An unmanned aerial vehicle-based farmland dynamic monitoring method and system

Abstract

The application discloses a kind of farmland dynamic monitoring method and system based on unmanned aerial vehicle, it is related to image processing technical field;Crop image is carried out artifact denoising processing and crop overexposure repair processing to obtain original crop image;Original crop image is substituted into preset weed crop segmentation model to obtain target crop image;Target crop image is substituted into focus recognition model to obtain focus type and focus position;Based on preset monitoring period, crop image is collected again to analyze, the focus data obtained by twice monitoring is associated and matched, and the growth state of crop is determined;After artifact denoising and overexposure repair of unmanned aerial vehicle collected image, the preset weed crop segmentation model can accurately distinguish crops and weeds, improve the purity of target crop extraction, and then the focus recognition model is used to accurately determine the type and position of the focus.Combined with the associated matching of multi-cycle monitoring data, the growth state of crops can be dynamically and comprehensively mastered, and the accuracy and efficiency of farmland monitoring are improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method and system for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs). Background Technology

[0002] With the rapid development of precision agriculture technology, real-time and dynamic monitoring of crop growth status in farmland has become a core requirement for improving agricultural production efficiency and reducing disaster losses. With its advantages of mobility, wide coverage and high data acquisition efficiency, drones have gradually replaced traditional manual inspections and become the mainstream technology carrier in the field of farmland monitoring. By carrying high-definition camera equipment to collect farmland images, they provide data support for crop growth status assessment.

[0003] However, in practical applications, the complex environment and equipment characteristics encountered when drones are used to photograph farmland from high altitudes result in widespread quality defects in the acquired crop images. This severely restricts the accuracy of subsequent monitoring and analysis. In farmland scenes, light intensity is easily affected by weather, time of day, and other factors, causing sudden changes. In addition, noise interference may be generated by the drone's onboard sensors during high-altitude operations, making crop areas prone to overexposure distortion. A large amount of key details such as the texture and outline of crops in highlighted areas are lost. Such image quality defects directly trigger a chain of problems in subsequent core technology aspects. In the crop and weed segmentation process, the feature boundaries between crops and weeds are blurred in overexposed areas, making it difficult for ordinary segmentation models to accurately distinguish between the two. This results in a large number of misclassified and missed areas in the segmentation results, leading to insufficient purity of the extracted target crop. In the crop lesion identification process, feature loss and noise interference caused by overexposure directly lead to the omission of small-scale lesions, while reducing the identifiability of lesion features. Furthermore, existing lesion identification models themselves also have technical bottlenecks, with insufficient multi-scale feature processing capabilities for crop lesions. They are unable to effectively integrate high-level semantic features and low-level spatial detail features, resulting in the loss of lesion edge features during feature fusion. Ultimately, this leads to a dilemma of low accuracy in lesion identification and low positioning accuracy.

[0004] Publication No. CN117649618A discloses a method, system, and electronic device for image target detection based on unmanned aerial vehicles (UAVs). This method acquires panoramic images of farmland captured by a high-resolution camera and infrared videos of farmland captured by an infrared camera. The panoramic images are processed through multi-layer convolution to obtain panoramic farmland channel feature vectors. At the same time, keyframes are extracted from the infrared videos and infrared image feature vectors are obtained through a spatial attention feature extractor. Finally, the two types of feature vectors are fused to achieve early warning of crop pests, thereby improving the monitoring effect of farmland pests.

[0005] However, existing technical solutions still have shortcomings. In the crop image acquisition stage, overexposure can lead to the loss of key details such as texture and contour in the highlight areas. This not only blurs the feature boundaries between crops and weeds, resulting in insufficient purity of the segmentation results, but also directly leads to the loss of feature information of small-scale lesions. At the same time, the existing models have weak multi-scale feature processing capabilities and cannot effectively fuse high-level semantic features used to determine the type of lesion with low-level spatial detail features used to locate the lesion. The loss of low-level spatial features caused by overexposure will further aggravate the information imbalance in the model feature fusion stage, resulting in the complete loss of lesion edge features. Ultimately, this leads to a significant decrease in the efficiency of crop lesion identification and low overall efficiency of farmland dynamic monitoring. Summary of the Invention

[0006] The purpose of this invention is to solve the problem that overexposure causes the loss of underlying spatial features, exacerbates the information imbalance in the model feature fusion stage, and ultimately leads to a significant decrease in the efficiency of crop lesion identification, resulting in low overall efficiency of farmland dynamic monitoring. Therefore, this invention proposes a farmland dynamic monitoring method and system based on unmanned aerial vehicles (UAVs).

[0007] In a first aspect of this invention, a method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs) is first proposed, the method comprising: Obtain crop images collected by drones, and perform artifact denoising processing on the crop images to obtain denoised crop images; The denoised crop image is subjected to crop overexposure repair processing to obtain the original crop image; Substitute the original crop map into the preset weed crop segmentation model to obtain the target crop map; Substitute the target crop image into the lesion identification model to obtain the lesion type and lesion location; Record the lesion type and location corresponding to all crop images, collect crop images again based on the preset monitoring cycle, and determine the lesion type and location corresponding to all crop images again; The lesion data obtained from the two monitoring sessions were correlated and matched to determine the crop growth status.

[0008] Based on the UAV-based dynamic monitoring method for farmland provided in this solution, the collected images are first subjected to artifact denoising and overexposure repair to restore key crop details. Then, based on the weed and crop segmentation and lesion identification model, the crops are accurately distinguished and the type and location of lesions are determined. Finally, by combining multi-period monitoring data matching, the crop growth status is dynamically grasped, which significantly improves the accuracy and efficiency of farmland monitoring.

