Forest grass landscape pattern identification method based on unmanned aerial vehicle image

Abstract

The invention discloses a forest and grass landscape pattern identification method based on unmanned aerial vehicle images, and the method comprises the steps: carrying out the definition and cloud shielding screening of images collected by an unmanned aerial vehicle through a random forest algorithm, and obtaining a high-quality first image data set; then, in combination with vegetation types and spatial distribution characteristics, landscape elements are classified by using a support vector machine, acquisition parameters are optimized, and a second image data set is generated. Further considering the influence of seasonal variation on spectral characteristics, selecting representative time phase images, and constructing a third image data set; and plaque information is calculated and fine plaques are fused to obtain fused plaque distribution data, and then multi-scale pattern indexes such as a shape index, an aggregation degree, a separation degree and a spreading degree are analyzed. And finally, indexes related to image definition are extracted, and a landscape pattern report is generated. According to the method, full-process automation from data screening, element classification, time phase selection to pattern analysis is realized, and the recognition precision and adaptability of the forest and grass landscape pattern are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of smart forestry technology, and in particular relates to a method for identifying forest and grassland landscape patterns based on UAV imagery. Background Technology

[0002] In fields such as ecological environmental protection, natural resource management, and sustainable land use planning, accurate and efficient identification and dynamic monitoring of forest and grassland landscape spatial patterns are of paramount practical significance. As a core component of terrestrial ecosystems, the distribution range, patch morphology, connectivity, and fragmentation of forest and grassland resources directly impact regional biodiversity maintenance, soil and water conservation, and ecological security assessments. Especially in macro-level decision-making processes such as addressing climate change and implementing major ecological restoration projects, timely acquisition of high-precision landscape pattern information is fundamental to scientific policy implementation.

[0003] Currently, monitoring of forest and grassland landscapes over large areas mainly relies on satellite remote sensing imagery and traditional manned aerial photography. While these technologies play a crucial role in acquiring macroscopic and periodic data, they also face significant limitations. First, satellite imagery is constrained by fixed revisit periods and spatial resolution, making it difficult to flexibly respond to sudden monitoring needs or acquire sufficient detail under complex terrain conditions; its data quality is also easily affected by atmospheric conditions (such as cloud cover), leading to missing or unevenly distributed data. Second, traditional aerial photography is costly and lacks operational flexibility, making it difficult to achieve high-frequency, low-cost data acquisition.

[0004] At the data processing and analysis level, existing methods typically employ pre-defined, fixed image processing workflows and parameters. When faced with massive amounts of image data with higher spatiotemporal resolution acquired through technologies such as drones, existing solutions exhibit significant limitations in adaptability: First, they lack intelligent pre-screening mechanisms for data quality, failing to automatically and effectively remove low-quality images caused by uneven lighting, cloud cover, or other factors, thus affecting the reliability of subsequent analysis. Second, when processing high-resolution images, they easily generate numerous fragmented patches, increasing analytical complexity and potentially obscuring true landscape pattern characteristics. Third, most methods do not adequately consider the scale dependence of landscape patterns, failing to establish flexible, multi-scale analysis frameworks adapted to specific vegetation types and land cover characteristics, and rarely systematically incorporate the impact of seasonal variations on the spectral characteristics of land cover, leading to reduced accuracy and comparability in cross-temporal analyses.

[0005] Therefore, existing technologies still face many bottlenecks in processing massive, multi-source, and high-resolution remote sensing data to achieve automated, high-precision, and highly adaptable forest and grassland landscape pattern identification. There is an urgent need for a solution that can connect the entire process from intelligent data screening and feature optimization extraction to accurate multi-scale pattern analysis. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for identifying forest and grassland landscape patterns based on UAV imagery, comprising:

[0007] The image data collected by the drone is obtained and filtered based on the sharpness index and the cloud and fog occlusion index to obtain the first image dataset;

[0008] Based on the vegetation type and spatial distribution characteristics in the first image dataset, the optimal acquisition parameters are determined and the landscape elements are classified to obtain the second image dataset.

[0009] From the second image dataset, representative temporal images are selected based on the influence of seasonal changes on spectral features to obtain the third image dataset;

[0010] Based on the landscape elements and vegetation types in the third image dataset, the initial information of the patches is calculated and the fragmented patches are fused to obtain the fused patch distribution data.

[0011] Based on the fused patch distribution data, the shape index and aggregation index are analyzed, and the calculation of the separation and spread indexes is optimized to obtain a multi-scale pattern index set.

[0012] From the multi-scale pattern index set, indicators related to image sharpness are extracted and their accuracy is judged to generate a landscape pattern report.

[0013] Optionally, the acquisition of image data collected by the drone is filtered based on sharpness and cloud / fog occlusion indicators to obtain a first image dataset, including:

[0014] Acquire image data collected by drones;

[0015] The random forest algorithm is used to predict each image data and output quantitative scores for sharpness, lighting conditions, cloud cover and overlap.

[0016] If the clarity index is higher than a preset threshold and the cloud cover is lower than a preset threshold, the image data is retained; otherwise, it is discarded.

[0017] The first image dataset is composed of the retained image data.

