A straw index evaluation method based on large-scale image mosaicking

Abstract

This invention, applicable to the field of agricultural information technology, provides a straw index assessment method based on large-scale image stitching. By combining image stitching technology with straw image segmentation technology, this method achieves high-precision estimation of straw coverage within a large area, significantly improving the accuracy of straw coverage estimation for large images. Using image stitching technology, the image resolution is largely preserved on the basis of existing single images, ensuring the integrity of details in large-scale images, thereby improving the accuracy of coverage estimation and effectively enhancing the accuracy and reliability of large-scale straw coverage assessment.

Description

Technical Field

[0001] The invention belongs to the field of agricultural information technology, and in particular relates to a straw index evaluation method based on large-scale image splicing. Background Art

[0002] In agricultural management, accurate assessment of straw cover is crucial for monitoring the effectiveness of straw return and scientifically managing soil fertility. With the widespread application of drones and image processing technology, straw cover assessment is gradually shifting from traditional manual sampling to image-based analysis. However, existing technologies still face numerous challenges and technical limitations when assessing straw cover over large areas.

[0003] Existing segmentation algorithms and image stitching techniques typically focus on high-precision processing within small areas, enabling target recognition and segmentation within a single image. However, when processing large areas, existing segmentation algorithms are prone to recognition errors and discontinuous segmentation due to the increased data volume and complex regional features. This not only affects the input quality of the segmentation algorithm but also limits the practical application of stitching technology in the agricultural sector.

[0004] At the same time, most current coverage assessment systems rely on fixed analysis processes and preset parameters, making them inflexible and incapable of adapting to actual needs. For example, when assessing coverage for a specific area (such as a region of interest, or ROI), existing methods lack interactive segmentation and dynamic analysis capabilities, making them difficult to meet the diverse needs of agricultural applications. This fixed process limits the technology's applicability and practicality.

[0005] Furthermore, existing coverage assessment methods for remote sensing or multispectral imagery are limited in their widespread adoption due to the complexity and high cost of acquisition equipment. While drone-based RGB imagery acquisition is simpler, its implementation still faces challenges in data quality and efficiency in large-scale regional analysis.

[0006] In summary, existing technologies can achieve high accuracy in small-scale image analysis, but still face significant challenges in large-scale regional assessments. Therefore, for large-scale assessments of straw coverage, an improved technical solution is urgently needed that can address these issues by optimizing the segmentation network, improving the accuracy of image stitching technology, and enhancing user flexibility. Summary of the Invention

[0007] The purpose of the embodiments of the present invention is to provide a straw index evaluation method based on large-scale image stitching, aiming to solve the problems raised in the above background technology.

[0008] The embodiment of the present invention is implemented as follows: a straw index evaluation method based on large-scale image splicing includes the following steps:

[0009] S1: UAV image acquisition;

[0010] Use drones to capture images of designated areas, follow a pre-set flight path to cover the target area, and collect straw image data; save the images in a fixed format and specific resolution, and attach metadata information;

[0011] S2: UAV image stitching processing;

[0012] The collected images are processed using image stitching technology to generate a complete image of the target area. The stitching process combines regional feature extraction and geometric correction strategies to ensure the alignment consistency and resolution consistency of the stitching results.

[0013] S3: stitching image segmentation algorithm;

[0014] Perform image segmentation on the stitched target area image to generate a complete mask image of the target area. The segmentation algorithm includes block processing of the stitched image, patch image segmentation and mask stitching restoration.

[0015] S4: calculation of stitched image metrics;

[0016] Based on the complete mask image of the target area generated by S3, the proportion of target pixels in the specified area is calculated using the specified area set by user interaction or automatically divided by the system to obtain the indicator results of the target area.

[0017] Further technical solutions, said S1 specifically includes:

[0018] S101: Using a drone to collect images of a designated area to obtain images of straw coverage; the drone comprehensively covers the target area according to a preset flight route, and the collected images are saved in a fixed format and a specific resolution, and necessary metadata information is attached;

[0019] S102: Dataset creation and annotation: First, create a dataset and use annotation tools to annotate the collected images to generate annotation files corresponding to the straw-covered areas, which are used as label data required for deep learning model training; then perform segmentation network training, based on the created dataset, train a dedicated segmentation network for automatically identifying straw-covered areas in images; during the training process, adjust network parameters and optimization strategies to ensure that the model can accurately complete the segmentation task.

