Ground object pixel distinguishing data model based on unmanned aerial vehicle spectrum image

Abstract

The invention discloses a ground object pixel distinguishing data model based on an unmanned aerial vehicle spectral image, and the data model comprises the following data connotation: carrying out the preprocessing of the spectral image, calculating at least two vegetation indexes of the preprocessed spectral image, carrying out the secondary data processing based on the obtained vegetation indexes, and obtaining a binary result of the vegetation indexes, the method comprises the following steps: sequentially constructing a confusion matrix, calculating a performance index, assigning a fuzzy number, constructing a fuzzy judgment matrix, calculating a fuzzy weight, and carrying out secondary weighting on a binary result of the vegetation index. Based on the product of the weight coefficient obtained through the secondary weighting operation in the step d and the vegetation index binarization result obtained in the step c and thresholding judgment, crop accurate identification algorithm implementation is carried out. According to the method, performance indexes of different vegetation indexes are combined, a reasonable weight is given to the fuzzy judgment matrix based on the image binarization result, the crop recognition precision can be effectively improved, and the recognition result is more accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of crop identification, and in particular to a ground object pixel differentiation data model based on unmanned aerial vehicle spectral images. Background Art

[0002] Crop identification, especially distinguishing crops from weeds and other background objects, is the key foundation for realizing intelligent agricultural technologies such as precision crop water management, agricultural production monitoring and precision pesticide application. With the continuous development of the concept of precision agriculture, how to efficiently use remote sensing images to obtain crop growth conditions and conduct intelligent management has become an important research direction in the agricultural field. Remote sensing images, especially drone spectral images, can provide detailed crop information due to their high temporal and spatial resolution, thus providing solid data support for crop water management, pest and disease control, and field operation decision-making. Accurate crop identification can not only provide timely and scientific decisions for agricultural production, but also optimize resource allocation, reduce agricultural inputs, and thus maximize crop yields and improve agricultural production efficiency.

[0003] However, in practical applications, due to the influence of soil, weeds and other natural environmental factors, drone spectral images often have background noise, which makes the distinction between crops and weeds extremely complicated. Especially when the crops and weeds are similar in morphology, the environmental conditions are changeable, and the interference of factors such as light and climate is large, the noise problem in the image becomes more significant, which greatly affects the accuracy of crop recognition. In addition, the current methods based on traditional image processing and feature extraction often rely on manually set features, which has great limitations when applied in complex agricultural environments. In recent years, with the progress of deep learning technology, crop recognition methods based on deep learning have gradually become a research hotspot. However, in large-scale agricultural applications, deep learning methods still face challenges such as high computational cost and difficulty in training data labeling. Therefore, in a dynamically changing agricultural environment, how to accurately extract crop information from complex drone spectral images is still a core problem that needs to be solved in current crop recognition technology. Summary of the invention

[0004] In order to solve the above problems, the present invention is specifically a ground object pixel differentiation data model based on UAV spectral imagery, which aims to improve the accuracy of crop identification by effectively identifying and distinguishing crops from weeds and other background objects.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A ground object pixel differentiation data model based on UAV spectral imagery, the data model includes the following data connotations:

[0007] a. Preprocessing the spectral image, wherein the preprocessing steps include image stitching, geographic matching and image cropping;

[0008] b. Calculate at least two vegetation indices of the preprocessed spectral image;

[0009] c. Perform secondary data processing based on the obtained vegetation index to obtain the binary result of the vegetation index;

[0010] d. Perform secondary weighting on the binary results of vegetation index by sequentially constructing confusion matrix, calculating performance index, assigning fuzzy numbers, constructing fuzzy judgment matrix, and calculating fuzzy weights;

[0011] e. An algorithm for accurate crop identification is implemented based on the product of the weight coefficient obtained by the secondary weighted operation in step d and the binarized result of the vegetation index obtained in step c and threshold discrimination.

[0012] Furthermore, the vegetation index in step b includes the normalized difference vegetation index NDVI, the soil adjusted vegetation index SAVI, the enhanced vegetation index EVI, the improved green vegetation index IGVI, the difference vegetation index DVI, the vegetation index ratio VIT, the greenness index GCI, and the difference normalized vegetation index DNDVI.

[0013] Furthermore, the specific process of step c includes finding the optimal threshold and obtaining a binarization result.

