex features, including healthy and diseased wheat and soil, over multiple time periods over the same area. During flight, the drone uses GNSS positioning and an inertial navigation system to ensure spatial consistency of the collected hyperspectral remote sensing images. All captured hyperspectral remote sensing images of wheat are combined into a raw dataset.

[0025] After obtaining the original dataset of the wheat field, first, the images in the original dataset are subjected to radiation correction, aiming to eliminate the radiation distortion in the image caused by factors such as lighting conditions. Radiation correction is an existing well-known technology, and its specific process will not be described in detail. Then, the radiation-corrected image is subjected to geometric correction, aiming to correct the image deformation and position offset caused by changes in the UAV's flight attitude, terrain undulations, and lens distortion, so that the image and geographic coordinates are accurately matched, and the spatial consistency of multi-phase images is guaranteed. Geometric correction is an existing well-known technology, and its specific process will not be described in detail. After radiation correction and geometric correction processing, the pre-processed wheat field hyperspectral remote sensing image dataset is obtained.

[0026] The lesion area enhancement module 102, (1) for each time period, divides the hyperspectral remote sensing image in the visible light band into healthy leaf areas, slightly lesion areas and seriously lesion areas based on the grayscale difference of each pixel point in the hyperspectral remote sensing image in the visible light band.

[0027] First, the preprocessed wheat field hyperspectral remote sensing image dataset is grayscaled and normalized. In this embodiment, the primary purpose is to monitor changes in crop color and chlorophyll levels. Therefore, the visible light band data from the preprocessed crop hyperspectral remote sensing image is grayscaled using the green channel inversion method. This allows darker green leaf features to exhibit lower grayscale values. The grayscaled image pixel values ​​are then normalized to [0, 1] using the minimum-maximum normalization method. Image grayscale and minimum-maximum normalization are well-known techniques, and the specific processes are not detailed here.

[0028] In dry winter climates, the soil surface lacks water and exhibits a light gray-brown color, with high reflectivity in the visible light band of hyperspectral images. Healthy wheat seedlings are rich in chlorophyll and have low reflectivity in the visible light band of hyperspectral images. When wheat seedlings develop disease, the chlorophyll in their leaves degrades, causing the leaves to appear yellow or even lighter in color, resulting in increased reflectivity in the visible light band of hyperspectral images.

[0029] Based on the above analysis, this embodiment performs thresholding on the hyperspectral remote sensing grayscale images for each time period. Based on the different grayscale values ​​in different regions, the hyperspectral remote sensing images are demarcated into healthy leaves, slightly diseased leaves, severely diseased leaves, and soil feature regions. The thresholding method in this embodiment uses multi-value thresholding, automatically deriving two grayscale thresholds, T1 and T2, using a histogram analysis method. This histogram analysis method is well known in the art, and the specific process is not described in detail here.

[0030] Specifically, for each time period, pixels with grayscale values ​​less than or equal to T1 in the hyperspectral remote sensing image of the visible light band are used as pixels of healthy leaves, pixels with grayscale values ​​greater than T1 and less than T2 are used as pixels of slightly diseased leaves, and pixels with grayscale values ​​greater than or equal to T2 are used as pixels of severely diseased leaves. Connected domain extraction is performed on all pixels of healthy leaves in the hyperspectral remote sensing image to obtain healthy leaf areas; connected domain extraction is performed on all pixels of slightly diseased leaves in the hyperspectral remote sensing image to obtain slightly diseased areas; and connected domain extraction is performed on all pixels of severely diseased areas in the hyperspectral remote sensing image to obtain severely diseased areas. Connected domain extraction is a well-known technique, and the specific process will not be repeated here.

[0031] (2) Determine the intensity of the change in the grayscale of each pixel in the neighborhood of each pixel in the severe lesion area in each time period in multiple historical time periods; perform skeleton extraction on the severe lesion area in each time period, obtain each skeleton map of the severe lesion area, and determine the average skeleton width of each skeleton map.

[0032] After thresholding the visible-band hyperspectral remote sensing image, severe lesions in areas where leaves are nearly wilted appear similar in grayscale to the soil. This makes it impossible to accurately distinguish between the actual severe lesions and soil using thresholding alone. Therefore, this embodiment combines time-series crop images and distinguishes between the actual severe lesions and soil based on changes in crop growth characteristics.