[0018] Optionally, the step of determining the optimal acquisition parameters and classifying landscape elements based on the vegetation type and spatial distribution characteristics in the first image dataset to obtain the second image dataset includes:

[0019] Based on the first image dataset, the correlation between vegetation type and spatial distribution is analyzed, the flight altitude parameter settings are adjusted, and optimized altitude configuration data is obtained.

[0020] Based on the optimized height configuration data and the image resolution requirements, a suitable acquisition resolution standard is calculated, and a matching image acquisition scheme is determined.

[0021] From the matched image acquisition scheme, the distribution information of landscape elements is extracted, and the landscape elements are accurately classified using a support vector machine to obtain a set of classified elements.

[0022] By analyzing the classified set of elements, a multi-scale pattern recognition framework is constructed, the spatial relationships of each element within the framework are determined, and a second image dataset is generated.

[0023] Optionally, the step of selecting representative temporal images from the second image dataset based on the influence of seasonal variations on spectral features to obtain a third image dataset includes:

[0024] From the second image dataset, spectral data of different time phases are extracted to obtain a separated temporal spectral set;

[0025] Analyze the correspondence between spectral features and pattern indicators in the aforementioned temporal spectrum set, compare the feature changes of each temporal data, and determine the variation rate of feature changes;

[0026] If the variation rate exceeds a preset threshold, the temporal spectrum set is filtered to obtain a representative temporal image subset.

[0027] The image data in the temporal image subset is classified using a support vector machine to obtain a set of image categories after classification.

[0028] A third image dataset is generated by integrating representative image data from different time periods.

[0029] Optionally, the step of calculating initial patch information and fusing fragmented patches based on landscape features and vegetation types in the third image dataset to obtain fused patch distribution data includes:

[0030] For the landscape elements and vegetation types in the third image dataset, the initial values ​​of the number of patches and patch area are calculated to obtain basic patch statistics.

[0031] If there are fragmented patches in the basic patch statistics due to high resolution, the random forest algorithm is used to classify and merge the fragmented patches to determine the optimized patch boundary data.

[0032] Based on the optimized patch boundary data, the number and area of ​​patches are recalculated to obtain the adjusted patch statistics.

[0033] Based on the adjusted patch statistics, a merged patch distribution layer is generated to obtain unified patch distribution information.

[0034] Optionally, the step of analyzing shape indices and clustering indices based on the fused patch distribution data, and optimizing the calculation of separation and spread indices, yields a multi-scale pattern index set, including:

[0035] The shape index and clustering degree are obtained from the fused patch distribution data;

[0036] The shape index distribution location is determined by dividing the patch distribution data into spatial grids, and the aggregation degree within each grid is calculated;

[0037] Support vector machine is used to classify and optimize separation and spread, resulting in optimized separation and spread values;

[0038] By resampling the patch distribution data at multiple scales, the shape index and clustering degree at different scales are obtained;

[0039] Based on the shape index and aggregation degree at different scales, and combined with the optimized separation degree and spread degree values, a multi-scale pattern index set is determined.

[0040] Optionally, the step of extracting indicators related to image sharpness from the multi-scale pattern indicator set, determining their accuracy, and generating a landscape pattern report includes:

[0041] From the set of multi-scale pattern indicators, a preliminary screening based on the correlation with image sharpness is performed to obtain a subset of indicators that are highly correlated with sharpness.

[0042] The subset of indicators is processed using a standardized format to convert the data and determine the formatted indicator dataset.

[0043] Based on the formatted index dataset, calculate the correlation coefficient between image sharpness and each index, and determine the influence weight of image sharpness.

[0044] If the influence weight of the image sharpness reaches a preset threshold, then the subset of indicators is confirmed to be suitable for the final pattern description, and a pattern feature set that can be used for analysis is obtained.

[0045] By combining the pattern feature set with landscape analysis methods, key descriptive information in the multi-scale pattern is extracted, the final pattern description content is determined, and a landscape pattern report is generated.

[0046] Optionally, the step of constructing a multi-scale pattern recognition framework by analyzing the classified element set, determining the spatial relationships of each element within the framework, and generating a second image dataset includes:

[0047] If the classified set of elements does not match the preset threshold, the acquisition parameters are fine-tuned, and the final classification output is obtained by processing the image data again.

[0048] By further analyzing the final classification output, a multi-scale pattern recognition framework is constructed to determine the spatial relationships of various elements within the framework, and the final second image dataset is generated.

[0049] Optionally, after generating a fused patch distribution layer based on the adjusted patch statistics to obtain unified patch distribution information, the method further includes:

[0050] Spatial overlay analysis is performed on the unified patch distribution information to determine whether there are any areas with abnormal distribution.

[0051] If abnormal areas exist, they are locally corrected to obtain the final patch distribution result.