[0009] Optionally, performing crop overexposure repair processing on the denoised crop image to obtain the original crop image includes: Calculate the brightness histogram of the denoised crop image, and determine the overexposure threshold based on the brightness histogram; Candidate overexposure points in the denoised crop image are determined based on the overexposure threshold and the differences in the RGB three channels; Connectivity analysis is performed on candidate overexposed points, and connected regions with an area greater than a preset number of pixels are retained. Holes in the regions are filled by morphological opening and closing operations, and edge burrs are removed to obtain an exposure mask. The denoised crop image is converted to grayscale to obtain a grayscale image, and the low-frequency and high-frequency components of the grayscale image are separated. The target exposure mask is obtained by performing crop exposure processing on the exposure mask based on the low-frequency components and the high-frequency components. The original crop image is obtained by correcting the denoised crop image based on the target exposure mask image.

[0010] This solution improves the clarity and realism of crop images by accurately locating and repairing overexposed areas while preserving low-frequency details and high-frequency textures, providing a high-quality image data foundation for subsequent visual tasks such as crop growth analysis and pest and disease identification.

[0011] Optionally, performing crop exposure processing on the exposure mask based on the low-frequency components and the high-frequency components to obtain the target exposure mask includes: By pre-setting the normal brightness range of crops, the low-frequency components corresponding to the exposure mask are brightness mapped so that their mean and variance are consistent with the normal crop brightness distribution, thus obtaining the low-frequency correction distribution. Calculate the channel imbalance and saturation deviation of the corresponding pixels in the exposure mask, and determine the heavily distorted area and the lightly distorted area in the exposure mask based on the channel imbalance and the saturation deviation. The low-frequency correction distribution is mapped onto the severe distortion region to obtain the corrected severe distortion region; For the slightly distorted area, the normal crop pixels at the edge of the distorted area are used as the reference. The RGB values ​​of the edge pixels are filled by bilinear interpolation, and then saturation compensation is performed to correct the slightly distorted area. The exposure mask is corrected based on the areas of severe distortion and mild distortion to obtain a corrected exposure mask. The high-frequency components are added to the corrected exposure mask using a feature fusion algorithm to obtain the target exposure mask.

[0012] The low-frequency components of the exposure mask are normalized and mapped to match the brightness distribution characteristics of normal crops. Simultaneously, based on channel imbalance and saturation deviation, severely and mildly distorted areas are accurately divided. Severely distorted areas are directly replaced with low-frequency corrected distributions, while mildly distorted areas undergo refined repair using bilinear interpolation of edge pixels combined with saturation compensation. Finally, high-frequency components are fused to restore texture details. This approach achieves accurate brightness correction in overexposed areas while preserving the texture information of the crop image to the greatest extent possible, effectively improving the realism and integrity of the repaired image and providing a high-quality image data foundation for subsequent crop lesion identification tasks.

[0013] Optionally, substituting the target crop image into the lesion identification model to obtain the lesion type and lesion location includes: The target crop map is substituted into the YOLOv10 backbone network to obtain high-scale, mid-scale, and low-scale maps; the high-scale map is the output of the fifth-layer C2f module; the mid-scale map is the output of the seventh-layer C2f module; and the low-scale map is the output of the eleventh-layer PSA module. A 1×1 convolution operation is performed on the high-scale image to obtain high-scale features; a 1×1 convolution operation is performed on the low-scale image to obtain low-scale features; Substituting the mesoscale map into a preset enhancement fusion model yields mesoscale enhanced features; After pixel resampling of the low-scale image, it is concatenated with the mid-scale enhancement feature and then substituted into the C3k2 module to obtain the first fused feature; The first fused feature is upsampled by pixel recombination and then concatenated with the high-scale feature before being substituted into the C3k2 module to obtain the first output feature; After performing a 1×1 convolution operation on the first output feature and downsampling it, it is fused with the first fused feature and the mesoscale enhancement feature to obtain the second fused feature; Substitute the second fusion feature into the C3k2 module to obtain the second output feature; The second output feature is downsampled and fused with the low-scale feature before being substituted into the C3k2 module to obtain the third output feature; The first, second, and third output features are input into the head network of YOLOv10 to obtain the lesion type and lesion location.

[0014] This solution is based on the YOLOv10 architecture and selectively uses feature maps from different layers of the backbone network as high-, medium-, and low-scale inputs. Through differentiated convolution, enhanced fusion, and multi-scale stitching operations, combined with the C3k2 module, the features are deeply refined. At the same time, a bidirectional fusion strategy of upsampling and downsampling is adopted to achieve accurate complementarity and enhancement of high-scale detailed features, medium-scale semantic features, and low-scale global features. This not only solves the problems of missed lesion detection and inaccurate localization in traditional single-scale feature detection, but also improves the model's recognition accuracy and localization efficiency for crop lesions of different sizes and shapes through feature reuse and targeted enhancement.

[0015] Optionally, substituting the mesoscale map into a preset enhancement fusion model to obtain mesoscale enhanced features includes: The mesoscale image is segmented into channels to obtain a first channel image and a second channel image; The first channel map is sequentially substituted into the convolutional layer and the multilayer perceptron layer, and then random path discarding is performed to obtain the first channel augmentation map; The first channel enhancement map and the second channel map are stitched together to obtain an initial mesoscale enhancement map; The initial mesoscale augmented image is subjected to a 3×3 convolution operation followed by global pooling to obtain global pooled features; The initial mesoscale augmented image is subjected to Y-axis pooling and X-axis pooling respectively to obtain Y-axis pooling features and X-axis pooling features; Perform 1×1 convolution operations on the global pooling feature, the Y-direction pooling feature, and the X-direction pooling feature respectively, and then input the Sigmoid activation function to obtain the global pooling weight, the Y-direction weight, and the X-direction weight; The Y-direction weights, X-direction weights, and global pooling weights are applied to the initial mesoscale augmentation map to obtain mesoscale augmentation features.