[0020] A further technical solution, in the S2, first, the scale-invariant feature change (SIFT) algorithm is used to detect feature points of multiple images taken by the drone, and corresponding feature descriptors are generated; secondly, the cluster matching algorithm is used to match the feature points between the images; then, the feature matching results are optimized based on the random sampling consensus (RANSAC) algorithm, the mismatched points are removed and the homography change matrix between the images is calculated to achieve geometric correction between the images; finally, the multi-band fusion method is used to fuse the overlapping areas of the stitching to eliminate the discontinuity of the image boundaries and ensure the alignment consistency and resolution consistency of the stitched images; after the stitching is completed, the generated complete target area image will be used as the input of the image segmentation algorithm in the subsequent steps.

[0021] Further technical solutions, said S3 specifically includes:

[0022] S301: Patch processing of the spliced ​​image;

[0023] The stitched target area image is divided into fixed-size, non-overlapping patches using a regular grid. The overlapping areas between patches are dynamically increased for areas with dense straw distribution or complex backgrounds. Patch areas with insufficient edge size are mirrored and filled, and the precise position coordinate information of each patch in the original image is recorded.

[0024] S302: Patch image segmentation;

[0025] Input the segmented patch image into the deep learning segmentation network trained in S1 to generate the corresponding mask image and visualize the result;

[0026] S303: Concat restoration of mask image;

[0027] The predicted mask maps are gradually spliced ​​and restored to the complete target area mask map, thereby generating a complete mask map corresponding to the original target area.

[0028] Further technical solution, said S4 specifically includes:

[0029] First, the designated area for calculating straw coverage is determined through user interaction or system automatic segmentation. Then, within the designated area, the number of target pixels is counted based on the straw target pixel values ​​marked in the mask image. The target pixel values ​​are binarized to specific values, and the statistical results are used to measure the coverage of the straw area within the specified range. Finally, the output results include coverage values, coverage distribution maps, and additional information.

[0030] The straw coverage area is calculated by calibrating the pixel distance between ridges in the image and the actual ridge spacing, and the straw image variation parameters are calculated through different convolution operations.

[0031] The present invention provides a straw index assessment method based on large-scale image stitching. By combining image stitching technology with straw image segmentation technology, this method achieves high-precision estimation of straw coverage within a large area, significantly improving the accuracy of straw coverage estimation for large images. Using image stitching technology, the image resolution is largely preserved on the basis of existing single images, ensuring the integrity of details in large-scale images, thereby improving the accuracy of coverage estimation and effectively enhancing the accuracy and reliability of large-scale straw coverage assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a straw index evaluation method based on large-scale image stitching provided by an embodiment of the present invention;

[0033] Figure 2 The straw coverage mask map predicted by the dedicated network;

[0034] Figure 3 This is the effect diagram of straw splicing;

[0035] Figure 4 It is the overall mask image restored by Concat;

[0036] Figure 5 Calculate the effect of the parameters. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0039] like Figure 1 As shown, a straw index evaluation method based on large-scale image stitching provided by an embodiment of the present invention includes the following steps:

[0040] S1: UAV image acquisition;

[0041] The designated area is imaged by a drone, and the target area is covered according to the preset flight route to provide high-quality straw image data for subsequent processing. The collected straw images are manually annotated using professional annotation tools to generate data labels for training. During the dataset production process, the images are adjusted to a uniform resolution to ensure the consistency of the annotation results and meet the input requirements of the subsequent segmentation network. Based on the annotated data, a dedicated straw image deep learning segmentation network is designed. Based on the classic encoder-decoder network, the network structure further integrates feature fusion modules and adaptive attention mechanisms to enhance the ability to extract straw features and improve the accuracy of image segmentation in complex environments. After the segmentation network training is completed, the straw images collected by the drone are input into the network to predict and generate a straw coverage mask map (mask map), as shown below. Figure 2 shown.

[0042] S2: UAV image stitching processing;

[0043] Since the images taken by drones have a fixed resolution and a single image covers a limited area, directly using it for target area segmentation may result in reduced accuracy. Using multiple images taken by drones, a complete target area image is generated through image stitching technology. The stitching process combines regional feature extraction and geometric correction strategies to ensure the alignment consistency and resolution consistency of the stitching results. The large-scale image obtained by stitching the drone images in S1 is as follows: Figure 3 shown.

[0044] S3: stitching image segmentation algorithm;

[0045] Perform image segmentation on the spliced ​​target area image to generate a complete mask image of the target area. This step includes the following sub-processes:

[0046] S301: Patch processing of the spliced ​​image;

[0047] The specific segmentation strategy is as follows: First, according to the standard input size accepted by the trained segmentation network (such as 512×512 pixels), the stitched target area image is divided into multiple fixed-size and non-overlapping image patches in a grid manner, so that the size of all patches is unified to adapt to the network input requirements; then, for the patch areas at the edge of the stitched image that cannot fully reach the standard size, a mirror padding strategy is used to fill the image to ensure that all patches are of the same size, thereby fully retaining the straw image features in the edge area; then, taking into account the special complexity of straw distribution, the overlap rate during segmentation is dynamically adjusted: in areas with dense straw and high background complexity, the overlap ratio between patches is appropriately increased to further improve the segmentation accuracy of these special areas; finally, during the segmentation process, a unique index identifier is assigned to each patch, and the precise position coordinates of each patch relative to the original stitched image (such as the coordinates of the upper left corner of the patch) are recorded in detail to ensure that the corresponding position of the original image can be accurately restored when the mask image is subsequently stitched.

