An unmanned aerial vehicle multi-spectral-based precise identification and missed detection supplement method for guayule

Abstract

The present application relates to a kind of based on unmanned aerial vehicle multispectral precision identification and missed detection of gueldenstaedtia pekinensis and check method, system and medium, the method includes: obtaining unmanned aerial vehicle multispectral image of research area or monitoring area in gueldenstaedtia pekinensis flowering period;The unmanned aerial vehicle multispectral image is pretreated, and ground reflectance orthoimage is generated;Based on normalized difference vegetation index extraction vegetation area, utilize white flower index to vegetation area is initially screened, obtain suspected gueldenstaedtia pekinensis white flower patch;Based on white flower true and false index, eliminate leaf back light confusion source of white color;Based on the pre-constructed liana texture index, remove similar wood misjudgment source of liana;Based on digital elevation model extraction microtopography factor, analyze the relationship between flowering lag and microtopography, guide differentiating supplementary flight, the result is fused with original recognition result, obtain final gueldenstaedtia pekinensis distribution chart.The present application significantly reduces the error recognition rate of gueldenstaedtia pekinensis, missed detection rate, finally realizes the high-precision, stable identification of gueldenstaedtia pekinensis.

Description

Technical Field

[0001] This invention relates to the field of ecological monitoring and remote sensing application technology, and in particular to a method for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging. Background Technology

[0002] Mikania microntha, a noxious invasive weed native to Central and South America, is known as a "plant killer" and poses a serious threat to forest ecosystems and agricultural and forestry production. Traditional surveys of Mikania microntha mainly rely on manual ground surveys or satellite remote sensing interpretation. Manual surveys are inefficient, costly, have limited coverage, and are subject to terrain limitations; while satellite remote sensing has a wide coverage, its spatial resolution is low (usually between meters and ten meters), making it difficult to distinguish Mikania microntha from other understory vegetation, and the revisit period is long, failing to capture the brief window of peak flowering.

[0003] In recent years, UAV multispectral remote sensing technology has become an emerging method for monitoring invasive plants due to its centimeter-level resolution, flexibility, and rich spectral information. Existing techniques typically utilize the high brightness characteristics of the abundant white flowers forming the canopy cover during the peak flowering period of Mikania micrantha (November to February of the following year) to construct a white flower index (such as WFI) for threshold segmentation. However, in complex forest contexts, this method suffers from the following significant drawbacks: 1. Severe confusion with white flowers: The undersides of the leaves of plants such as white crane vine, white catalpa, and fine round vine are silvery white or pinkish white and have extremely high reflectivity under strong light; the white flowers of plants such as white beggar-ticks that bloom at the same time have spectral characteristics that highly overlap with the white flowers of Mikania micrantha, resulting in a misjudgment rate (confusion misclassification rate) of over 11.3%.

[0004] 2. Misidentification of similar species: Woody vines such as *Symplocos buergeriana* and *Symplocos nigra* often coexist with *Mikania micrantha*. Before flowering, both are green and difficult to distinguish in images, leading to a large number of "green" *Mikania micrantha* plants being missed or misidentified as other vegetation.

[0005] 3. Systemic missed detections: Due to insufficient sunlight and high humidity in shady slope and valley areas, the flowering period of Mikania micrantha lags behind that of sunny slopes by approximately 18.5 days. In the main cruise (mid-November) imagery, Mikania micrantha in these areas had not yet bloomed and lacked the characteristic white flowers, resulting in a missed detection rate as high as 8.2%. In addition, Mikania micrantha outside its flowering period completely lacks the characteristic white flowers and cannot be identified by current technology.

[0006] 4. Large fluctuations in accuracy: Due to the influence of climate, the matching degree between the peak flowering period and the flight is different in different years, resulting in large fluctuations in the annual recognition accuracy (86.5%-98.7%), and poor stability in business applications.