[0052] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0053] Compared with the prior art, the present invention has the following advantages and technical effects:

[0054] This invention addresses the challenges of insufficient image clarity, cloud and fog obstruction, seasonal spectral variation, and fragmented patch distribution in operational scenarios involving massive amounts of image data collected by drones. It constructs a logically interconnected solution. The invention uses a random forest algorithm to filter image data, ensuring that image clarity and cloud / fog obstruction meet threshold requirements. It then combines this with support vector machine classification of landscape elements, optimizes flight altitude and resolution parameters, extracts seasonal spectral features, selects representative temporal images, addresses the fragmented patch distribution, calculates spatial pattern indicators, and finally generates a landscape pattern report and updates the identification framework. Through multi-algorithm collaboration and a dynamic feedback mechanism across multiple scales, this invention significantly improves the accuracy and adaptability of landscape pattern identification, achieving full automation from data filtering to pattern optimization, and providing efficient technical support for precise monitoring in complex terrain environments. Attached Figure Description

[0055] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the first image dataset generation process according to an embodiment of the present invention. Detailed Implementation

[0058] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0059] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment provides a method for forest and grassland landscape pattern recognition based on UAV imagery, including:

[0062] The image data collected by the drone is obtained and filtered based on the sharpness index and the cloud and fog occlusion index to obtain the first image dataset;

[0063] Based on the vegetation type and spatial distribution characteristics in the first image dataset, the optimal acquisition parameters are determined and the landscape elements are classified to obtain the second image dataset.

[0064] From the second image dataset, representative temporal images are selected based on the influence of seasonal changes on spectral features to obtain the third image dataset;

[0065] Based on the landscape elements and vegetation types in the third image dataset, the initial information of the patches is calculated and the fragmented patches are fused to obtain the fused patch distribution data.

[0066] Based on the fused patch distribution data, the shape index and aggregation index are analyzed, and the calculation of the separation and spread indexes is optimized to obtain a multi-scale pattern index set.

[0067] From the multi-scale pattern index set, indicators related to image sharpness are extracted and their accuracy is judged to generate a landscape pattern report.

[0068] As a feasible implementation method, the specific steps include:

[0069] S101. Acquire massive amounts of image data collected by the drone, and analyze the data for indicators such as sharpness, lighting conditions, cloud and fog obstruction, and overlap rate using a random forest algorithm. Determine if the sharpness is higher than a preset threshold and the cloud and fog obstruction is lower than a preset threshold, then retain the image to obtain the first image dataset.

[0070] like Figure 2 As shown, a massive amount of image data collected by a drone is acquired. A random forest algorithm is used to predict the performance of each image, outputting quantitative scores for sharpness, lighting conditions, cloud / fog occlusion, and overlap. If the sharpness score is higher than a preset threshold and the cloud / fog occlusion is lower than a preset threshold, the image data is retained; otherwise, it is discarded. The retained image data forms the first image dataset.

[0071] In the feasible process of UAV image data processing, acquiring massive amounts of image data is the primary task. When performing aerial photography missions, UAVs capture a large number of photos at a high frequency, forming a dataset containing thousands or even hundreds of thousands of images. Due to factors such as shooting environment and equipment performance, the quality of this image data varies greatly, thus requiring algorithmic screening and evaluation. The Random Forest algorithm, as a powerful machine learning method, can comprehensively analyze multiple features of images and output quantitative indicators. This algorithm constructs multiple decision trees to score indicators such as image sharpness, lighting conditions, cloud and fog occlusion, and overlap rate. The scoring range is typically set from 0 to 1, with higher values ​​indicating better quality. This method can effectively reduce human intervention and improve the objectivity and efficiency of the evaluation.

[0072] Specifically, for the sharpness metric evaluation, the Random Forest algorithm analyzes features such as edge sharpness and noise level of the image. Assuming a sharpness threshold of 0.8, if an image has a sharpness score of 0.85, it indicates good detail and meets the retention criteria. For the cloud / fog occlusion metric, the algorithm analyzes the image's grayscale distribution and contrast to determine the degree of occlusion. Assuming a threshold of 0.3, if an image has a cloud / fog occlusion score of 0.2, it indicates minimal occlusion and meets the requirements. This dual threshold screening ensures that the retained images meet the standards for key quality metrics, laying the foundation for subsequent analysis.

[0073] In assessing lighting conditions, the algorithm focuses on the uniformity of brightness distribution in an image. An image taken in the early morning might have a lighting score of 0.6, indicating weak but still acceptable light. Conversely, an image taken at midday might have a score of 0.9, showing excellent lighting conditions. This quantitative analysis helps filter out data suitable for further processing, avoiding detail loss due to insufficient lighting.

[0074] Specifically, the overlap rate metric focuses on the degree of content repetition between adjacent images. During drone aerial photography, overlapping areas are typically planned to ensure complete coverage. If the overlap rate score between two images is 0.7, it indicates good content continuity, which is beneficial for subsequent image stitching and 3D modeling. The algorithm achieves this evaluation through techniques such as feature point matching, ensuring the continuity of the dataset.

[0075] In constructing the first image dataset, only images that pass all the above-mentioned criteria are retained. Assuming the initial dataset contains 10,000 images, after filtering by dual thresholds for sharpness and cloud / fog occlusion, 6,000 images are retained. These images also exhibit high quality in terms of illumination and overlap, ultimately forming a high-quality dataset. This dataset can be used for subsequent tasks such as geographic information analysis and disaster monitoring, significantly improving application effectiveness. This systematic selection method not only ensures data quality but also reduces the consumption of storage and computing resources by redundant data, resulting in higher processing efficiency and more accurate analysis results.