[0014] Furthermore, the search for the optimal threshold is specifically to traverse different thresholds and perform binary segmentation on the image pixels, wherein one class includes all pixels greater than the current threshold, and the other class includes all pixels less than the current threshold. After each segmentation, the corresponding inter-class variance is calculated, and the threshold t that maximizes the inter-class variance is selected. * as the optimal segmentation point; the binarization result is obtained specifically as follows: for each pixel in the image, its binarization result is determined according to its corresponding vegetation index value and the optimal threshold value found, and more specifically, if the value corresponding to the vegetation index of a certain pixel is greater than or equal to the optimal threshold value, then the pixel is assigned a value of 1, indicating that it belongs to a certain category; if the vegetation index value of a certain pixel is less than the optimal threshold value, then the pixel is assigned a value of 0, indicating that it belongs to another category.

[0015] Furthermore, the confusion matrix constructed in step d is specifically to obtain ground truth data through field survey and remote sensing acquisition, and the binarization result of each vegetation index is finely compared with the ground truth data at the pixel level to construct a confusion matrix containing four key dimensions.

[0016] Furthermore, the performance indicators calculated in step d specifically include calculating accuracy, precision, recall and F1-score.

[0017] Furthermore, the fuzzy number assignment in step d is specifically to compare the obtained multiple vegetation indices in pairs, quantify the differences between them, and use fuzzy language values ​​to represent the relative importance of each comparison object. Further, according to the size relationship of these fuzzy language values, they are mapped to corresponding triangular fuzzy numbers.

[0018] Furthermore, the fuzzy judgment matrix is ​​constructed in step d as follows: each pair of vegetation indices is filled into the judgment matrix with fuzzy language values. For n vegetation indices, an n×n fuzzy judgment matrix is ​​constructed. Each element in the matrix represents the relative importance of the corresponding vegetation indices. In the matrix, when a vegetation index is compared with itself, its corresponding fuzzy value is set to a specific value representing a completely self-equivalent relationship.

[0019] Furthermore, the fuzzy weights calculated in step d are weighted twice for the binarized results of the vegetation index as follows: first, all fuzzy numbers in each column are added together to obtain the column sum, and then each element is divided by the column sum of the column in which it is located to obtain the normalized fuzzy weight matrix of each column; then, the normalized elements of each row are added together and averaged to obtain a preliminary fuzzy weight, and then the lower bound, median and upper bound of the triangular fuzzy numbers corresponding to the preliminary fuzzy weights of each row are added together and divided by 3; finally, the weights are normalized so that the sum of all weights is equal to 1.

[0020] Furthermore, step e includes calculating weighted scores and adaptive threshold classification, wherein the weighted score calculation is specifically to multiply the weight corresponding to each vegetation index by its binarization result, and then add these products; the adaptive threshold classification is specifically to first calculate the local statistical characteristics of each pixel and its neighborhood pixels, including the local mean and the local standard deviation, and then use the adjustment factor to control the influence of the local variance in the threshold calculation, and obtain the adaptive threshold by multiplying the local mean plus the adjustment factor by the local standard deviation, and determine the classification result of the pixel according to the adaptive threshold. If the weighted score of the pixel is greater than or equal to the adaptive threshold, then the pixel is judged to be a "crop" category; if the weighted score of the pixel is less than the adaptive threshold, then the pixel is judged to be a "background" category.

[0021] Compared with the prior art, the present invention has the following technical advances:

[0022] The present invention fully mines the characteristic information of crops by calculating multiple vegetation indices, such as the normalized difference vegetation index NDVI and the soil adjusted vegetation index SAVI, and then performs data processing on these vegetation indices, determines the optimal threshold, performs binary segmentation on the image pixels according to the threshold, and obtains the binarization result, so as to accurately locate the crop area. Then, through a series of steps such as rigorously constructing a confusion matrix, calculating performance indicators, assigning fuzzy numbers, constructing a fuzzy judgment matrix, and calculating fuzzy weights, the weight of the vegetation index is determined, and the weighted operation is implemented accordingly, which effectively improves the accuracy of crop identification. Finally, based on an in-depth analysis of the local statistical characteristics (including local mean and local standard deviation) of each pixel and its neighboring pixels, the adjustment factor is used to control the influence of the local variance in the threshold calculation, and an adaptive threshold is obtained, thereby accurately distinguishing the crop from the background (such as weeds, soil, etc.) area. The entire crop identification process is efficient and smooth with low resource consumption, which significantly improves the overall effect of crop identification and its practicality in actual agricultural production and other scenarios, and strongly promotes the development and application of crop precision identification technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0024] In the attached picture:

[0025] Figure 1 This is a crop classification and recognition result diagram obtained in Example 2 of the present invention;

[0026] Figure 2 This is the test field distribution map in Example 2 of the present invention;

[0027] Figure 3 This is a schematic diagram of a sensor mounted on a drone in Example 2 of the present invention; DETAILED DESCRIPTION

[0028] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0029] Embodiment 1:

[0030] The present invention provides a ground object pixel differentiation data model based on unmanned aerial vehicle spectral images, and the data model includes the following data connotations:

[0031] a. Preprocessing the remote sensing spectral images collected by the UAV, wherein the preprocessing steps include image stitching, geographic matching and image cropping.