[0007] Therefore, there is an urgent need for a precise identification and supplementary detection method for Mikania micrantha that can overcome interference from complex backgrounds, effectively reduce confusion and misclassification rates, and specifically address the problem of missed detection caused by micro-topography. Summary of the Invention

[0008] This invention provides a method for accurate identification and re-detection of Mikania micrantha based on UAV multispectral imaging. It aims to effectively distinguish white confusion sources from similar species by constructing the White Flower Authenticity Index (FSI) and the Vine Texture Index (LTI). By analyzing the quantitative relationship between micro-topographic factors and flowering lag, it guides differentiated re-flight, significantly reducing the false detection rate, and ultimately achieving high-precision and stable identification of Mikania micrantha.

[0009] This invention provides a method for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging, the method comprising the following steps: Step S10: Obtain UAV multispectral images of the study area or monitoring area during the peak flowering period of Mikania micrantha, wherein the images contain reflectance data of at least five bands: blue, green, red, red edge, and near-infrared. Step S20: Preprocess the UAV multispectral image to generate an orthophoto image of the ground surface reflectance; Step S30: Extract vegetation areas based on normalized difference vegetation index, and use the constructed white flower index to perform preliminary screening on the extracted vegetation areas to obtain suspected white flower patches of Mikania micrantha. Step S40: Based on the pre-constructed white flower authenticity index, the texture features of suspected white flower patches of Mikania micrantha are finely distinguished to eliminate white sources of confusion that are reflected on the underside of leaves. Step S50: Based on the pre-constructed vine texture index, perform fine feature differentiation on the finely differentiated patches and remove misidentified sources of similar woody vines; Step S60: Extract micro-topographic factors based on digital elevation model, analyze the relationship between flowering lag and micro-topography, establish a missed detection risk prediction model, and guide differentiated supplementary flight. The micro-topographic factors include slope aspect, slope, altitude and topographic humidity index. Step S70: Repeat steps S30 to S50 on the supplementary flight image, and fuse the results with the original recognition results to obtain the final Mikania distribution map.

[0010] A further technical solution of the present invention is that step S30 includes: Step S301, construct the White Flower Index (WFI), wherein the formula for calculating the White Flower Index (WFI) is: WFI = (Blue + Green + Red) / 3 Wherein, Blue refers to the reflectance value of the blue light band, Green refers to the reflectance value of the green light band, and Red refers to the reflectance value of the red light band; Step S302: Vegetation areas are extracted using the Normalized Difference Vegetation Index (NDVI). The formula for calculating the NDVI is as follows: NDVI = (NIR - Red) / (NIR + Red) NIR refers to the reflectance in the near-infrared band, and Red refers to the reflectance in the red band. Step S303: Patches with a white flower index (WFI) greater than 0.45 and a three-band standard deviation less than 0.05 are designated as white flower patches of Mikania micrantha.

[0011] A further technical solution of the present invention is that step S40 includes: Step S401: To address white confusion and misjudgment of similar species, a white flower authenticity index (FSI) is constructed based on spectral and textural differences to quantify the strength of the red edge effect. The calculation formula is as follows: FSI = (NIR – RedEdge) / (NIR + RedEdge) Wherein, NIR is the reflectance in the near-infrared band, and RedEdge is the reflectance in the red-edge band.

[0012] A further technical solution of the present invention is that step S50 includes: Step S501, construct the vine texture index LTI, the calculation formula is as follows: LTI = GLCM contrast ratio / homogeneity GLCM contrast is a measure of the distance of the numerical distribution in the gray-level co-occurrence matrix from the diagonal in image processing, i.e., the drasticness of local gray-level changes; GLCM contrast The calculation formula is:

[0013] in, It refers to weight; the greater the difference in grayscale, the higher the weight. It is a grayscale pair The probability of occurrence; Homogeneity measures the degree to which the numerical distribution of gray-level co-occurrence matrices in image processing is concentrated near the diagonal, reflecting the uniformity of texture. Its calculation formula is:

[0014] Among them, the denominator The function is that the greater the grayscale difference, the more... The smaller the weight.