[0076] S102. For the first image dataset, adjust the relationship between flight altitude and image resolution according to vegetation type and spatial distribution characteristics, use support vector machine to classify the landscape elements in the dataset, determine the optimal acquisition parameters under the multi-scale pattern recognition framework, and obtain the second image dataset:

[0077] Based on the first image dataset, the correlation between vegetation type and spatial distribution is analyzed, and the flight altitude parameters are adjusted to obtain optimized altitude configuration data. For the optimized altitude configuration data, combined with image resolution requirements, a suitable acquisition resolution standard is calculated to determine a matching image acquisition scheme. From the matched image acquisition scheme, the distribution information of landscape elements is extracted, and a support vector machine is used to accurately classify the elements, obtaining a classified element set. If the classified element set does not match the preset threshold, the acquisition parameters are fine-tuned, and the image data is processed again to obtain a more accurate classification output. Through further analysis of the classification output, a multi-scale pattern recognition framework is constructed to determine the spatial relationships of each element within the framework, generating the final second image dataset.

[0078] It is feasible to examine the correlation between vegetation type and spatial distribution using statistical analysis and spatial autocorrelation methods based on the first image dataset.

[0079] Specifically, the images in the dataset are first labeled with vegetation types, and then the spatial aggregation index of different vegetation types, such as Moran's I value, is calculated to quantify their spatial distribution patterns.

[0080] In one possible implementation, if it is found that coniferous forests are mainly concentrated in the higher-altitude northern regions, while broad-leaved forests are more distributed in the southern lowlands, this correlation can directly guide the adjustment of the drone's flight altitude.

[0081] The flight altitude was optimized from a uniform 150 meters to 180 meters in the northern region to capture finer details of the coniferous canopy, while maintaining 120 meters in the south to cover a wider range of broadleaf forest distribution, thus creating optimized altitude configuration data. This adjustment helps improve the ability of subsequent imagery to capture the vertical structure of vegetation.

[0082] It should be noted that, for the optimized height configuration data, when combined with image resolution requirements, the appropriate acquisition resolution standard can be determined by calculating the ground sampling distance.

[0083] To achieve a ground resolution of 5 cm / pixel, and considering an optimized altitude of 180 meters, using a common drone camera with a focal length of 24mm, the required flight speed and shutter parameters are calculated using proportional relationships. This allows for the determination of a suitable image acquisition scheme, such as a flight path with 80% lateral overlap and 70% longitudinal overlap. This approach ensures uniform resolution at different altitudes, effectively reducing image quality variations caused by altitude changes.

[0084] In one embodiment, when extracting the distribution information of landscape elements from new images acquired after the matching image acquisition scheme is executed, the locations of elements such as water bodies, woodlands, and bare land can first be identified through edge detection and texture analysis. Then, a support vector machine is used for precise classification.

[0085] By training the feature vectors using radial basis function kernels, the pixel-level classification accuracy is improved to over 92%, resulting in a classified feature set. This classification process significantly improves the accuracy of feature boundaries and avoids misclassification caused by traditional thresholding methods.

[0086] If the classified feature set shows that the proportion of water exceeds a preset threshold of 15%, it indicates that there may be residual clouds or fog or reflection interference caused by improper altitude. In this case, the acquisition parameters are fine-tuned, such as reducing the flight altitude by 10 meters and increasing lateral lighting compensation. By reprocessing the image data, a more accurate classification output is obtained. This iterative fine-tuning can quickly converge to the ideal classification result and reduce the proportion of invalid data.

[0087] Specifically, by further analyzing the classification output, when constructing a multi-scale pattern recognition framework, landscape pattern indices such as patch density and connectivity index can be calculated at multiple scales such as 10 meters, 50 meters and 200 meters to determine the spatial relationships of various elements within the framework.

[0088] The study revealed high fragmentation of forest patches at small scales, while overall connectivity was strong at large scales, reflecting the hierarchical distribution characteristics of vegetation. Based on this framework, high-quality images were selected and integrated to generate the final second image dataset. This dataset offers richer spatial representation, providing more reliable foundational data for subsequent ecological monitoring and change detection, and supporting precise analysis of vegetation dynamics.

[0089] S103. Extract the impact of seasonal variation on spectral features from the second image dataset. By comparing the correspondence between spectral features and pattern indicators at different time phases, determine if the spectral feature variation rate exceeds a preset threshold, then select representative time phase images to obtain the third image dataset.

[0090] Starting with the second image dataset, data extraction methods are used to separate spectral data from different time phases, based on the correlation between seasonal variations and spectral features, to obtain a set of separated temporal spectra. For the separated temporal spectra, the correspondence between spectral features and pattern indicators is analyzed. By comparing the feature changes of each time phase, the rate of variation of feature changes is determined. If the rate of variation exceeds a preset threshold, the temporal spectra are filtered to obtain a representative subset of temporal images. Starting from this representative subset, image data meeting the criteria are extracted, constructing a preliminary image filtering result. Based on the preliminary image filtering result, the influence distribution of spectral features is analyzed, and support vector machines are used to classify the image data, obtaining a set of classified image categories. Using the classified image category set, representative image data from different time phases are integrated to generate the final third image dataset.

[0091] Feasible practice is to analyze the correlation between seasonal variations and spectral characteristics when processing the second image dataset, starting with the spectral reflectance characteristics of vegetation-covered areas. Assuming that differences in chlorophyll content between spring and autumn image data lead to significant variations in spectral reflectance values ​​in the green band, with spring potentially showing higher reflectance values ​​(e.g., 0.45) and autumn potentially decreasing to 0.3, this difference reflects the seasonal changes in vegetation growth, providing a basis for subsequent data extraction. During analysis, comparing spectral curves from different time phases can preliminarily determine which bands are most sensitive to seasonal variations.