[0032] Among them, image stitching, as the first and most critical step, contains many complex technical details. When performing image acquisition tasks, drones are usually affected by factors such as their flight path, coverage, and the characteristics of the imaging equipment they carry, and they usually obtain multiple spectral images with certain overlapping areas. The first step in image stitching is to accurately extract and analyze the features of each image. Common methods include feature point detection and description based on algorithms such as SIFT (Scale Invariant Feature Transform) and SURF (Speeded Up Robust Features). After completing feature extraction, based on the correspondence between feature points, mathematical models such as the least squares method are used to calculate the geometric transformation matrix between images, thereby achieving accurate alignment of images in spatial position and angle. Subsequently, in the image fusion and stitching stage, the weighted average method is used to fuse the pixel values ​​of the overlapping areas. Specifically, different weights are assigned according to the distance between the pixels and the edge of the image, so as to ensure that the stitched images have a natural transition and seamless docking, and finally form a complete and continuous large image, which shows the overall situation of the study area in an all-round and blind spot-free manner, and provides comprehensive basic data support for subsequent in-depth data analysis.

[0033] After the image stitching is completed, the geographic matching work can be carried out. The geographic matching process involves a series of high-precision technical means and complex data processing processes, aiming to ensure that each image can be accurately positioned in a specific geographic coordinate system and maintain a high degree of consistency with other images. Specifically, the geographic matching process includes the following key steps: First, the initial geographic coordinate information when the image is taken is obtained using the built-in global positioning system (GPS) positioning module of the drone. This initial positioning provides basic data for subsequent precise calibration. However, GPS data alone is often difficult to meet the high-precision requirements, so it is necessary to combine ground control points (GCPs) for further geographic coordinate correction. GCPs usually select landmark features with known precise geographic coordinates in the study area, such as road intersections and building corners. These control points are identified and extracted on the image, and matched and compared with their known geographic coordinates to achieve precise positioning. After the feature point matching is completed, geometric correction models such as affine transformation and polynomial transformation are used to accurately calibrate the image's geographic coordinates. Affine transformation is suitable for processing linear deformations such as rotation, scaling and translation of images, while polynomial transformation can correct more complex nonlinear distortions and improve matching accuracy. In order to optimize the transformation parameters, optimization algorithms such as the least squares method are often used to reduce systematic errors and random errors and improve the overall accuracy and robustness of geographic matching.

[0034] In addition, geographic information system (GIS) technology plays a key role in the geographic matching process. By importing the calibrated image geographic coordinate information into the GIS platform, technicians can perform overlay analysis with existing geographic layers (such as topographic maps, administrative division maps, land use maps, etc.). This process not only helps to verify the geographic matching effect of the image, but also identifies and corrects potential matching deviations through spatial consistency testing and error analysis, ensuring the accuracy and consistency of the image at the geographic information level.

[0035] Finally, the application of image cropping technology. Large-scale images processed by splicing and geographic matching contain a lot of complex information, but not all of them are closely related to the research objectives. Therefore, image cropping accurately extracts the target area that needs to be focused on from the large image according to the pre-set research objectives. Specifically, the cropping process is first based on the boundary coordinate information of the study area, which may come from the previous field measurement data or existing geographic data. Subsequently, the cropping tool in the image processing software is used to set precise cropping parameters, including the coordinates of the upper left and lower right corners of the cropping area, and the cropping shape (such as rectangle, polygon, etc.). For the need for cropping based on specific spectral features, spectral analysis technology must also be combined. For example, by analyzing the characteristic bands of different objects in the spectral curve, the corresponding threshold range is set to screen out pixel areas that meet specific spectral characteristics for cropping.

[0036] b. Calculate at least one vegetation index of the preprocessed spectral image, specifically including:

[0037] (1) Normalized Difference Vegetation Index (NDVI): Calculated by the near-infrared band reflectance and the red band reflectance, the difference between the near-infrared band reflectance and the red band reflectance is calculated, and the difference is divided by the sum of the two to obtain the Normalized Difference Vegetation Index (NDVI). This index can reflect information such as the growth status and coverage of vegetation;

[0038] (2) Soil Adjusted Vegetation Index (SAVI): It also requires the near-infrared band reflectance and the red band reflectance. A soil correction factor L is introduced in the calculation process, with a value of 0.5. The difference between the near-infrared band reflectance and the red band reflectance is calculated. Then, the difference is multiplied by (1 + L), and the result is divided by the sum of the near-infrared band reflectance, the red band reflectance and the soil correction factor to obtain the soil adjusted vegetation index. This index takes into account the influence of soil background and can more accurately reflect vegetation information.