[0033] For the hyperspectral remote sensing image acquired in the tth time period, other hyperspectral remote sensing images in the adjacent time periods can reflect the state changes of the wheat seedlings in the area. For a slightly lesioned area, the lesioned area will experience two states in the time period after the tth time period: recovery or worsening of the lesion. Therefore, it can be concluded that the grayscale value of the same pixel will change in different time periods. In addition, the impact of the lesion can also cause the leaf to shrink. When the leaf is severely lesioned, it will wither, resulting in a decrease in the width of the leaf skeleton feature.

[0034] If the pixel If a more severe lesion occurs in the subsequent time period, then the pixel The grayscale of the pixels in the neighborhood of the corresponding pixel in the subsequent image will increase. Based on this, this embodiment calculates the change intensity of each pixel in the severe lesion area in the hyperspectral remote sensing image of the visible light band in each time period. The specific expression is: Where, is the change intensity of the p-th pixel in the severe lesion area in the hyperspectral remote sensing image of the visible light band in the t-th time period, is the mean grayscale value of the pixels in the neighborhood of the pixel at the same position as the p-th pixel in the hyperspectral remote sensing image of the visible light band in the i-th historical time period before the t-th time period; is the mean grayscale value of the pixels in the neighborhood of the pixel at the same position as the p-th pixel in the hyperspectral remote sensing image of the visible light band in the i+1th historical time period before the t-th time period, and n is the number of preset historical time periods in the t-th time period. In this embodiment, n=5, and implementers can set it according to actual conditions.

[0035] It should be noted that the neighborhood in this embodiment is the eight neighborhoods of the pixel point. The implementer can determine the neighborhood range of the pixel point according to actual conditions, and this embodiment does not impose any limitation on this.

[0036] At the same time, the actual severely diseased areas of leaves will shrink and narrow over time, while the soil area does not exhibit this characteristic. Therefore, this embodiment uses the K3M skeleton extraction algorithm to extract skeleton images of severely diseased areas in each time period. These skeleton images include skeleton images of severely diseased leaves and cracks in the soil. The skeleton images are then transformed using the Euclidean distance transform to calculate the average skeleton width of each skeleton image. The K3M skeleton extraction algorithm and Euclidean distance transform are well-known technologies, and the specific process is not repeated here.

[0037] (3) Analyze the change trend of the average skeleton width of the skeleton graphs corresponding to the matched skeleton graphs in the hyperspectral remote sensing images of multiple historical time periods in the severe lesion areas of each time period, and combine the change intensity of the pixel points in the adjacent areas of each skeleton graph to divide the severe lesion areas of each time period into real severe lesion areas and soil areas.

[0038] For each time period, if the mth skeleton in the severe lesion area in the hyperspectral remote sensing image is a real wheat seedling skeleton with severe lesions, then as the wheat seedling lesions further deteriorate and shrink over time, its skeleton width will become narrower, causing the measurement of the skeleton width to gradually decrease. Therefore, this embodiment performs a linear fit on the average skeleton width of the corresponding matching skeleton graphs in the hyperspectral remote sensing images of the historical n time periods for any skeleton graph in the severe lesion area of ​​each time period to obtain the slope of the fitting line. In addition, the pixel points with a change intensity greater than a preset threshold in the severe lesion area of ​​each time period are obtained, recorded as mutation pixels, and the adjacent mutation pixels are merged to obtain the mutation area. The fitting line is obtained by linear fitting using the least squares method, and the implementer can choose other existing feasible linear fitting algorithms.

[0039] If the slope is negative and the corresponding area of ​​any skeleton image is connected to or intersects with the mutation area, the corresponding area of ​​any skeleton image is regarded as the real severe lesion area, and the area in the severe lesion area that is not the real severe lesion area is regarded as the soil area. Figure 2 shown.

[0040] It should be noted that the preset threshold is set according to the characteristics of the crop, representing the change in chlorophyll level of the crop leaves from a healthy state to a diseased state in the pixel grayscale value. In this embodiment, the preset threshold is set to 0.1, and the implementer can set it according to the actual situation. This embodiment does not impose any restrictions on this.

[0041] The field crop monitoring module 103, (1) based on the regions that have been divided in the hyperspectral remote sensing image, determines the grayscale thresholds that can distinguish the regions; when using the Canny edge detection algorithm to perform edge detection on the hyperspectral remote sensing image in the visible light band, the Sobel operator and the Krisch operator are integrated in the process of calculating the gradient of each pixel to obtain the optimized gradient value of each pixel; and the distribution weights of the Sobel operator and the Krisch operator for each pixel in each region of the hyperspectral remote sensing image when calculating the gradient are determined by using the grayscale thresholds.