[0092] When using data extraction methods to separate spectral data from different time phases, near-infrared data for a specific vegetation type, such as coniferous forest, can be extracted for summer and winter. Summer data may show higher near-infrared reflectance values, indicating vigorous vegetation growth, while winter data shows a significant decrease. This separation method helps to form a clear temporal spectral set, laying the foundation for subsequent analysis.

[0093] It should be noted that during the separation process, bands with significant seasonal differences can be prioritized to improve the data's discriminative power.

[0094] Analysis of the correlation between spectral characteristics and pattern indicators can focus on the relationship between vegetation cover and spectral values. Suppose that in a certain area, the infrared values ​​of phases with high vegetation cover are generally above 0.6, while those with low vegetation cover are below 0.4. By comparing these data, the rate of variation of these characteristics can be derived. If the rate of variation exceeds a preset threshold, such as 15%, the temporal spectral set needs to be filtered to retain the subsets of data with significant changes. This filtering helps to focus on the most representative image data.

[0095] When constructing preliminary image screening results, image sharpness and lighting conditions can be combined as screening criteria. For example, image data with cloud cover exceeding 20% ​​can be excluded, retaining only images with uniform lighting. This approach ensures the accuracy of subsequent classification processing. Analyzing the influence of spectral characteristics on distribution allows observation of the distribution differences of a certain band value in different regions, thus providing a reference for classification.

[0096] When using support vector machines to classify image data, the data can be categorized into three main types based on vegetation: broadleaf forest, coniferous forest, and grassland. By analyzing the differences in spectral characteristics among these types—for example, broadleaf forests generally have lower reflectance values ​​in the red light band than coniferous forests—a relatively accurate classification can be achieved, resulting in a set of classified image categories. This classification method helps to more clearly present the spatial distribution characteristics of vegetation at different time phases.

[0097] When integrating representative image data from different time periods to generate a third image dataset, priority should be given to integrating data from time periods with significant changes in vegetation cover. For example, combining spring and autumn image data can generate a comprehensive dataset that reflects dynamic changes in vegetation. This integration approach can provide more comprehensive data support for subsequent landscape pattern analysis and help reveal the spatial evolution patterns of vegetation in different seasons.

[0098] S104. Based on the landscape elements and vegetation types in the third image dataset, calculate preliminary values ​​for the number and area of ​​patches. Use the random forest algorithm to process the fragmented patch problem caused by high resolution, and obtain the fused patch distribution data:

[0099] For the landscape elements and vegetation types in the third image dataset, initial values ​​for the number and area of ​​patches are calculated to obtain basic patch statistics. If fragmented patches due to high resolution exist in the basic patch statistics, a random forest algorithm is used to classify and merge these fragmented patches to determine optimized patch boundary data. Based on the optimized patch boundary data, the number and area of ​​patches are recalculated to obtain adjusted patch statistics. For the adjusted patch statistics, a merged patch distribution layer is generated to obtain unified patch distribution information. Spatial overlay analysis is performed on the unified patch distribution information to determine if there are any anomalous distribution areas. If anomalous areas exist, they are locally corrected to obtain the final patch distribution results. After obtaining the final patch distribution results, corresponding distribution feature data is generated to determine the spatial distribution pattern of landscape elements and vegetation types.

[0100] Feasible practice is to begin processing landscape features and vegetation types in third-party image datasets by starting with basic patch statistics. Initial calculations of patch number and area can be performed using image segmentation techniques to divide the landscape features in the dataset into multiple independent units. For example, in a forest area, assuming initial statistics show 200 patches with a total area of ​​5000 hectares, some patches, due to detail capture in high-resolution imagery, appear to be only 0.5 hectares in size, making them appear overly fragmented. Such fragmented patches may affect the accuracy of subsequent analysis and therefore require further processing.

[0101] When classifying and merging fragmented patches using the random forest algorithm, judgments can be made based on the patch's texture features and surrounding environmental attributes. Suppose that in the aforementioned forest area, there are 10 patches with an area less than 1 hectare. Algorithm analysis reveals that 8 of these patches share the same vegetation type as adjacent larger patches, such as coniferous forest, and are spatially closely connected. Based on this, these smaller patches are merged into the larger patches, forming new boundary data. This merging method reduces data redundancy and improves the continuity of patch boundaries.

[0102] When recalculating the number and area of ​​patches based on the optimized patch boundary data, it may be found that the number of patches decreases from 200 to 180, while the total area remains unchanged, but the average area of ​​an individual patch increases. This adjusted statistical result better reflects the actual landscape distribution characteristics, providing a more reliable basis for subsequent analysis. Next, when generating a merged patch distribution layer, patches of different vegetation types can be distinguished by color, visually displaying the spatial distribution of elements such as forests and grasslands.

[0103] When conducting spatial overlay analysis to identify areas of distribution anomalies, if a patch density is found to be significantly higher than the surrounding area, and the vegetation types exhibit an unnatural intermingling pattern, such as overly fragmented boundaries between forest and grassland patches, local correction techniques can be used. These techniques, combined with historical imagery data or field verification, can adjust the patch classification in that area to ensure the rationality of the distribution information. This correction helps improve the reliability of the final distribution results.