[0015] A further technical solution of the present invention is that step S60 includes: Step S601: Mark the missed plaques; Step S602: Extract the multi-temporal phenological characteristics of the missed detection patches; Step S603: Extract the terrain features of the missed patches; Step S604: Using terrain features as the independent variable and the number of days of delayed flowering as the dependent variable, perform multiple regression analysis to establish a model for predicting the risk of missed detection, and see to what extent terrain can predict the number of days of delayed flowering.

[0016] A further technical solution of the present invention is that, after step S50 or step S70, it further includes: step S80, using an object-oriented classification method to extract the spectral features, texture features and geometric features of each object, using training samples to construct a random forest classifier, and outputting four categories: white flowers of Mikania micrantha, other white patches, non-white flower vegetation and non-vegetation.

[0017] A further technical solution of the present invention is that step S80, which employs an object-oriented classification method to extract the spectral features, texture features, and geometric features of each object, includes: Step S801: Using eCognition multi-scale segmentation, adjacent pixels or small objects are merged into large objects from top to bottom until a preset heterogeneity threshold is met. In step S802, after segmentation, eCognition calculates attributes for each object and extracts the spectral features, texture features, and geometric features of each object.

[0018] A further technical solution of the present invention is that the step S80, which involves constructing a random forest classifier using training samples and outputting four categories: white Mikania micrantha flowers, other white patches, non-white flower vegetation, and non-vegetation, includes: Step S803: Use eCognition to segment the entire drone image into irregularly shaped "image objects"; Step S804: Label the category of each training object to form a training set; Step S805: Extract the feature vector of each object: spectral features, texture features, and geometric features, and attach a category label for each feature; Step S806: Input the feature vectors and labels of all training objects into the random forest algorithm to train the random forest model.

[0019] To achieve the above objectives, the present invention also proposes a system for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging. The system includes a memory, a processor, and a program for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging stored in the memory. The program for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging is executed by the processor to perform the steps described in the above embodiments.

[0020] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a program for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging. When the UAV multispectral imaging program is run by a processor, the steps described in the above embodiments are executed.

[0021] The beneficial effects of this invention, based on the multispectral imaging method, system, and medium for accurate identification and missed detection of Mikania micrantha from unmanned aerial vehicles (UAVs), are as follows: This invention effectively distinguishes white confusion sources from similar species by constructing the White Flower Authenticity Index (FSI) and the Vine Texture Index (LTI). By analyzing the quantitative relationship between micro-topographic factors and flowering lag, it guides differentiated replanting, significantly reducing the false recognition rate and false negative rate of Mikania micrantha, and ultimately achieving high-precision and stable identification of Mikania micrantha. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a preferred embodiment of the method for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging in this invention. Figure 2 This is a detailed flowchart of step S30; Figure 3 This is a detailed flowchart of step S60; Figure 4 This is a detailed flowchart of step S80; Figure 5 This is a schematic diagram of the hardware architecture of the Mikania micrantha accurate identification and missed detection supplementation system based on UAV multispectral imaging of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0024] This invention proposes a method for accurate identification and re-detection of Mikania micrantha based on UAV multispectral imaging, such as... Figures 1 to 5 As shown, a preferred embodiment of the method for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging of the present invention includes the following steps: Step S10: Obtain UAV multispectral images of the study area or monitoring area during the peak flowering period of Mikania micrantha, wherein the images contain reflectance data of at least five bands: blue, green, red, red edge, and near-infrared.