[0016] This scheme achieves precise weighted enhancement of mesoscale image features through channel segmentation and differential enhancement, random path dropping to suppress overfitting, and 3×3 convolution to extract local features. Then, through multi-dimensional pooling and weight allocation in the global, X and Y directions, it preserves the local texture details and global spatial distribution features of the image, while adaptively highlighting key semantic information, effectively improving the representational ability of mesoscale enhanced features.

[0017] In a second aspect of this invention, a farmland dynamic monitoring system based on unmanned aerial vehicles (UAVs) is proposed, comprising: The noise reduction module is used to acquire crop images collected by the drone and perform artifact noise reduction processing on the crop images to obtain denoised crop images. The overexposure repair module is used to perform crop overexposure repair processing on the denoised crop image to obtain the original crop image. The target crop map generation module is used to substitute the original crop map into a preset weed crop segmentation model to obtain the target crop map. The lesion identification module is used to input the target crop image into the lesion identification model to obtain the lesion type and lesion location; The lesion re-identification module is used to record the lesion type and lesion location corresponding to all crop images, collect crop images again based on a preset monitoring cycle, and re-determine the lesion type and lesion location corresponding to all crop images. The crop growth status determination module is used to correlate and match lesion data obtained from two monitoring sessions to determine the crop growth status.

[0018] Optionally, the overexposure repair module includes: The overexposure threshold calculation module is used to calculate the brightness histogram of the denoised crop image and determine the overexposure threshold based on the brightness histogram. The candidate overexposure point determination module is used to determine the candidate overexposure points in the denoised crop image based on the overexposure threshold and the difference between the RGB three channels. The exposure mask generation module is used to perform connected region analysis on candidate overexposed points, retain connected regions with an area greater than a preset number of pixels, fill holes in the region through morphological opening and closing operations, and remove edge burrs to obtain the exposure mask. The low- and high-frequency component determination module is used to perform grayscale conversion on the denoised crop image to obtain a grayscale image, and to separate the low-frequency and high-frequency components of the grayscale image; The target exposure mask generation module is used to perform crop exposure processing on the exposure mask based on the low-frequency components and the high-frequency components to obtain the target exposure mask; The original crop image generation module is used to correct the denoised crop image based on the target exposure mask image to obtain the original crop image.

[0019] Optionally, the target exposure mask generation module includes: The low-frequency correction distribution determination module is used to perform brightness mapping on the low-frequency components corresponding to the exposure mask image by preset normal crop brightness range, so that its mean and variance are consistent with the normal crop brightness distribution to obtain the low-frequency correction distribution; The distortion area determination module is used to calculate the channel imbalance and saturation deviation of the corresponding pixels in the exposure mask, and determine the heavily distorted area and the lightly distorted area in the exposure mask based on the channel imbalance and the saturation deviation. The severe distortion region determination module is used to map the low-frequency correction distribution to the severe distortion region to obtain the severe distortion region. The module for determining the slightly distorted area is used to fill the RGB values ​​of the edge pixels of the distorted area with normal crop pixels at the edge of the distorted area as a reference, and then perform saturation compensation to obtain the corrected slightly distorted area. The corrected exposure mask generation module is used to correct the exposure mask based on the corrected severe distortion area and the corrected mild distortion area to obtain a corrected exposure mask. The target exposure mask determination module is used to add the high-frequency components to the corrected exposure mask using a feature fusion algorithm to obtain the target exposure mask.

[0020] Optionally, the lesion identification module includes: A multi-scale extraction module is used to input the target crop map into the backbone network of YOLOv10 to obtain high-scale, medium-scale, and low-scale maps; the high-scale map is the output of the fifth-layer C2f module; the medium-scale map is the output of the seventh-layer C2f module; and the low-scale map is the output of the eleventh-layer PSA module. The low-scale feature determination module is used to perform a 1×1 convolution operation on the high-scale image to obtain high-scale features; and to perform a 1×1 convolution operation on the low-scale image to obtain low-scale features. The mesoscale enhancement feature determination module is used to substitute the mesoscale map into a preset enhancement fusion model to obtain mesoscale enhancement features; The first fusion feature generation module is used to perform pixel re-sampling on the low-scale image, then concatenate it with the mid-scale enhancement feature, and then substitute it into the C3k2 module to obtain the first fusion feature; The first output feature generation module is used to perform pixel recombination and upsampling on the first fused feature, then concatenate it with the high-scale feature and then substitute it into the C3k2 module to obtain the first output feature; The second fusion feature generation module is used to perform a 1×1 convolution operation on the first output feature, downsample it, and then fuse it with the first fusion feature and the mesoscale enhancement feature to obtain the second fusion feature. The second output feature generation module is used to substitute the second fused feature into the C3k2 module to obtain the second output feature; The third output feature generation module is used to downsample the second output feature and fuse it with the low-scale feature before substituting it into the C3k2 module to obtain the third output feature; The lesion identification module is used to input the first output feature, the second output feature, and the third output feature into the head network of YOLOv10 to obtain the lesion type and lesion location.