[0104] When generating the final distribution characteristic data, the spatial pattern of landscape elements can be further analyzed, such as the clustering of forest patches or the dispersion of grassland patches. Taking a forest area as an example, suppose the analysis shows that forest patches are mainly concentrated in the north, accounting for 70% of the total area, while grasslands are scattered in the south. A clear understanding of this spatial distribution pattern helps in understanding the ecological characteristics of landscape elements and provides data support for subsequent planning. Overall, the above method, through gradual optimization and correction, ensures the scientific validity and practicality of the patch distribution results.

[0105] It is worth noting that the method described in this embodiment has good scalability and compatibility. To meet the need for refined extraction of forest internal structural parameters, a forest stand 3D structure reconstruction and analysis module can be further integrated on the basis of the third image dataset (or synchronously acquired UAV multi-view images). Specifically, high-precision 3D point clouds can be generated using structure-of-motion (SfM) technology, and key forest stand parameters can be automatically extracted based on this: for example, canopy closure can be calculated by analyzing the vertical projection of the canopy point cloud; tree height can be obtained by identifying the elevation difference between the apex of a single tree and the ground point; and canopy width can be estimated by analyzing the horizontal distribution of the canopy point cloud of a single tree. These forest stand structural parameters can be used as new attribute dimensions to perform spatial correlation and fusion analysis with the aforementioned landscape type patches, thereby adding a description of "spatial differentiation of forest stand quality" to the final landscape pattern report, realizing a comprehensive and three-dimensional landscape analysis from macroscopic two-dimensional pattern to microscopic three-dimensional structure, from type classification to quality assessment, greatly enhancing the application value of the method in complex scenarios such as forest carbon sink assessment, biomass estimation, and habitat quality evaluation.

[0106] S105. For the fused patch distribution data, analyze the spatial pattern indicators of shape index and clustering degree. Optimize the separation and spread calculations of the data using support vector machines to obtain a multi-scale pattern indicator set:

[0107] Shape indices and clustering degrees are obtained from the fused patch distribution data. The distribution locations of shape indices are determined by dividing the patch distribution data into spatial grids. The clustering degree within each grid is calculated based on the shape index distribution locations. Support vector machines are used to classify and optimize separation and spread, resulting in optimized separation and spread values. If the optimized separation value is higher than the clustering value, high-separation regions are marked; otherwise, low-separation regions are marked. Shape indices and clustering degrees at different scales are obtained by resampling patch distribution data at multiple scales. A multi-scale pattern index set is determined by combining the shape indices and clustering degrees at different scales with the optimized separation and spread values.

[0108] Feasible methods for processing fused patch distribution data include spatial analysis to obtain shape indices and clustering degrees. Shape indices reflect the geometric complexity of patches and are typically related to the irregularity of patch boundaries, while clustering degree indicates the spatial concentration of patches. For example, within a landscape feature distribution area, if a certain type of vegetation patch has a shape index of 2.5, indicating relatively tortuous boundaries, and a clustering degree of 0.8, suggesting a relatively concentrated patch distribution, then it can be preliminarily determined that the vegetation distribution within this area exhibits a certain degree of spatial continuity.

[0109] In one possible implementation, spatial gridding is a crucial method for determining the distribution of shape indices. Assuming the study area is divided into 100 grids, each with an area of ​​1 square kilometer, calculating the shape index of patches within each grid reveals that some grids have indices as high as 3.0, indicating complex patch morphologies in these areas, possibly due to irregular boundaries caused by topographic relief or human disturbance. For aggregation, the distribution density of patches within each grid can be statistically analyzed. If the aggregation degree within a grid is 0.9, it indicates a high concentration of vegetation patches in that area, potentially related to a suitable growth environment.

[0110] When using Support Vector Machines (SVMs) to classify and optimize separation and spread, separation can be understood as the degree of spatial isolation between patches, while spread reflects the spatial expansion trend of patches. Assuming an initial separation value of 0.7 and a spread value of 0.4, after SVM classification optimization, the separation value is adjusted to 0.75 and the spread value to 0.45. If the optimized separation value is higher than the clustering value (e.g., a clustering value of 0.6), the area is marked as a high-separation area, indicating that there may be significant spatial breaks between patches, requiring further analysis of the reasons, such as whether it is affected by road segmentation or land use changes.

[0111] In one possible implementation, multi-scale resampling of patch distribution data can help reveal distribution characteristics at different scales. Assuming a small scale (grid area of ​​0.5 square kilometers) with a shape index of 2.8 and a clustering degree of 0.7, while at a large scale (grid area of ​​5 square kilometers), the shape index decreases to 2.2 and the clustering degree increases to 0.85. This indicates that at larger scales, patch morphologies tend to be more regular and the distribution more concentrated. Combining the optimized separation value of 0.75 and the spread value of 0.45, a multi-scale pattern index set can be constructed to analyze the spatial distribution patterns of landscape elements.