[0039] (3) Enhanced Vegetation Index (EVI): The near-infrared band reflectance, red band reflectance and blue band reflectance are used for calculation, and a series of coefficients are also considered. Among them, the adjustment coefficient is usually taken as 1, the gain factor is usually 2.5, and the two atmospheric correction coefficients are taken as 6 and 7.5 respectively; first calculate the difference between the near-infrared band reflectance and the red band reflectance, then divide it by the near-infrared band reflectance plus the sum of the red band reflectance, blue band reflectance and adjustment coefficient calculated with a specific coefficient, and finally multiply it by the gain factor to obtain the enhanced vegetation index. This index can monitor vegetation changes more sensitively and effectively reduce the interference of atmospheric and other factors on the results;

[0040] (4) Improved Green Vegetation Index (IGVI): Calculated using near-infrared band reflectance and green band reflectance, first calculate the difference between the near-infrared band reflectance and the green band reflectance, and then divide the difference by the sum of the two to get the improved green vegetation index. This index is mainly used to reflect the relevant characteristics of vegetation and evaluate vegetation conditions from another perspective;

[0041] (5) Difference Vegetation Index (DVI): Only the near-infrared band reflectance and the red band reflectance need to be used for calculation. The difference between the near-infrared band reflectance and the red band reflectance is the difference vegetation index. This index intuitively reflects the difference between vegetation and non-vegetation.

[0042] (6) Vegetation index ratio (VIT): The calculation involves the reflectance of the near-infrared band, the short-wave infrared band and the red band. First, the difference between the near-infrared band reflectance and the red band reflectance is calculated, and the ratio of the difference to the sum of the near-infrared band reflectance and the red band reflectance is calculated. Then, the difference between the short-wave infrared band reflectance and the near-infrared band reflectance is calculated and multiplied by the above ratio to obtain the vegetation index ratio. This index can more comprehensively reflect the status of vegetation by integrating the reflectance of multiple bands;

[0043] (7) Greenness Index (GCI): It is calculated based on the near-infrared band reflectance and the green band reflectance. First, the ratio between the near-infrared band reflectance and the green band reflectance is calculated, and then 1 is subtracted from the ratio to get the greenness index. This index is mainly used to measure the greenness of vegetation and reflect the health of vegetation.

[0044] (8) Difference Normalized Difference Vegetation Index (DNDVI): It is necessary to use the near-infrared band reflectance, short-wave infrared band reflectance and red band reflectance to calculate the difference between the near-infrared band reflectance and the red band reflectance, and then ratio the difference with the sum of the two. Then, the difference between the short-wave infrared band reflectance and the red band reflectance is calculated, and the difference is ratioed with the sum of the two. The difference between the two ratios is the difference normalized difference vegetation index. This index can be used to analyze the difference characteristics of vegetation under different band combinations and reflect the changes in vegetation under different backgrounds.

[0045] c. Perform secondary data processing based on the obtained vegetation index to obtain the binary result of the vegetation index, specifically:

[0046] Each vegetation index image is processed based on threshold segmentation to achieve binarization. First, the possible threshold range is traversed and a series of candidate thresholds are systematically selected. For each candidate threshold, the pixels in the image are divided into two categories: those above and those below the threshold. Subsequently, the inter-class variance between these two categories of pixels is calculated. This indicator reflects the degree of discreteness and discrimination of the distribution of the two categories of pixels. The larger the inter-class variance, the better the classification effect and the more obvious the difference between the two categories. By comparing the inter-class variances corresponding to all candidate thresholds, the threshold that maximizes the inter-class variance is selected as the optimal threshold. This optimal threshold is recorded as t * .

[0047] After this step, the binarization result is obtained. For each pixel in the image, its position under the kth vegetation index is (x, y), and the pixel value at this position is expressed by f k (x,y). According to this pixel value and the optimal threshold t * The size relationship determines its final binary result B k (x,y). Specifically, if the f of this pixel k The value of (x,y) is greater than or equal to the optimal threshold t * , then the binarization result of this pixel is B k (x,y) is set to 1; if the f k The value of (x,y) is less than the optimal threshold t * , then the binarization result of this pixel is B k(x,y) is set to 0. In this way, the binarization of the entire image is completed, and the image originally with multiple pixel values ​​is turned into an image with only two pixel values ​​of 0 and 1, thereby achieving accurate division of image pixels, facilitating further analysis and processing of the image, such as extracting areas of interest or performing image recognition. The advantage of this method is that it automatically finds the best segmentation point through the principle of maximizing the global inter-class variance without manually setting the threshold, making the entire binarization process more objective and accurate, and can better serve subsequent research and applications. For example, in vegetation monitoring, it can more clearly distinguish between vegetation areas and non-vegetation areas, or in the agricultural field, it can more accurately divide crop areas from other areas.