[0021] Optionally, the mesoscale enhancement feature determination module includes: The channel segmentation module is used to perform channel segmentation on the mesoscale image to obtain a first channel image and a second channel image; The first channel enhancement map determination module is used to sequentially substitute the first channel map into the convolutional layer and the multilayer perceptron layer and then randomly discard the path to obtain the first channel enhancement map. An initial mesoscale enhancement map generation module is used to stitch the first channel enhancement map and the second channel map to obtain an initial mesoscale enhancement map; The global pooling feature generation module is used to perform a 3×3 convolution operation on the initial mesoscale augmented map and then perform global pooling processing to obtain global pooling features. The XY-direction pooling module is used to perform Y-direction pooling and X-direction pooling on the initial mesoscale augmented image to obtain Y-direction pooling features and X-direction pooling features, respectively. The XY weight determination module is used to perform 1×1 convolution operations on the global pooling feature, the Y-direction pooling feature and the X-direction pooling feature respectively, and then input the Sigmoid activation function to obtain the global pooling weight, the Y-direction weight and the X-direction weight; The mesoscale enhancement feature generation module is used to apply the Y-direction weights, the X-direction weights, and the global pooling weights to the initial mesoscale enhancement map to obtain mesoscale enhancement features.

[0022] The beneficial effects of this invention are: This invention proposes a method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs). First, artifact denoising and overexposure repair are performed on images acquired by the UAV, effectively restoring key details such as crop texture and contours in highlight areas, laying a high-quality data foundation for subsequent processing. Based on the repaired images, a pre-set weed-crop segmentation model can accurately distinguish between crops and weeds, improving the purity of target crop extraction. Then, a lesion identification model is used to accurately determine the type and location of lesions. Finally, by combining the correlation and matching of multi-period monitoring data, the growth status of crops can be dynamically and comprehensively grasped, significantly improving the accuracy and efficiency of farmland monitoring. This solves the monitoring dilemma caused by image quality defects and insufficient model processing capabilities in traditional solutions. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 A flowchart illustrating a method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for repairing crop overexposure provided in an embodiment of the present invention; Figure 3 A schematic diagram of a lesion identification model provided in an embodiment of the present invention; Figure 4 This is a framework diagram of a farmland dynamic monitoring system based on unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0026] 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.

[0027] This invention provides a method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs). See also... Figure 1 , Figure 1 A flowchart illustrating a method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention. The method includes the following steps: S101: Obtain crop images collected by the drone, and perform artifact denoising processing on the crop images to obtain denoised crop images. S102, Perform crop overexposure repair processing on the denoised crop image to obtain the original crop image; S103, Substitute the original crop map into the preset weed crop segmentation model to obtain the target crop map; S104, Substitute the target crop image into the lesion identification model to obtain the lesion type and lesion location; S105, record the lesion type and lesion location corresponding to all crop images, collect crop images again based on the preset monitoring cycle, and determine the lesion type and lesion location corresponding to all crop images again; S106, correlate and match the lesion data obtained from the two monitoring sessions to determine the crop growth status.

[0028] In one implementation, the two-step preprocessing of first removing artifacts and then repairing overexposure can effectively restore key details such as the texture and outline of crops in the highlight area, solve the problem of distortion and feature loss in UAV farmland images, and provide reliable raw data support for subsequent model analysis; the artifact removal here is done through existing technical means, which will not be described in detail.

[0029] In one implementation, the preset weed crop segmentation model is a commonly used model on the market, such as the MobileSeg series, SegNet, DeepLab series, etc.; the preset monitoring cycle is determined by technical personnel.

[0030] In one implementation, the lesion data obtained from two monitoring sessions are correlated and matched. If the lesion area continues to shrink and there are no new lesions, it indicates that the crop growth is improving and the disease is under control. If the lesion area continues to expand and the proportion of new lesions is greater than 30%, it indicates that the crop growth is deteriorating and the disease is spreading. If the original lesions disappear and there are no new lesions, the crop has recovered to health. If the original lesions do not change significantly and a small number of new lesions appear, the crop growth is stable and the spread of the disease should be monitored.

[0031] In one embodiment, see Figure 2 , Figure 2 A flowchart is provided for a method of remediating crop overexposure, including... S1021, Calculate the brightness histogram of the denoised crop image, and determine the overexposure threshold based on the brightness histogram; S1022, determine candidate overexposed points in the denoised crop image based on the overexposure threshold and the difference between the RGB three channels; S1023, perform connected region analysis on candidate overexposed points, retain connected regions with an area greater than a preset number of pixels, fill holes in the region through morphological opening and closing operations, and remove edge burrs to obtain an exposure mask image. S1024, perform grayscale conversion on the denoised crop image to obtain a grayscale image, and separate the low-frequency and high-frequency components of the grayscale image; S1025, the target exposure mask is obtained by performing crop exposure processing on the exposure mask based on the low-frequency and high-frequency components; S1026, Correct the denoised crop image based on the target exposure mask image to obtain the original crop image.

[0032] In one implementation, the overexposure threshold is determined based on the brightness histogram, candidate overexposure points are screened by combining the differences in the RGB three channels, and the mask is optimized by connected component analysis and morphological operations. This can accurately distinguish between the real overexposure area and the crop highlight area, avoiding the misjudgment of normal highlight area as overexposure area for ineffective repair. In particular, after the rice ears are filled, they are pale yellow and reflect light in patches under sunlight, which can easily lead to highlight overflow.

[0033] In one implementation, determining the overexposure threshold based on the brightness histogram specifically involves dynamically determining the threshold based on the bimodal distribution characteristics of the brightness histogram of the input image. If the histogram shows a normal brightness peak and an overexposure peak, the valley between the two peaks is taken as the initial threshold; if it is a single peak, the sum of the brightness mean and twice the standard deviation is taken as the threshold.

[0034] In one implementation, a strategy of separating high and low frequency components of grayscale images is adopted. Low frequency components are used to ensure a smooth brightness transition in overexposed areas, while high frequency components are used to restore details such as crop texture and leaf veins. This solves the problem of detail loss after brightness restoration in traditional overexposure restoration, providing high-quality image data for subsequent weed segmentation and lesion identification. Morphological opening and closing operations can fill holes in overexposed areas and remove edge burrs, resulting in a well-defined edge in the target exposure mask image. No patchy restoration marks will appear when the image is corrected, and the brightness and texture transition of the restored image are natural, with strong visual consistency with non-overexposed areas.