[0048] S302: Patch image segmentation;

[0049] The segmented patch image is fed into the deep learning segmentation network trained in S1 to generate the corresponding mask image and visualize the results. This step ensures that each small image patch can be independently segmented with high precision, thereby improving the segmentation accuracy of the overall spliced ​​image.

[0050] S303: Mask image concat restoration

[0051] like Figure 4 As shown, the overall mask image restored by Concat is obtained. According to the patch position information recorded in S301, the predicted mask images are gradually concatenated and restored to a complete target area mask image, thereby generating a complete mask image corresponding to the original target area.

[0052] S4: calculation of stitched image metrics;

[0053] Based on the complete mask image of the target area generated by S3, the proportion of target pixels in the specified area is calculated using the specified area set by user interaction or automatically divided by the system to obtain the indicator results of the target area.

[0054] As a preferred embodiment of the present invention, the S1 specifically includes:

[0055] S101: Use drones to capture images of a designated area, obtaining high-quality images of straw cover and providing baseline data for subsequent processing. The drones will follow a pre-set flight path to fully cover the target area, ensuring the integrity and consistency of the captured image data. To improve image quality, it is recommended to complete the acquisition under suitable environmental conditions, such as during periods of good lighting or stable weather. The acquired images must be saved in a fixed format and a specific resolution, with the necessary metadata attached (such as acquisition time, location, and device ID). These images will serve as the raw input for subsequent dataset creation and annotation.

[0056] S102: Dataset creation and annotation: First, create a dataset, use professional annotation tools to annotate the collected images, and generate annotation files corresponding to the straw-covered areas. These annotation files serve as the label data required for deep learning model training. To ensure data quality, all images need to be adjusted to a specific size and fixed pixel format to meet the requirements of subsequent network input. Then perform segmentation network training. Based on the prepared dataset, a dedicated segmentation network is trained to automatically identify straw-covered areas in images. During the training process, the network parameters and optimization strategies are adjusted to ensure that the model can accurately complete the segmentation task. The trained segmentation network will be used in the subsequent image segmentation steps.

[0057] As a preferred embodiment of the present invention, the S2 specifically includes:

[0058] Image stitching technology is used to process multiple images captured by drones to generate a complete target area image. First, the scale-invariant feature transform (SIFT) algorithm is used to detect feature points in the multiple drone images and generate corresponding feature descriptors. Second, a cluster matching algorithm is used to match feature points between images. Then, the feature matching results are optimized based on the random sampling consensus (RANSAC) algorithm to remove mismatched points and calculate the homography change matrix between the images to achieve geometric correction between the images. Finally, a multi-band fusion method is used to fuse the overlapping stitching areas to eliminate discontinuities at the image boundaries and ensure alignment and resolution consistency of the stitched images. After stitching is completed, the generated complete target area image will serve as the input for the image segmentation algorithm in the subsequent steps.

[0059] As a preferred embodiment of the present invention, the S4 specifically includes:

[0060] Based on the complete mask map of the target area generated by S3, the straw coverage in the target area is quantitatively analyzed to obtain the straw coverage rate estimation result. Figure 5Figure 2 shows a specialized image-based straw index calculation software application. First, the designated area for straw coverage calculation is determined through user interaction or automatic system segmentation. User interaction allows the selection of a region of interest (ROI) through a graphical interface, while automatic segmentation generates a designated area based on preset regional boundaries, geographic information, or an algorithm. Within the designated area, the target straw pixels are counted based on their values, which are binarized to specific values. The statistical results are used to measure the extent of straw coverage within the designated area.

[0061] Finally, if Figure 5 The figure shows the actual application results of this method. The calculation results are presented in numerical and graphical form, including coverage values ​​(e.g., as a percentage), coverage distribution maps, and other additional information (such as image time, location, or regional boundaries). The straw cover area is further calculated by calibrating the pixel distance between ridges in the image with the actual ridge spacing. Straw image variation parameters are calculated using different convolution operations to reflect the actual straw uniformity. These results can provide a scientific basis for agricultural management and analysis of straw distribution and coverage effectiveness.