[0048] d. The binary results of the vegetation index are weighted twice by sequentially constructing a confusion matrix, calculating performance indicators, assigning fuzzy numbers, constructing a fuzzy judgment matrix, and calculating fuzzy weights.

[0049] (1) Constructing a confusion matrix includes obtaining ground truth data, dividing the data set, and constructing a confusion matrix. The methods for obtaining ground truth data mainly include field surveys and remote sensing acquisition. A professional survey team conducts in-depth field surveys in the target area. Based on multi-dimensional visual information such as crop growth morphology, texture characteristics, color differences, and distribution patterns, combined with GPS positioning equipment, the geographical location information of each crop and background (such as weeds, soil, etc.) is accurately recorded, and their category attributes are marked in detail. In addition, high-resolution ground measurement instruments are used to measure the spectral reflectance of crops and backgrounds in a specific area to obtain field spectral feature data. These data are used to supplement and calibrate remote sensing data. In terms of remote sensing acquisition, multispectral and hyperspectral sensors carried on drones or satellites are used to image the study area multiple times at different time nodes and lighting conditions. By screening, fusing, and analyzing these massive remote sensing image data, spectral feature information related to crops and backgrounds is extracted, and combined with terrain, land use type, and other data in the geographic information system (GIS), the distribution range of different landforms is further determined. The multi-source data obtained from comprehensive field surveys and remote sensing collection are cleaned, sorted and integrated to finally form complete ground truth data.

[0050] In order to improve the accuracy and generalization ability of the model during performance evaluation, it is crucial to divide the data set scientifically and reasonably. First, the entire data set is divided into a training set and a test set according to a specific ratio, of which 70% of the data is designated as a training set and 30% of the data is designated as a test set. Stratified sampling technology is used to ensure that the distribution of samples of each category in the training set is similar to the overall data set based on the proportion of different categories in the data set. In this way, during the training process, the model can fully access representative samples of each category, so as to learn the feature differences and internal laws between different categories more comprehensively.

[0051] In the stage of building the confusion matrix, the binarization result of each vegetation index needs to be finely compared with the ground truth data at the pixel-by-pixel level. The specific steps are as follows: First, read the binarization result image and the corresponding ground truth data image, and accurately align the two in the same geographic coordinate system to ensure that the position of each pixel in the two images corresponds one to one. Then, starting from the upper left corner of the image, scan and compare pixel by pixel by row. For each pixel, check whether its classification value in the binarization result (for example, 0 for background and 1 for crop) is consistent with the true category marked in the ground truth data. If it is consistent, it is recorded as a correct classification; if it is inconsistent, it is recorded as an incorrect classification. In this process, in order to improve the efficiency and accuracy of the comparison, multi-threaded parallel processing technology is used to compare multiple pixels at the same time. Through such meticulous pixel-by-pixel comparison, a confusion matrix with four key dimensions is finally constructed.

[0052] True Positives (TP): During the comparison process, we find out the pixels that are classified as "crop" in the binarization results and are confirmed to be "crop" by the ground truth data. We count the number of these pixels, which is the number of true positives. It reflects the accuracy of the model in correctly classifying crop pixels.

[0053] False Positives (FP): At the same time, we also need to find out the pixels that are classified as "crop" in the binarization result, but in fact should be background according to the ground truth data, and count the number of these pixels, which is the number of false positives. It means that the model mistakenly classifies some pixels that should be background as crop pixels.

[0054] True Negatives (TN): Next, find out the pixels that are classified as "background" in the binarization result and confirmed to be "background" by the ground truth data. The number of these pixels is the number of true negatives. It reflects the accuracy of the model in correctly classifying background pixels.

[0055] False Negative (FN): Finally, identify the pixels that are classified as "background" in the binarization result but should be crops according to the ground truth data, and count the number of these pixels. This is the number of false negatives. It reflects that the model misclassifies some pixels that should be crops as background pixels.

[0056] Table 1 Confusion Matrix Table Based on Binarization Result

[0057] Predict positive class (1) Predict negative class (0) Actual positive class (1) TP FN Actual negative class (0) FP TN

[0058] This confusion matrix not only intuitively reflects the accuracy of the model in pixel classification but also provides a key data basis for subsequent calculation of performance metrics, and is one of the core bases for evaluating the performance of the model.

[0059] (2) Calculate Performance Metrics

[0060] Accuracy: Accuracy is used to measure the proportion of correct classifications overall. The calculation method is to add the number of true positives (TP) and true negatives (TN), and then divide by the sum of the numbers of the four types of pixels: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). This metric allows us to understand the overall accuracy of the model in all pixel classifications.