[0035] In one implementation, pixels within the crop semantic region that simultaneously meet the following two conditions are considered overexposed: first, the brightness is greater than the exposure threshold; second, the difference between the RGB three channels is less than 10. The preset number is determined by the technicians and is set to 50 by default. The low-frequency and high-frequency components of the grayscale image are separated by Gaussian filtering with an initial σ value of 1.

[0036] In one embodiment, obtaining a target exposure mask by performing crop exposure processing on the exposure mask image based on low-frequency and high-frequency components includes: By pre-setting the normal brightness range of crops, the low-frequency components corresponding to the exposure mask are brightness mapped so that their mean and variance are consistent with the normal crop brightness distribution, thus obtaining the low-frequency correction distribution. Calculate the channel imbalance and saturation deviation of the corresponding pixels in the exposure mask, and determine the heavily distorted and lightly distorted areas in the exposure mask based on the channel imbalance and saturation deviation. The low-frequency correction distribution is mapped onto the severe distortion region to obtain the corrected severe distortion region; For the slightly distorted area, the normal crop pixels at the edge of the distorted area are used as the reference. The RGB values ​​of the edge pixels are filled by bilinear interpolation, and then saturation compensation is performed to correct the slightly distorted area. The exposure mask is corrected by correcting the severely distorted area and the slightly distorted area to obtain the corrected exposure mask; The target exposure mask is obtained by adding high-frequency components to the corrected exposure mask using a feature fusion algorithm.

[0037] In one implementation, a field scene with natural lighting without strong direct sunlight or shadows is selected. At least 50 such crop images corresponding to the same growth cycle of the current species are collected. Then, the crop area is converted into a grayscale image, and the brightness values ​​of all pixels are extracted to form a brightness dataset. The mean and standard deviation of the brightness of the dataset are calculated. The minimum value is the mean brightness minus twice the standard deviation, and the maximum value is the mean brightness plus twice the standard deviation.

[0038] In one implementation, the channel imbalance and saturation deviation of the corresponding pixels in the exposure mask are calculated using the formula D=|RG|+|GB|+|BR|. The channel imbalance is calculated using the formula ΔS=S-S0, where S0 is the saturation of the crop image corresponding to the same growth cycle of the current species under conditions of no strong direct light and no shadow occlusion, and S is the current saturation. The severely distorted area is defined as (D<10 and ΔS>0.5), the slightly distorted area is defined as (10≤D<20 and 0.3≤ΔS≤0.5), and the remaining areas are not processed.

[0039] In one implementation, the low-frequency correction distribution is mapped to the heavily distorted region to obtain the corrected heavily distorted region. The RGB mean (R0, G0, B0) of the corresponding crop is called from the preset feature library (which stores the RGB mean of the crop in the crop image corresponding to the same growth cycle of the current species in a field with no direct sunlight and no shadow). The normalized low-frequency components are mapped to the intervals [R0-20, R0+20], [G0-20, G0+20], and [B0-20, B0+20].

[0040] In one implementation, a severely distorted area and a slightly distorted area are divided based on channel imbalance and saturation deviation, and different repair strategies are adopted accordingly: the severely distorted area achieves overall brightness correction through low-frequency brightness mapping, while the slightly distorted area completes local detail restoration through edge pixel interpolation and saturation compensation, avoiding over-correction or under-correction of overexposed areas caused by a one-size-fits-all repair.

[0041] In one implementation, low-frequency components are mapped based on a preset normal crop brightness range, so that the mean and variance of the brightness of the overexposed area after correction are consistent with the brightness distribution of the normal crop. This solves the problem of overexposed areas being too dark or too bright after correction in traditional restoration, and ensures the naturalness of the image brightness. The separated high-frequency components are added back to the corrected exposure mask map through a feature fusion algorithm, avoiding the loss of key texture information during low-frequency correction, and providing clear detail support for subsequent weed segmentation and lesion identification.

[0042] In one implementation, the weight q of the heavily distorted region is defined as 0.7, and the high-frequency components are added to the corrected exposure mask using the formula A(x,y)=qM(x,y)+(1-q)T(x,y) (this formula only adds the heavily distorted and lightly distorted regions). A is the target exposure mask, M is the high-frequency component, and T is the corrected exposure mask.

[0043] In one embodiment, substituting the target crop image into the lesion identification model to obtain the lesion type and lesion location includes: Substituting the target crop map into the YOLOv10 backbone network yields high-scale, mid-scale, and low-scale maps; the high-scale map is the output of the fifth-layer C2f module; the mid-scale map is the output of the seventh-layer C2f module; and the low-scale map is the output of the eleventh-layer PSA module. High-scale features are obtained by performing a 1×1 convolution operation on the high-scale image; low-scale features are obtained by performing a 1×1 convolution operation on the low-scale image. Substituting the mesoscale map into the preset enhancement fusion model yields the mesoscale enhanced features; After pixel resampling of the low-scale image, it is concatenated with the mid-scale enhancement feature and then substituted into the C3k2 module to obtain the first fused feature; The first fused feature is upsampled by pixel recombination and then concatenated with the high-scale feature before being substituted into the C3k2 module to obtain the first output feature; After downsampling the first output feature by performing a 1×1 convolution operation, it is fused with the first fused feature and the mesoscale enhancement feature to obtain the second fused feature; Substitute the second fused feature into the C3k2 module to obtain the second output feature; The second output feature is downsampled and fused with the low-scale feature before being substituted into the C3k2 module to obtain the third output feature; The first, second, and third output features are input into the head network of YOLOv10 to obtain the lesion type and lesion location.