[0042] Furthermore, based on the separated real serious lesion area and soil area, a grayscale threshold for distinguishing the real serious lesion area from the soil area can be obtained through histogram analysis, which is recorded as T3, where T3>T2>T1.

[0043] In the visible light band, the reflectance of yellowed leaves approaches that of soil. Furthermore, the edges of diseased wheat seedling leaves exhibit irregular features, potentially appearing tilted or jagged. This results in spatial features with weak gradient amplitudes and complex edge directions. The traditional gradient calculation method used in the Canny edge detection algorithm relies on a single directional gradient, which can easily miss weak edges. This makes it difficult to distinguish boundaries in the visible light band and prevents accurate differentiation of yellowed wheat leaves from the soil background.

[0044] Based on this, this embodiment corrects the gradient amplitude and direction calculation of the Canny edge detection algorithm.

[0045] For the wheat field hyperspectral remote sensing image dataset Any pixel in an image In the gradient and direction calculation of the Canny algorithm, the Sobel operator is used to calculate the gradient strength and gradient direction The calculation of the Sobel operator gradient is an existing technology, and its detailed process will not be repeated here.

[0046] The Krisch operator uses an 8-directional convolution kernel (the convolution kernel size is 3×3) to perform multi-directional gradient calculations, and the gradient strength of the Krisch operator can be obtained. and gradient direction , which can detect edges in all directions in the image. Krisch operator gradient calculation is an existing technology, and its detailed process will not be repeated here.

[0047] Therefore, this embodiment integrates the Sobel operator and the Krisch operator to correct the gradient amplitude and direction calculation of the Canny edge detection algorithm. The gradient calculation expression of each pixel point in each area of ​​the hyperspectral remote sensing image is: Where, is the optimized gradient amplitude of each pixel, is the gradient amplitude calculated by Sobel operator at each pixel, is the gradient amplitude calculated by Krisch operator at each pixel, is the weight assigned to the gradient magnitude, is the gradient direction calculated by Sobel operator for each pixel, is the gradient direction calculated by Krisch operator for each pixel, is the weight assigned to the gradient direction, It is a hyperparameter that adjusts the proportion of the gradient strength and direction of the pixel points calculated by the Krisch operator and the Sobel operator in the Canny algorithm.

[0048] Furthermore, hyperparameters These values ​​are not fixed; they adapt adaptively when calculating gradient strength and direction based on the areas demarcated by different grayscale thresholds. In hyperspectral remote sensing images, if wheat seedlings are severely diseased, the tips of their leaves will become lighter and the shape of their edges will change. That is, the grayscale value of pixels in healthy leaf areas is less than the grayscale value of pixels in areas with mild disease, and less than the grayscale value of pixels in areas with actual severe disease. The grayscale value of pixels in areas with actual severe disease is even close to the grayscale intensity of dry soil.

[0049] When calculating the gradient intensity and gradient direction of the pixel points in different regions of the hyperspectral remote sensing image, this embodiment uses the optimized threshold of this region To adjust , optimize the threshold The value of is: For the healthy leaf area, the optimized threshold is , for the mild lesion area, the optimized threshold is the mean of T1 and T2. For the real serious lesion area, the optimized threshold is the mean of T2 and T3. For the soil area, the optimized threshold is .

[0050] but and The value of is: Wherein, C is a preset constant. It is set according to the characteristics of crops. Different crops have different chlorophyll contents, so the leaf colors are not exactly the same. The size of C is set according to the grayscale mean value of the grayscale image of the normal leaves of each crop. In this embodiment, wheat images are used and C is set to 0.3. is the optimized threshold value of each region in the hyperspectral remote sensing image of each time period, and min() is the minimum value function.

[0051] It should be understood that The smaller, The larger the value, the more likely it is to suppress soil texture noise in the healthy leaf area using the Sobel operator as the dominant information. The bigger, The smaller it is, the purpose is to use the Krisch operator as the dominant information to enhance the detection of multi-directional weak edges in the diseased leaf area.

[0052] The smaller, The larger the value, the more likely it is that the edge of the healthy leaf area will be detected based on the gradient direction of the Sobel operator. The bigger, The smaller it is, the purpose is to focus on the gradient direction of the Krisch operator and improve the accuracy of edge detection for complex scenes such as diseased leaf areas and soil.