[0061] Precision: Precision focuses on the proportion of pixels that the model predicts as "crops" and are actually "crops". The calculation method is to divide the number of true positives (TP) by the sum of the numbers of true positives (TP) and false positives (FP). It helps evaluate how accurately the model identifies the actual crop pixels among the pixels predicted as crops.

[0062] Recall: Recall refers to the proportion of pixels that are actually "crops" and are correctly identified as "crops" by the model. The calculation method is to divide the number of true positives (TP) by the sum of the numbers of true positives (TP) and false negatives (FN). This metric can reflect the model's ability to correctly identify actual crop pixels.

[0063] F1-score: F1-score is the harmonic mean of precision and recall, and it can comprehensively measure the precision and integrity of the model. The calculation process is to first calculate the product of precision and recall, then multiply by 2, and then divide by the sum of precision and recall. This metric can more comprehensively evaluate the comprehensive performance of the model in classifying crop pixels.

[0064] (3) Assign Values to Data Using Triangular Fuzzy Numbers Based on Performance Metrics

[0065] When comparing multiple vegetation indices, the multiple vegetation indices obtained are first paired two by two, aiming to identify the differences between them and quantify these differences in the form of specific numerical values.

[0066] Fuzzy language values ​​are used to describe the relative importance of vegetation indices during the comparison process. For example, when comparing two vegetation indices, if they are considered to be of similar importance in the research dimension, they can be described as "equal"; if one vegetation index is slightly better than the other, it can be described as "slightly important"; when the importance of the two vegetation indices differs significantly, they can be described as "very important". These fuzzy language values ​​provide an intuitive and qualitative way to effectively express the ambiguity and uncertainty of the relative importance of vegetation indices.

[0067] According to the size relationship of these fuzzy language values, they are mapped to corresponding triangular fuzzy numbers. Triangular fuzzy numbers are a commonly used way to describe uncertainty. They are represented by three values, namely the minimum possible value, the most likely value, and the maximum possible value. They are denoted as If we consider that two vegetation indices are in an "equal" relationship, then the minimum possible value, the most likely value, and the maximum possible value of the corresponding triangular fuzzy numbers will be relatively close, because their importance is considered to be similar; if it is a "slightly important" relationship, then the minimum possible value, the most likely value, and the maximum possible value will change accordingly according to this relative importance to reflect the difference between them.

[0068] Table 2 Triangular fuzzy number mapping table based on vegetation index performance differences

[0069]

[0070] (4) Constructing the fuzzy judgment matrix

[0071] For each pair of vegetation indices (e.g. and ), and their relative importance is judged based on the experts’ knowledge. The experts, relying on their expertise and rich experience in remote sensing and agriculture, systematically evaluate the relationship between the vegetation indices.

[0072] Then, the corresponding elements in the judgment matrix are filled with the aforementioned fuzzy language values. For example, when experts evaluate the relative importance of vegetation index NDVI and EVI as "NDVI is slightly more important than EVI", this relative importance can be converted into a specific fuzzy number according to the pre-set rules, thereby converting this qualitative judgment into a quantitative fuzzy relationship.

[0073] for vegetation index, we will compare and fill the elements one by one, and finally form an n×n fuzzy judgment matrix In this matrix, each element Vegetation Index and In particular, when When any vegetation index is compared with itself, its corresponding fuzzy number is strictly defined as , which means that in this particular comparison, it is completely equal to itself and there is no difference. In this way, the fuzzy judgment matrix is ​​constructed. A ˜ = [ a ˜ i j ] .

[0074] (5) Calculate fuzzy weight

[0075] First, the fuzzy matrix should be normalized. , add up all the fuzzy numbers in this column and get a column sum The calculation method is to take the first Line Fuzzy number of columns from arrive Add them all up. Then, use each element Divide by the column sum of the column it is in , so we get the normalized fuzzy weight matrix , where each element

[0076] Next, for the normalized fuzzy weight matrix Each line of , calculate a preliminary fuzzy weight The calculation method is to take all the elements in this row from arrive Add them up, then divide by , that is, taking the average value.

[0077] Get the initial fuzzy weight After that, the centroid method is used for defuzzification. The centroid method is to calculate the initial fuzzy weight The lower bound , median and upper bound Add them up and divide by 3 to get a definite value .

[0078] Finally, we need to get the weight Normalize it. The calculation method is to use each weight Divide by all weights After such processing, the final weight satisfies the condition that the sum of all weights is equal to 1, that is, These weights can be used to more reasonably consider the importance of different vegetation indices in tasks such as crop identification.