[0044] In one implementation, see [link to implementation details]. Figure 3 , Figure 3 A schematic diagram of a lesion identification model is provided; different levels of YOLOv10 modules are selected for output (high, medium and low scale images). The high-scale image focuses on leaf-level micro-lesions (such as pinhead-sized lesions), the medium-scale image covers plant-level lesion areas, and the low-scale image captures the distribution characteristics of lesions at the plot level, solving the problem of missing small lesions and misjudging large lesions due to single-scale features; all fusions are performed to calculate the average.

[0045] In one implementation, only 1×1 convolution operations are performed on high- and low-scale images. While compressing the feature dimension and reducing the amount of computation, the core features such as texture and contour of the lesion are fully preserved, avoiding feature redundancy caused by complex convolution and adapting to the real-time inference requirements of UAVs. For the medium-scale image, a preset enhancement fusion model is introduced separately to specifically enhance the contrast features between crop lesions and normal tissues.

[0046] In one implementation, a bidirectional fusion link with forward progression and reverse feedback is constructed: forward, low-scale features are gradually integrated into medium and high-scale features through pixel recombination and upsampling, realizing the transfer of low-level spatial details to high-level semantics; reverse, high-scale features are backpropagated to medium and low-scale feature layers through 1×1 convolution downsampling, completing the optimization of low-level details by high-level semantics. This bidirectional closed-loop fusion mode effectively avoids the problem of detail loss or semantic ambiguity in the feature transfer of existing models, and enables the edge contour of the lesion (low-level details) to form a precise association with the category attribute (high-level semantics).

[0047] In one implementation, existing models are prone to loss of crop lesion edge features during feature fusion. The core reason is the lack of enhancement processing for key scale features. In this embodiment, a preset enhancement fusion model is introduced separately for the mesoscale map to specifically enhance the edge contrast features between lesions and normal tissues. The C3k2 module is then used as the enhancement unit for the fused features. Its efficient feature aggregation capability can firmly lock the lesion edge information and avoid feature dilution during the fusion process.

[0048] In one embodiment, substituting the mesoscale map into a preset enhancement fusion model to obtain mesoscale enhanced features includes: Channel segmentation of the mesoscale image yields the first channel image and the second channel image; The first channel image is obtained by substituting it into the convolutional layer and the multilayer perceptron layer in sequence and then discarding random paths. The first channel augmentation map and the second channel map are stitched together to obtain the initial mesoscale augmentation map; The initial mesoscale augmentation map is subjected to a 3×3 convolution operation followed by global pooling to obtain global pooling features. The initial mesoscale augmented image is processed by Y-axis pooling and X-axis pooling respectively to obtain Y-axis pooling features and X-axis pooling features; Perform 1×1 convolution operations on the global pooling features, Y-direction pooling features, and X-direction pooling features respectively, and then input the Sigmoid activation function to obtain the global pooling weights, Y-direction weights, and X-direction weights; The Y-direction weights, X-direction weights, and global pooling weights are applied to the initial mesoscale augmentation map to obtain mesoscale augmentation features.

[0049] In one implementation, existing models suffer from the easy loss of edge features of crop lesions and low recognition accuracy. The core problem lies in the insufficient enhancement of key mesoscale features and the lack of coordinated optimization of channel and spatial features. The mesoscale map, as the core carrier connecting high-scale details and low-scale global features, achieves precise activation of crop lesion features through innovative design of channel differentiation enhancement and multi-dimensional spatial weighting. By splitting the mesoscale map into first and second channel maps through channel segmentation, the first channel map is specifically subjected to convolution to extract details and multilayer perceptron to enhance semantic links, accurately capturing features that are significantly different from those of crop lesions and healthy tissues. The random path discarding operation further eliminates invalid channel interference, making the first channel enhanced map a core feature carrier rich in lesion information, laying a high-quality foundation for subsequent fusion.

[0050] One implementation addresses the problem of one-sided spatial feature capture and blurred lesion edges in existing models by constructing a global, X / Y 3D pooling weight system: global pooling is performed on the initial mesoscale enhancement image to obtain overall feature weights, while separate pooling in the X and Y directions accurately captures the spatial distribution features of lesions in the horizontal and vertical dimensions. After 1×1 convolution dimensionality reduction and Sigmoid activation, the three types of weights can respectively weight the initial image from the three dimensions of global importance, horizontal extension, and vertical contour, so that key features such as lesion edges and lesion distribution are accurately enhanced, improving the features corresponding to defects.

[0051] In one implementation, the Y-direction weights, X-direction weights, and global pooling weights are applied to the initial mesoscale enhanced image, respectively; Y-direction weighting: the Y-direction weights are multiplied by the initial features after being broadcast along the height dimension; X-direction weighting: the X-direction weights are multiplied by the above results after being broadcast along the width dimension; global weighting: the global pooling weights are multiplied by the above results after being broadcast along the spatial dimension.

[0052] Based on the same inventive concept, this invention also provides a drone-based dynamic monitoring system for farmland. See also Figure 4 , Figure 4 A framework diagram of a UAV-based farmland dynamic monitoring system provided for embodiments of the present invention includes: The denoising module is used to acquire crop images collected by the drone and perform artifact denoising processing on the crop images to obtain denoised crop images. The overexposure repair module is used to repair overexposure of crops in denoised crop images to obtain the original crop images. The target crop map generation module is used to input the original crop map into a preset weed crop segmentation model to obtain the target crop map; The lesion identification module is used to input the target crop image into the lesion identification model to obtain the lesion type and lesion location; The lesion re-identification module is used to record the lesion type and lesion location corresponding to all crop images, collect crop images again based on a preset monitoring cycle, and re-determine the lesion type and lesion location corresponding to all crop images. The crop growth status determination module is used to correlate and match lesion data obtained from two monitoring sessions to determine the crop growth status.