[0025] As one implementation scheme, this embodiment can use the DJI Phantom 4 Multispectral, which integrates one visible light RGB (Red, Green, Blue) camera and five multispectral cameras to acquire reflectance data for five bands: Blue (center wavelength 450 nm), Green (center wavelength 560 nm), Red (center wavelength 650 nm), Red Edge (center wavelength 730 nm), and Near-Infrared (NIR, center wavelength 840 nm). The drone is equipped with a Real-Time Kinematic (RTK) module and a light intensity sensor. The flight altitude is 50-80 m, the Ground Sampling Distance (GSD) is 5 cm, the forward overlap is 80%, and the lateral overlap is 70%. Data was collected in November 2023, November 2024, and November 2025, and surface reflectance orthophotos were generated by DJI Terra radiometric correction and orthophoto stitching.

[0026] Step S20: Preprocess the UAV multispectral image to generate a surface reflectance orthophoto image.

[0027] Step S30: Extract vegetation areas based on normalized difference vegetation index, and use the constructed white flower index to perform preliminary screening on the extracted vegetation areas to obtain suspected white flower patches of Mikania micrantha.

[0028] In this embodiment, considering that the white flowers of Mikania micrantha exhibit high reflectance in the blue, green, and red bands and that the values ​​are similar, a White Flower Index (WFI) is constructed based on this.

[0029] Step S30 specifically includes: Step S301, construct the White Flower Index (WFI), wherein the formula for calculating the White Flower Index (WFI) is: WFI = (Blue + Green + Red) / 3 In this system, Blue refers to the reflectance value in the blue light band (center wavelength 450 nm), Green refers to the reflectance value in the green light band (center wavelength 560 nm), and Red refers to the reflectance value in the red light band (center wavelength 650 nm). The White Flower Index (WFI) calculates the average reflectance of each pixel in the blue, green, and red visible light bands. Since the white flowers of Mikania micrantha have high reflectance in all three bands, a higher WFI value indicates that the pixel appears whiter and brighter.

[0030] Step S302: Vegetation areas are extracted using the Normalized Difference Vegetation Index (NDVI). The formula for calculating the NDVI is as follows: NDVI = (NIR - Red) / (NIR + Red) NIR refers to the reflectance in the near-infrared band, and Red refers to the reflectance in the red band.

[0031] Healthy vegetation (or typical vegetation) has a high NIR, low Red, and a relatively large NDVI value. In order to first remove non-vegetated areas such as bare soil, water bodies, and buildings, leaving only vegetated areas for subsequent analysis, this embodiment sets NDVI > 0.45.

[0032] Step S303: Patches with a white flower index (WFI) greater than 0.45 and a three-band standard deviation less than 0.05 are designated as white flower patches of Mikania micrantha.

[0033] "Three bands" refers to the reflectance values ​​of the three visible light bands: Blue, Green, and Red. Standard deviation is a statistical concept used to measure the degree of dispersion among these three values.

[0034] The Red, Green, and Blue values ​​of pure white or neutral gray objects will be very close, with a standard deviation approaching 0. To ensure that the selected bright pixels are "neutral white" rather than other light-colored objects that are bluish, greenish, or yellowish, and thus to more accurately locate the white flowers of Mikania micrantha, this embodiment sets the standard deviation to <0.05.

[0035] Step S40: Based on the pre-constructed white flower authenticity index, the texture features of suspected white flower patches of Mikania micrantha are finely distinguished, and the reflective white color on the underside of the leaves is eliminated as a source of confusion. True white flowers of Mikania micrantha do not have a red edge effect, and the FSI is usually <0.25; although the undersides of leaves of plants such as White Crane Vine, White Catalpa, and Fine Round Vine are white, they contain chlorophyll, and the red edge is significantly enhanced, with an FSI >0.30.

[0036] In this embodiment, a False-white Spectral Index (FSI) is constructed based on spectral and texture differences to address white confusion and misidentification of similar species.

[0037] Step S40 specifically includes: Step S401: To address white confusion and misjudgment of similar species, a white flower authenticity index (FSI) is constructed based on spectral and textural differences to quantify the strength of the red edge effect. The calculation formula is as follows: FSI = (NIR – RedEdge) / (NIR + RedEdge) Wherein, NIR is the reflectance of the near-infrared band, and in this embodiment, the DJI Phantom 4 Multispectral Edition has a center wavelength of 840 nm. RedEdge is the reflectance of the red edge band, with a center wavelength of 730 nm.