[0079] e. The algorithm for accurate crop identification is implemented based on the product of the weight coefficient obtained by the secondary weighted operation in step d and the binary result of the vegetation index obtained in step c and the threshold discrimination, specifically:

[0080] (1) Calculate weighted score

[0081] Binarization result B of vegetation index obtained in step c k (x, y) and the weight coefficient w obtained after the second weighting operation in step d k , calculate the weighted score of each pixel (x, y). Specifically, for each discrete pixel in the image space (represented by the coordinates (x, y)), a multi-dimensional weighted superposition operation is required to obtain its corresponding weighted score. The weighted score is an important basis for judging the category of the pixel. Its calculation depends not only on the weight factor, but also on the spatial position of the pixel, the surrounding structure (neighborhood structure) and the relevant features of the pixel itself.

[0082] When the weighted score is higher, it means that more important vegetation indices support that the pixel belongs to the crop category; on the contrary, if the weighted score is lower, it means that more or more critical vegetation indices tend to classify the pixel as the background area.

[0083] The formula for calculating the weighted score for each pixel (x, y) is:

[0084]

[0085] For the kth vegetation index, it has the weight And the binarization result , by multiplying the weight of each vegetation index with its corresponding binary result and accumulating n products, the weighted score of this pixel is finally obtained .

[0086] (2) Adaptive threshold classification

[0087] We need to calculate the local statistical characteristics of each pixel and its surrounding pixels (neighborhood pixels), and adaptively determine the classification threshold of each pixel based on these characteristics. The specific local adaptive threshold calculation formula is: Here is the average value (local mean) of the pixel (x, y) and its neighborhood area. is the value fluctuation size (local standard deviation) of the pixel point (x, y) and its neighborhood area. is an adjustment factor that controls the influence of the local standard deviation in calculating the threshold.

[0088] Based on adaptive threshold To determine the classification result of the pixel point (x, y) , the calculation formula of the classification result is:

[0089]

[0090] If the weighted score of the pixel (x, y) Greater than or equal to the adaptive threshold of this pixel , then this pixel is judged as the "crop" category, represented by 1; if the weighted score of the pixel (x, y) Adaptive threshold smaller than this pixel , then this pixel is judged as the "background" category and is represented by 0.

[0091] Example 2

[0092] like Figure 2 As shown in the figure, the pepper spectral images obtained were identified in Wuwei City, Gansu Province. The total area of ​​the experimental field is about 2 mu, with a size of 68m long and 20.2m wide. The entire experimental area is evenly divided into 46 plots, each with a size of 4.6 m×3.2m. The planting layout of peppers in each plot is 0.4m in row spacing and 0.3m in plant spacing, with a total of 120 pepper seedlings planted.

[0093] like Figure 3 As shown in the figure, two drones (DJI M600 Pro and DJI M300 RTK, DJI Sky City, Shenzhen, China) produced by Shenzhen DJI Innovations Technology Co., Ltd. were used to obtain hyperspectral data of pepper. The drone has a transmission distance of 15 km and is suitable for large-area image acquisition of farmland crops. Its maximum flight time is 55 minutes and the maximum flight speed is 23m / s. The spectral range of the S185 hyperspectral imager is 450-950nm, the spectral resolution is 4 nm, and there are 138 spectral channels in total.

[0094] Based on the vegetation index calculation formula given above, multiple images corresponding to the normalized difference vegetation index NDVI, soil adjusted vegetation index, enhanced vegetation index EVI, improved green vegetation index IGVI, difference vegetation index DVI, vegetation index ratio VIT, greenness index GCI and difference normalized vegetation index DNDVI were obtained.

[0095] Each type of vegetation index image is binarized and the corresponding binarized image is associated. In order to further improve the classification accuracy, a corresponding weight value is assigned to each pixel in each binarized image based on the fuzzy analytic hierarchy process. By integrating the determined weights, all binarization results are weighted to calculate the comprehensive score of each pixel. Figure 1 As shown in the figure, based on these comprehensive scores and adaptive thresholds, the final classification at the pixel level is performed to achieve accurate identification of crop types. The experimental results show that the method of directly binarizing using a single vegetation index is around 77% to 83% in terms of both user accuracy (User's Accuracy) and overall accuracy (Overall Accuracy), while the proposed method exceeds 88%. The specific performance of pepper recognition obtained by the proposed method is shown in the figure. This shows that the method has significant advantages in improving the multi-plant fusion index and classification accuracy, and can effectively identify and distinguish crops from weeds and other background objects, thereby improving the accuracy of crop recognition.

[0096] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of protection of the claims of the present invention.