[0053] Based on the UAV-based dynamic monitoring system for farmland provided by this invention, the system first performs artifact denoising and overexposure repair on the acquired images to restore key crop details. Then, based on weed and crop segmentation and lesion identification models, it accurately distinguishes crops and determines the type and location of lesions. Finally, by combining multi-cycle monitoring data matching, it dynamically grasps the crop growth status, significantly improving the accuracy and efficiency of farmland monitoring.

[0054] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: Obtain crop images collected by drones, and perform artifact denoising processing on the crop images to obtain denoised crop images; The denoised crop image is subjected to crop overexposure repair processing to obtain the original crop image; Substitute the original crop map into the preset weed crop segmentation model to obtain the target crop map; Substitute the target crop image into the lesion identification model to obtain the lesion type and lesion location; Record the lesion type and location corresponding to all crop images, collect crop images again based on the preset monitoring cycle, and determine the lesion type and location corresponding to all crop images again; The lesion data obtained from the two monitoring sessions were correlated and matched to determine the crop growth status. The original crop image obtained by performing crop overexposure repair processing on the denoised crop image includes: Calculate the brightness histogram of the denoised crop image, and determine the overexposure threshold based on the brightness histogram; Candidate overexposure points in the denoised crop image are determined based on the overexposure threshold and the differences in the RGB three channels; Connectivity analysis is performed on candidate overexposed points, and connected regions with an area greater than a preset number of pixels are retained. Holes in the regions are filled by morphological opening and closing operations, and edge burrs are removed to obtain an exposure mask. The denoised crop image is converted to grayscale to obtain a grayscale image, and the low-frequency and high-frequency components of the grayscale image are separated. The target exposure mask is obtained by performing crop exposure processing on the exposure mask based on the low-frequency components and the high-frequency components. The original crop image is obtained by correcting the denoised crop image based on the target exposure mask image. The target exposure mask is obtained by performing crop exposure processing on the exposure mask based on the low-frequency components and the high-frequency components, including: By pre-setting the normal brightness range of crops, the low-frequency components corresponding to the exposure mask are brightness mapped so that their mean and variance are consistent with the normal crop brightness distribution, thus obtaining the low-frequency correction distribution. Calculate the channel imbalance and saturation deviation of the corresponding pixels in the exposure mask, and determine the heavily distorted area and the lightly distorted area in the exposure mask based on the channel imbalance and the saturation deviation. The low-frequency correction distribution is mapped onto the severe distortion region to obtain the corrected severe distortion region; For the slightly distorted area, the normal crop pixels at the edge of the distorted area are used as the reference. The RGB values ​​of the edge pixels are filled by bilinear interpolation, and then saturation compensation is performed to correct the slightly distorted area. The exposure mask is corrected based on the areas of severe distortion and mild distortion to obtain a corrected exposure mask. The high-frequency components are added to the corrected exposure mask using a feature fusion algorithm to obtain the target exposure mask.

2. The method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Substituting the target crop image into the lesion identification model yields the lesion type and lesion location, including: The target crop map is substituted into the YOLOv10 backbone network to obtain high-scale, mid-scale, and low-scale maps; the high-scale map is the output of the fifth-layer C2f module; the mid-scale map is the output of the seventh-layer C2f module; and the low-scale map is the output of the eleventh-layer PSA module. A 1×1 convolution operation is performed on the high-scale image to obtain high-scale features; a 1×1 convolution operation is performed on the low-scale image to obtain low-scale features; Substituting the mesoscale map into a preset enhancement fusion model yields mesoscale enhanced features; After pixel resampling of the low-scale image, it is concatenated with the mid-scale enhancement feature and then substituted into the C3k2 module to obtain the first fused feature; The first fused feature is upsampled by pixel recombination and then concatenated with the high-scale feature before being substituted into the C3k2 module to obtain the first output feature; After performing a 1×1 convolution operation on the first output feature and downsampling it, it is fused with the first fused feature and the mesoscale enhancement feature to obtain the second fused feature; Substitute the second fusion feature into the C3k2 module to obtain the second output feature; The second output feature is downsampled and fused with the low-scale feature before being substituted into the C3k2 module to obtain the third output feature; The first, second, and third output features are input into the head network of YOLOv10 to obtain the lesion type and lesion location.

3. The method for dynamic monitoring of farmland based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, Substituting the mesoscale map into a preset enhancement fusion model yields the following mesoscale enhanced features: The mesoscale image is segmented into channels to obtain a first channel image and a second channel image; The first channel map is sequentially substituted into the convolutional layer and the multilayer perceptron layer, and then random path discarding is performed to obtain the first channel augmentation map; The first channel enhancement map and the second channel map are stitched together to obtain an initial mesoscale enhancement map; The initial mesoscale augmented image is subjected to a 3×3 convolution operation followed by global pooling to obtain global pooled features; The initial mesoscale augmented image is subjected to Y-axis pooling and X-axis pooling respectively to obtain Y-axis pooling features and X-axis pooling features; Perform 1×1 convolution operations on the global pooling feature, the Y-direction pooling feature, and the X-direction pooling feature respectively, and then input the Sigmoid activation function to obtain the global pooling weight, the Y-direction weight, and the X-direction weight; The Y-direction weights, X-direction weights, and global pooling weights are applied to the initial mesoscale augmentation map to obtain mesoscale augmentation features.