[0112] The marking of highly segregated areas allows for further speculation about their causes, which may be due to natural environmental heterogeneity or human disturbance. Multi-scale analysis reveals that highly segregated areas at small scales may be related to local topography, while those at large scales may be associated with land use planning. This analytical approach helps to understand the formation mechanisms of landscape patterns from different spatial perspectives, providing data support for subsequent vegetation protection or landscape planning, and enhancing the depth of understanding of complex spatial distributions.

[0113] S106. Extract standardized output formats related to image sharpness from the multi-scale pattern index set. If the accuracy of the index set reaches a preset threshold, confirm the final pattern description and obtain a landscape pattern report.

[0114] An initial set of indicator data is obtained from multi-scale pattern data. Preliminary screening based on image sharpness correlation yields a subset of indicators highly correlated with sharpness. This subset is then processed using a standardized format to unify the data structure and dimensions, resulting in a formatted indicator dataset. Based on this dataset, correlation coefficients between image sharpness and each indicator are calculated to obtain correlation assessment results and determine the influence weight of image sharpness. If the influence weight of image sharpness reaches a preset threshold, the subset is deemed suitable for the final pattern description, resulting in a pattern feature set usable for analysis. Using this feature set and landscape analysis methods, key descriptive information from the multi-scale patterns is extracted to determine the final pattern description content. Based on this final pattern description, structured landscape analysis results are generated, completing the mapping between multi-scale patterns and image sharpness, and yielding a complete analytical output.

[0115] The landscape pattern report generated based on the above process is not simply a list of indicators, but a comprehensive result integrating spatial distribution and quantitative analysis. The report systematically elucidates the landscape characteristics of the study area, mainly covering four levels: First, at the landscape composition level, the report clarifies the area proportions and spatial distribution patterns of various elements such as forests, shrublands, and grasslands; second, at the spatial configuration level, it quantifies the morphological complexity, spatial aggregation and dispersion, and landscape connectivity of patches through a series of indicators such as shape index, aggregation degree, separation degree, and spread degree; third, the report conducts multi-scale heterogeneity analysis, revealing how the above pattern characteristics evolve with changes in observation scale and identifying key areas of spatial heterogeneity; finally, the report includes a dedicated section on data quality correlation analysis, clearly indicating the potential impact of image sharpness on the calculation results of each key pattern indicator, thereby providing a self-assessment and explanation of the reliability of the conclusions. The report is presented in a format combining thematic maps, statistical charts, and structured text, and can be directly used to guide practical work such as ecological protection planning and forest management scheme development.

[0116] In practice, when processing multi-scale pattern data, an initial set of indicators can be extracted from the raw data, focusing on indicators related to image sharpness. For example, if multiple indicators such as shape index and distribution density are extracted from the data, and initial screening reveals a high correlation between the shape index and image sharpness, then the shape index can be included in the indicator subset. This screening process can be based on a correlation threshold; for instance, indicators with a correlation coefficient greater than 0.8 are retained.

[0117] Standardization of subsets of indicators can convert data with different dimensions into a unified range, such as normalizing all indicator values ​​to the range of 0 to 1. For example, if the original shape index ranges from 2 to 10 and the distribution density ranges from 50 to 200, standardization adjusts both to the 0-1 range, facilitating subsequent analysis. This approach helps eliminate dimensional differences and improves the fairness of data comparison.

[0118] When calculating the correlation coefficients between image sharpness and various indicators, statistical analysis methods can be used to evaluate the degree of association between each indicator and sharpness. Assuming the correlation coefficient of the shape index is 0.85 and its distribution density is 0.6, and if a preset threshold of 0.7 is set, the shape index is considered to have a significant impact on sharpness. This evaluation helps to clarify which indicators are more representative in pattern description.

[0119] If the weighting of image sharpness reaches a preset threshold, it can be confirmed that the subset of indicators is suitable for the final pattern description. For example, if the weighting value is 0.75, which is higher than the threshold of 0.7, then this subset is used to construct the pattern feature set. This confirmation process ensures the reliability of subsequent analyses.

[0120] When extracting key descriptive information by combining pattern feature sets with landscape analysis methods, attention can be paid to the spatial distribution characteristics of the pattern. For example, if shape indices show a clear concentrated distribution at a certain scale, this may reflect a specific landscape structure and help to more accurately describe pattern characteristics.

[0121] When generating structured landscape analysis results, the correlation mapping between multi-scale patterns and image sharpness can be compiled into a report format. Assuming the analysis results show that areas with high sharpness tend to correspond to patterns with higher shape indices, this provides an important reference for subsequent landscape planning. This mapping relationship helps reveal the impact of image quality on pattern analysis.

[0122] The final analysis output can be presented in the form of visual charts, such as heatmaps showing the spatial correspondence between clarity and pattern indicators. This intuitive presentation facilitates understanding of the analysis results and supports decision-making. Through the above multi-faceted processing and analysis, the comprehensiveness and accuracy of pattern description can be effectively improved, providing more valuable data support for landscape research.