[0053] Finally, for each time period, the Canny edge detection algorithm after the modified gradient calculation is used to perform edge detection on the hyperspectral remote sensing image in the visible light band to obtain the edge detection results. The Canny edge detection algorithm is a well-known technology, and the specific process is not described in detail here.

[0054] (2) Combining the hyperspectral remote sensing images of each time period with the results of edge detection, a neural network model is used to monitor the growth status of field crops.

[0055] The edge detection results of the hyperspectral remote sensing images of the visible light bands in each time period are mapped to the hyperspectral remote sensing images of each band, and the hyperspectral remote sensing images with edge marks in each band of each time period are obtained. The edge marking processing is performed on the hyperspectral remote sensing images of all time periods in the wheat field hyperspectral remote sensing image dataset, and the edge-marked hyperspectral remote sensing image dataset of wheat fields is obtained, which realizes the edge enhancement of the hyperspectral remote sensing images.

[0056] The edge-marked wheat field hyperspectral remote sensing image dataset is input into the neural network model for training, and the chlorophyll content is used as the output parameter of the neural network model. By monitoring the chlorophyll content of crops, the growth status of crops can be monitored, and the accuracy of chlorophyll content monitoring of crops can be improved. The block diagram of the field crop monitoring module is shown in the figure. Figure 3 shown.

[0057] In this embodiment, the neural network model adopts the ResNet18 network, the model optimizer adopts the Adam optimizer, and the loss function adopts the mean square error loss function. The implementer can choose other existing feasible neural network models. Monitoring the chlorophyll content of crops through the neural network model is an existing well-known technology, and the specific process will not be described in detail.

[0058] It should be noted that the order of the embodiments of the present application is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0059] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0060] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A low-altitude remote sensing monitoring system for field crop phenotypes, characterized in that: The system comprises: Remote sensing image acquisition module, which collects hyperspectral remote sensing images of field crops in the same area at different time periods; The lesion area enhancement module divides the hyperspectral remote sensing image in the visible light band into healthy leaf areas, slightly lesion areas, and severely lesion areas based on the grayscale difference of each pixel in the hyperspectral remote sensing image in the visible light band for each time period; Determining the intensity of change in the grayscale of pixels in the neighborhood of each pixel in the severe lesion area in each time period over multiple historical time periods; performing skeleton extraction on the severe lesion area in each time period to obtain skeleton images of the severe lesion area, and determining the average skeleton width of each skeleton image; Analyze the change trend of the average skeleton width of each skeleton graph in the severe lesion area in each time period and the corresponding matched skeleton graph in the hyperspectral remote sensing images of multiple historical time periods, and combine the change intensity of the pixel points in the adjacent areas of each skeleton graph to divide the severe lesion area in each time period into a real severe lesion area and a soil area; The field crop monitoring module determines the grayscale thresholds that can distinguish each region based on the regions that have been divided in the hyperspectral remote sensing image. When using the Canny edge detection algorithm to perform edge detection on the hyperspectral remote sensing image in the visible light band, the Sobel operator and the Krisch operator are integrated in the process of calculating the gradient of each pixel to obtain the optimized gradient value of each pixel. The grayscale thresholds are used to determine the distribution weight of the Sobel operator and the Krisch operator when calculating the gradient of each pixel in each region of the hyperspectral remote sensing image. The hyperspectral remote sensing images of each time period are combined with the edge detection results and a neural network model is used to monitor the growth status of field crops.

2. A low-altitude remote sensing monitoring system for field crop phenotypes according to claim 1, characterized in that: The hyperspectral remote sensing image in the visible light band is divided into healthy leaf areas, slightly lesion areas and seriously lesion areas, including: The histogram analysis of the hyperspectral remote sensing image in the visible light band is performed, and the multi-threshold processing method is used to obtain two grayscale thresholds, which are recorded as T1 and T2, respectively, where T1 < T2; Based on the thresholds T1 and T2, the pixels in the hyperspectral remote sensing image in the visible light band are divided into healthy leaf areas, slightly lesion areas, and severely lesion areas.

3. A low-altitude remote sensing monitoring system for field crop phenotypes according to claim 2, characterized in that: The pixel area with grayscale value less than or equal to T1 in the hyperspectral remote sensing image of the visible light band is regarded as the healthy leaf area, the pixel area with grayscale value greater than T1 and less than T2 is regarded as the mild lesion area, and the pixel area with grayscale value greater than or equal to T2 is regarded as the severe lesion area.