4. A farmland dynamic monitoring system based on unmanned aerial vehicles (UAVs), characterized in that, The system includes: The noise reduction module is used to acquire crop images collected by the drone and perform artifact noise reduction processing on the crop images to obtain denoised crop images. The overexposure repair module is used to perform crop overexposure repair processing on the denoised crop image to obtain the original crop image. The target crop map generation module is used to substitute the original crop map into a preset weed crop segmentation model to obtain the target crop map. The lesion identification module is used to input the target crop image into the lesion identification model to obtain the lesion type and lesion location; The lesion re-identification module is used to record the lesion type and lesion location corresponding to all crop images, collect crop images again based on a preset monitoring cycle, and re-determine the lesion type and lesion location corresponding to all crop images. The crop growth status determination module is used to correlate and match the lesion data obtained from two monitoring sessions to determine the crop growth status. The overexposure repair module includes: The overexposure threshold calculation module is used to calculate the brightness histogram of the denoised crop image and determine the overexposure threshold based on the brightness histogram. The candidate overexposure point determination module is used to determine the candidate overexposure points in the denoised crop image based on the overexposure threshold and the difference between the RGB three channels. The exposure mask generation module is used to perform connected region analysis on candidate overexposed points, retain connected regions with an area greater than a preset number of pixels, fill holes in the region through morphological opening and closing operations, and remove edge burrs to obtain the exposure mask. The low- and high-frequency component determination module is used to perform grayscale conversion on the denoised crop image to obtain a grayscale image, and to separate the low-frequency and high-frequency components of the grayscale image; The target exposure mask generation module is used to perform crop exposure processing on the exposure mask based on the low-frequency components and the high-frequency components to obtain the target exposure mask; The original crop image generation module is used to correct the denoised crop image based on the target exposure mask image to obtain the original crop image; The target exposure mask generation module includes: The low-frequency correction distribution determination module is used to perform brightness mapping on the low-frequency components corresponding to the exposure mask image by preset normal crop brightness range, so that its mean and variance are consistent with the normal crop brightness distribution to obtain the low-frequency correction distribution; The distortion area determination module is used to calculate the channel imbalance and saturation deviation of the corresponding pixels in the exposure mask, and determine the heavily distorted area and the lightly distorted area in the exposure mask based on the channel imbalance and the saturation deviation. The severe distortion region determination module is used to map the low-frequency correction distribution to the severe distortion region to obtain the severe distortion region. The module for determining the slightly distorted area is used to fill the RGB values ​​of the edge pixels of the distorted area with normal crop pixels at the edge of the distorted area as a reference, and then perform saturation compensation to obtain the corrected slightly distorted area. The corrected exposure mask generation module is used to correct the exposure mask based on the corrected severe distortion area and the corrected mild distortion area to obtain a corrected exposure mask. The target exposure mask determination module is used to add the high-frequency components to the corrected exposure mask using a feature fusion algorithm to obtain the target exposure mask.

5. A farmland dynamic monitoring system based on unmanned aerial vehicles (UAVs) according to claim 4, characterized in that, The lesion identification module includes: A multi-scale extraction module is used to input the target crop map into the backbone network of YOLOv10 to obtain high-scale, medium-scale, and low-scale maps; the high-scale map is the output of the fifth-layer C2f module; the medium-scale map is the output of the seventh-layer C2f module; and the low-scale map is the output of the eleventh-layer PSA module. The low-scale feature determination module is used to perform a 1×1 convolution operation on the high-scale image to obtain high-scale features; and to perform a 1×1 convolution operation on the low-scale image to obtain low-scale features. The mesoscale enhancement feature determination module is used to substitute the mesoscale map into a preset enhancement fusion model to obtain mesoscale enhancement features; The first fusion feature generation module is used to perform pixel re-sampling on the low-scale image, then concatenate it with the mid-scale enhancement feature, and then substitute it into the C3k2 module to obtain the first fusion feature; The first output feature generation module is used to perform pixel recombination and upsampling on the first fused feature, then concatenate it with the high-scale feature and then substitute it into the C3k2 module to obtain the first output feature; The second fusion feature generation module is used to perform a 1×1 convolution operation on the first output feature, downsample it, and then fuse it with the first fusion feature and the mesoscale enhancement feature to obtain the second fusion feature. The second output feature generation module is used to substitute the second fused feature into the C3k2 module to obtain the second output feature; The third output feature generation module is used to downsample the second output feature and fuse it with the low-scale feature before substituting it into the C3k2 module to obtain the third output feature; The lesion identification module is used to input the first output feature, the second output feature, and the third output feature into the head network of YOLOv10 to obtain the lesion type and lesion location.

6. A farmland dynamic monitoring system based on unmanned aerial vehicles (UAVs) according to claim 5, characterized in that, The mesoscale enhanced feature determination module includes: The channel segmentation module is used to perform channel segmentation on the mesoscale image to obtain a first channel image and a second channel image; The first channel enhancement map determination module is used to sequentially substitute the first channel map into the convolutional layer and the multilayer perceptron layer and then randomly discard the path to obtain the first channel enhancement map. An initial mesoscale enhancement map generation module is used to stitch the first channel enhancement map and the second channel map to obtain an initial mesoscale enhancement map; The global pooling feature generation module is used to perform a 3×3 convolution operation on the initial mesoscale augmented map and then perform global pooling processing to obtain global pooling features. The XY-direction pooling module is used to perform Y-direction pooling and X-direction pooling on the initial mesoscale augmented image to obtain Y-direction pooling features and X-direction pooling features, respectively. The XY weight determination module is used to perform 1×1 convolution operations on the global pooling feature, the Y-direction pooling feature and the X-direction pooling feature respectively, and then input the Sigmoid activation function to obtain the global pooling weight, the Y-direction weight and the X-direction weight; The mesoscale enhancement feature generation module is used to apply the Y-direction weights, the X-direction weights, and the global pooling weights to the initial mesoscale enhancement map to obtain mesoscale enhancement features.

Images