[0123] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0124] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0125] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying forest and grass landscape pattern based on unmanned aerial vehicle image, characterized in that, The method comprises the following steps: acquiring image data collected by a UAV, screening based on a definition index and a cloud and fog blocking index to obtain a first image data set; determining optimal collection parameters and classifying landscape elements according to the vegetation type and spatial distribution characteristics in the first image data set to obtain a second image data set; selecting representative phase images based on the influence of seasonal changes on spectral characteristics from the second image data set to obtain a third image data set; calculating patch initial information and fusing fine patches according to the landscape elements and vegetation types in the third image data set to obtain fused patch distribution data; analyzing shape index and aggregation index and optimizing the calculation of separation index and spread index for the fused patch distribution data to obtain a multi-scale pattern index set; extracting indexes related to image definition from the multi-scale pattern index set and judging the accuracy to generate a landscape pattern report.

2. The method of claim 1, wherein, The method of acquiring image data collected by a UAV, screening based on a definition index and a cloud and fog blocking index to obtain a first image data set comprises the following steps: acquiring image data collected by a UAV; outputting quantification scores of definition index, lighting condition, cloud and fog blocking and overlap rate index by predicting each image data through a random forest algorithm; if the definition index is higher than a preset threshold and the cloud and fog blocking is lower than a preset threshold, retaining the image data, otherwise discarding it; composing a first image data set from the retained image data.

3. The method of claim 1, wherein, The method of determining optimal collection parameters and classifying landscape elements according to the vegetation type and spatial distribution characteristics in the first image data set to obtain a second image data set comprises the following steps: analyzing the correlation between vegetation type and spatial distribution according to the first image data set, adjusting the parameter setting of flight height to obtain optimized height configuration data; calculating suitable collection resolution standards in combination with image resolution requirements for the optimized height configuration data to determine a matched image collection scheme; extracting distribution information of landscape elements from the matched image collection scheme, accurately classifying the landscape elements by support vector machine to obtain a classified element set; constructing a multi-scale pattern recognition framework by analyzing the classified element set, judging the spatial relationship of each element in the framework, and generating a second image data set.

4. The method of claim 1, wherein, The method of selecting representative phase images based on the influence of seasonal changes on spectral characteristics from the second image data set to obtain a third image data set comprises the following steps: extracting spectral data of different phases from the second image data set to obtain a separated phase spectral set; analyzing the corresponding relationship between spectral characteristics and pattern indexes in the phase spectral set, comparing the characteristic changes of each phase data to determine the variation rate value of the characteristic changes; if the variation rate value exceeds a preset threshold, screening the phase spectral set to obtain a representative phase image subset; classifying the image data in the phase image subset by support vector machine to obtain a classified image category set; integrating representative image data of different phases to generate a third image data set.

5. The method of claim 1, wherein, The patch initial information is calculated according to the landscape elements and vegetation types in the third image data set, and the fine patch is fused, to obtain fused patch distribution data, including: The initial values of patch number and patch area are calculated for the landscape elements and vegetation types in the third image data set, to obtain basic patch statistical information; If there are fine patches caused by high resolution in the basic patch statistical information, a random forest algorithm is used to classify and merge the fine patches, to determine the optimized patch boundary data; According to the optimized patch boundary data, the patch number and area information are recalculated to obtain adjusted patch statistical results; According to the adjusted patch statistical results, a fused distribution patch distribution layer is generated to obtain unified patch distribution information.

6. The method of claim 1, wherein, The shape index and aggregation degree index are analyzed for the fused patch distribution data, and the calculation of the separation degree and spread degree index is optimized to obtain a multi-scale pattern index set, including: The shape index and aggregation degree are obtained from the fused patch distribution data; The shape index distribution position is determined by dividing the patch distribution data into spatial grids, and the aggregation degree in each grid is calculated; The support vector machine is used to classify and optimize the separation degree and spread degree to obtain the optimized separation degree value and spread degree value; The shape index and aggregation degree under different scales are obtained by multi-scale resampling the patch distribution data; According to the shape index and aggregation degree under different scales, combined with the optimized separation degree value and spread degree value, the multi-scale pattern index set is determined.

7. The method of claim 1, wherein, From the multi-scale pattern index set, the indexes related to image clarity are extracted and the accuracy is judged to generate a landscape pattern report, including: From the multi-scale pattern index set, the indexes related to image clarity are preliminarily screened to obtain an index subset highly related to clarity; The index subset is processed by data conversion in a standardized format to determine a formatted index data set; According to the formatted index data set, the correlation coefficient between image clarity and each index is calculated to determine the influence weight of image clarity; If the influence weight of image clarity reaches a preset threshold, it is confirmed that the index subset is suitable for final pattern description, and a pattern feature set that can be used for analysis is obtained; Through the pattern feature set, combined with the landscape analysis method, the key description information in the multi-scale pattern is extracted to determine the final pattern description content, and a landscape pattern report is generated.

8. The method of claim 3, wherein, The multi-scale pattern recognition framework is constructed by analyzing the classified element set, the spatial relationship of each element in the framework is judged, and a second image data set is generated, including: If the classified element set does not match the preset threshold, the collection parameters are fine-tuned, and the final classification output is obtained by processing the image data again; The multi-scale pattern recognition framework is constructed by further analyzing the final classification output, the spatial relationship of each element in the framework is judged, and the final second image data set is generated.

9. The method of claim 5, wherein, The method further comprises: performing spatial superposition analysis on the unified patch distribution information to determine whether there is an abnormal distribution area; if there is an abnormal area, performing local correction on the abnormal area to obtain a final patch distribution result.

10. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, The processor implements the method of any one of claims 1-9 when executing the computing program.

Images