4. The low-altitude remote sensing monitoring system for field crop phenotypes according to claim 1, characterized in that: The expression of the change intensity is: Where, is the change intensity of the p-th pixel in the severe lesion area in the hyperspectral remote sensing image of the visible light band in the t-th time period, is the mean grayscale value of the pixels in the neighborhood of the pixel at the same position as the p-th pixel in the hyperspectral remote sensing image of the visible light band in the i-th historical time period before the t-th time period; is the grayscale mean of the pixels in the neighborhood of the pixel at the same position as the p-th pixel in the hyperspectral remote sensing image of the visible light band in the i+1-th historical time period before the t-th time period, and n is the number of preset historical time periods in the t-th time period.

5. The low-altitude remote sensing monitoring system for field crop phenotypes according to claim 1, characterized in that: The dividing of the severe lesion area in each time period into a real severe lesion area and a soil area includes: Performing a linear fit on the average skeleton width of the skeleton graphs corresponding to the matched skeleton graphs in the hyperspectral remote sensing images of multiple historical time periods in any skeleton graph in the severe lesion area in each time period to obtain the slope of the fitting line, obtaining the pixel points in the severe lesion area in each time period with a change intensity greater than a preset threshold, recording them as mutation pixels, and merging adjacent mutation pixels to obtain the mutation area; The real serious lesion area and the soil area are divided according to the positive and negative relationship of the slope and the mutation area.

6. A low-altitude remote sensing monitoring system for field crop phenotypes according to claim 5, characterized in that: The division of the real serious lesion area and the soil area includes: If the slope is negative and the corresponding area of ​​any skeleton image is connected to or intersects with the mutation area, the corresponding area of ​​any skeleton image is taken as the real severe lesion area, and the area in the severe lesion area that is not the real severe lesion area is taken as the soil area.

7. The low-altitude remote sensing monitoring system for field crop phenotypes according to claim 2, characterized in that: The step of determining grayscale thresholds capable of distinguishing the regions includes: Based on the separated real serious lesion area and soil area, a grayscale threshold for distinguishing the real serious lesion area from the soil area can be obtained through histogram analysis, which is recorded as T3, where T3>T2>T1.

8. A low-altitude remote sensing monitoring system for field crop phenotypes according to claim 7, characterized in that: The method of determining the distribution weights of the Sobel operator and the Krisch operator for each pixel point in each region of the hyperspectral remote sensing image when calculating the gradient by each grayscale threshold includes: Based on the grayscale thresholds in the hyperspectral remote sensing image of the visible light band of each time period, the optimized threshold of each area in the hyperspectral remote sensing image of the visible light band of each time period is determined; The gradient calculation expression of each pixel point in each area of ​​the hyperspectral remote sensing image is: Where, is the optimized gradient amplitude of each pixel, is the gradient amplitude calculated by Sobel operator at each pixel, is the gradient amplitude calculated by Krisch operator at each pixel, is the weight assigned to the gradient magnitude, is the gradient direction calculated by Sobel operator for each pixel, is the gradient direction calculated by Krisch operator for each pixel, Assign weights to the gradient directions; The distribution weight of the gradient amplitude and the distribution weight of the gradient direction of each pixel are determined according to the optimized threshold of each area in the hyperspectral remote sensing image.

9. A low-altitude remote sensing monitoring system for field crop phenotypes according to claim 8, characterized in that: The step of determining the optimized threshold value for each region in the hyperspectral remote sensing image of the visible light band in each time period includes: For the healthy leaf area, the optimized threshold is T1; for the slightly lesion area, the optimized threshold is the average of T1 and T2; for the truly severe lesion area, the optimized threshold is the average of T2 and T3; for the soil area, the optimized threshold is T3.

10. The low-altitude remote sensing monitoring system for field crop phenotypes according to claim 8, characterized in that: Determining the distribution weight of the gradient magnitude and the distribution weight of the gradient direction of each pixel includes: Calculate the ratio of a preset constant to the optimized threshold of each region in the hyperspectral remote sensing image, determine the minimum value between the ratio and the natural number 1, and the distribution weight of the gradient amplitude and the distribution weight of the gradient direction of each pixel point are both the minimum value corresponding to the region to which the pixel point belongs.

Images