[0062] This method combines image stitching with straw image segmentation to achieve high-precision estimation of straw cover over large areas, significantly improving the accuracy of straw cover estimation for large images. Image stitching, based on existing single images, preserves image resolution to a high degree, ensuring the integrity of details across large areas and thus improving the accuracy of cover estimation. Experimental results demonstrate that this method effectively improves the accuracy and reliability of large-scale straw cover assessment.

[0063] This method is not only applicable to estimating straw cover but also has strong scalability and can be extended to assess the cover of other crops or objects, such as weed cover, in the future. With the continued development of drone technology, this method, combined with different types of image data (such as multispectral and thermal infrared imagery) and interpretable artificial intelligence (xAI) methods, can further construct more complex and accurate regional environmental monitoring models, providing more intelligent and diverse detection and assessment solutions for the agricultural sector. In the future, by incorporating advanced deep learning algorithms, the applicability and practicality of the estimation model will be further improved, allowing this method to be applied to even more scenarios.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A straw index evaluation method based on large-scale image splicing, characterized in that: The following steps are involved: S1: UAV image acquisition; Use drones to collect images of designated areas, cover the target area according to the preset flight route, and collect straw image data; Save images in a fixed format and specific resolution, and append metadata information; S2: UAV image stitching processing; The multiple images collected are processed using image stitching technology to generate a complete image of the target area. The stitching process combines regional feature extraction and geometric correction strategies to ensure the alignment consistency and resolution consistency of the stitching results. S3: stitching image segmentation algorithm; Perform image segmentation on the spliced ​​target area image to generate a complete mask image of the target area. The segmentation algorithm includes block processing of the spliced ​​image, patch image segmentation and mask splicing restoration; S4: calculation of stitched image metrics; Based on the complete mask image of the target area generated by S3, the proportion of target pixels in the specified area is calculated using the specified area set by user interaction or automatically divided by the system to obtain the indicator results of the target area.

2. The straw index evaluation method based on large-scale image splicing according to claim 1 is characterized in that: Said S1 specifically includes: S101: Using a drone to collect images of a designated area to obtain images of straw coverage; the drone comprehensively covers the target area according to a preset flight route, and the collected images are saved in a fixed format and a specific resolution, and metadata information is attached; S102: Dataset creation and annotation: First, create a dataset and use annotation tools to annotate the collected images to generate annotation files corresponding to the straw-covered areas, which are used as label data required for deep learning model training; then perform segmentation network training, based on the created dataset, train a dedicated segmentation network for automatically identifying straw-covered areas in images; during the training process, adjust network parameters and optimization strategies to ensure that the model can accurately complete the segmentation task.

3. The straw index evaluation method based on large-scale image splicing according to claim 1 is characterized in that: In the S2, first, a scale-invariant feature change algorithm is used to detect feature points of multiple images taken by the drone, and corresponding feature descriptors are generated; secondly, a cluster matching algorithm is used to match feature points between images; then, the feature matching results are optimized based on a random sampling consistency algorithm, false matching points are removed, and the homography change matrix between images is calculated to achieve geometric correction between images; finally, a multi-band fusion method is used to fuse the overlapping areas of the stitching to eliminate the discontinuity of the image boundaries and ensure the alignment consistency and resolution consistency of the stitched images; after the stitching is completed, the generated complete target area image will be used as the input of the image segmentation algorithm in the subsequent steps.

4. The straw index evaluation method based on large-scale image splicing according to claim 1 is characterized in that: The S3 specifically includes: S301: Patch processing of the spliced ​​image; The stitched target area image is divided into fixed-size, non-overlapping patches using a regular grid. The overlapping areas between patches are dynamically increased for areas with dense straw distribution or complex backgrounds. Patch areas with insufficient edge size are mirrored and filled, and the precise position coordinate information of each patch in the original image is recorded. S302: Patch image segmentation; Input the segmented Patch image into the deep learning segmentation network trained in S1 to generate the corresponding mask image; S303: Concat restoration of mask image; The predicted mask maps are gradually spliced ​​and restored to the complete target area mask map, thereby generating a complete mask map corresponding to the original target area.

5. The straw index evaluation method based on large-scale image splicing according to claim 1 is characterized in that: The S4 specifically includes: First, the designated area for calculating straw coverage is determined through user interaction or system automatic segmentation. Then, within the designated area, the number of target pixels is counted based on the straw target pixel values ​​marked in the mask image. The target pixel values ​​are binarized to specific values, and the statistical results are used to measure the coverage of the straw area within the specified range. Finally, the output results include coverage values, coverage distribution maps, and additional information. The straw coverage area is calculated by calibrating the pixel distance between ridges in the image and the actual ridge spacing, and the straw image variation parameters are calculated through different convolution operations.

Images