Claims

1. A ground object pixel differentiation data model based on UAV spectral imagery, characterized in that: The data model includes the following data content: a. Preprocessing the spectral image, wherein the preprocessing steps include image stitching, geographic matching and image cropping; b. Calculate at least two vegetation indices of the preprocessed spectral image; c. Perform secondary data processing based on the obtained vegetation index to obtain the binary result of the vegetation index; d. Perform secondary weighting on the binary results of vegetation index by sequentially constructing confusion matrix, calculating performance index, assigning fuzzy numbers, constructing fuzzy judgment matrix, and calculating fuzzy weights; e. An algorithm for accurate crop identification is implemented based on the product of the weight coefficient obtained by the secondary weighted operation in step d and the binarized result of the vegetation index obtained in step c and threshold discrimination.

2. The ground object pixel differentiation data model based on drone spectral imagery according to claim 1 is characterized in that: The vegetation index in step b includes the normalized difference vegetation index NDVI, the soil adjusted vegetation index SAVI, the enhanced vegetation index EVI, the improved green vegetation index IGVI, the difference vegetation index DVI, the vegetation index ratio VIT, the greenness index GCI, and the difference normalized vegetation index DNDVI.

3. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1 is characterized in that: The specific process of step c includes finding the optimal threshold and obtaining the binarization result.

4. The ground object pixel differentiation data model based on drone spectral imagery according to claim 3 is characterized in that: The specific method of finding the optimal threshold is to traverse different thresholds and perform binary segmentation on the image pixels, wherein one class includes all pixels greater than the current threshold, and the other class includes all pixels less than the current threshold. After each segmentation, the corresponding inter-class variance is calculated, and the threshold t that maximizes the inter-class variance is selected. * as the optimal segmentation point; the binarization result is obtained specifically as follows: for each pixel in the image, its binarization result is determined according to its corresponding vegetation index value and the optimal threshold value found, and more specifically, if the value corresponding to the vegetation index of a certain pixel is greater than or equal to the optimal threshold value, then the pixel is assigned a value of 1, indicating that it belongs to a certain category; if the vegetation index value of a certain pixel is less than the optimal threshold value, then the pixel is assigned a value of 0, indicating that it belongs to another category.

5. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1 is characterized in that: The construction of the confusion matrix in step d is specifically to obtain ground truth data through field surveys and remote sensing acquisition, and to compare the binarized results of each vegetation index with the ground truth data in a pixel-by-pixel manner to construct a confusion matrix containing four key dimensions.

6. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1 is characterized in that: The performance indicators calculated in step d specifically include calculating accuracy, precision, recall and F1-score.

7. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1 is characterized in that: The fuzzy number assignment in step d is specifically to compare the obtained multiple vegetation indices in pairs, quantify the differences between them, and use fuzzy language values ​​to represent the relative importance of each comparison object. Further, according to the size relationship of these fuzzy language values, they are mapped to corresponding triangular fuzzy numbers.

8. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1 is characterized in that: The fuzzy judgment matrix constructed in step d is specifically to fill each pair of vegetation indices with fuzzy language values ​​into the judgment matrix. For n vegetation indices, an n×n fuzzy judgment matrix is ​​constructed. Each element in the matrix represents the relative importance between corresponding vegetation indices. In the matrix, when a vegetation index is compared with itself, its corresponding fuzzy value is set to a specific value representing a complete self-equivalence relationship.

9. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1 is characterized in that: The calculation of the fuzzy weight in step d is to perform secondary weighting on the binarization result of the vegetation index. Specifically, all the fuzzy numbers in each column are first added to obtain the column sum, and then each element is divided by the column sum of the column in which it is located to obtain the normalized fuzzy weight matrix of each column; then the normalized elements of each row are added and averaged to obtain a preliminary fuzzy weight, and then the lower bound, median and upper bound of the triangular fuzzy number corresponding to the preliminary fuzzy weight of each row are added and divided by 3; finally, the weights are normalized so that the sum of all weights is equal to 1.

10. The ground object pixel differentiation data model based on unmanned aerial vehicle spectral image according to claim 1, characterized in that: The step e includes calculating weighted scores and adaptive threshold classification. The weighted score calculation is specifically to multiply the weight corresponding to each vegetation index by its binarization result, and then add these products; the adaptive threshold classification is specifically to first calculate the local statistical characteristics of each pixel and its neighboring pixels, including the local mean and the local standard deviation, and then use the adjustment factor to control the influence of the local variance in the threshold calculation, and obtain the adaptive threshold by adding the local mean and the adjustment factor multiplied by the local standard deviation. The classification result of the pixel is determined according to the adaptive threshold. If the weighted score of the pixel is greater than or equal to the adaptive threshold, then the pixel is determined to be a "crop" category; If the weighted score of a pixel is less than the adaptive threshold, then the pixel is classified as "background".

Images