[0038] The formula for calculating the White Flower Authenticity Index (FSI) is similar to that for the Normalized Difference Vegetation Index (NDVI), except that the red band is replaced with the red edge band. Its physical meaning is to quantify the strength of the "red edge effect".

[0039] Healthy green plants exhibit a steep gradient of reflectivity between red and near-infrared light, known as the "red edge." Therefore, the reflectivity of the Red Edge is significantly higher than that of red light, resulting in a relatively small (NIR - Red Edge) difference.

[0040] The white petals of Mikania micrantha do not contain chlorophyll and have no photosynthetic activity. The red edge effect is extremely weak or even absent, so the difference between its NIR and Red Edge is relatively larger, and its FSI value is lower.

[0041] Therefore, this embodiment sets the following: the true white flowers of Mikania micrantha have an FSI of <0.25, while the confusing plants (such as Crane Vine, Catalpa bungei, and Slender Round Vine) with white leaves but still containing chlorophyll have an FSI of >0.30, thus effectively distinguishing the two.

[0042] Step S50: Based on the pre-constructed vine texture index, perform fine feature differentiation on the finely differentiated patches and eliminate misidentified sources of similar woody vines.

[0043] This embodiment addresses the issues of white confusion and misidentification of similar species by constructing a Liana Texture Index (LTI) based on spectral and textural differences. Mikania micrantha has small leaves (LTI < 18), while woody vines such as Pteris vittata and Vinegarita have rough leaves (LTI > 25).

[0044] The decision rule in this embodiment is as follows: when WFI>0.45 and FSI<0.25 and LTI<18, it is determined to be white chamomile; when FSI>0.30 or LTI>25, it is directly determined to be a non-target confusing compound.

[0045] Step S50 specifically includes: Step S501, construct the vine texture index LTI, the calculation formula is as follows: LTI = GLCM contrast ratio / homogeneity GLCM contrast is a measure of the distance of the numerical distribution in the gray-level co-occurrence matrix from the diagonal in image processing, i.e., the drasticness of local gray-level changes; GLCM contrast The calculation formula is:

[0046] in, It refers to weight; the greater the difference in grayscale, the higher the weight. It is a grayscale pair The probability of occurrence; Homogeneity measures the degree to which the numerical distribution of gray-level co-occurrence matrices in image processing is concentrated near the diagonal, reflecting the uniformity of texture. Its calculation formula is:

[0047] Among them, the denominator The function is that the greater the grayscale difference, the more... The smaller the weight.

[0048] In this embodiment, both the GLCM contrast and homogeneity parameters originate from texture analysis based on the gray-level co-occurrence matrix (GLCM) in image processing. The core basis of GLCM is the gray-level co-occurrence matrix (GLCM), which is not the original image but a statistical matrix that records the frequency of specific gray-level value pairs in the image.

[0049] The construction process (taking an image object in eCognition as an example) is as follows: 1. Select the analysis band: usually near-infrared or green band.

[0050] 2. Gray-level compression: Compress reflectance values ​​(e.g., 0-1) to a limited number of gray levels (e.g., 64 levels, i.e., 0-63) to reduce computational load.

[0051] 3. Define spatial relationships: Set a direction (e.g., 0° means horizontally adjacent) and a distance (usually 1 pixel).

[0052] 4. Statistical analysis of grayscale pairs: Iterate through all pixels within the object, and for each pair of adjacent pixels that meet the spatial relationship... In the matrix ( OK, (Column) position incremented by 1.

[0053] 5. The final generated square matrix size is: gray level × gray level, and each element in the matrix... This indicates that the grayscale value within the object is... The pixels and grayscale are The number of times (or normalized probability) that pixels are adjacent in a specified direction.

[0054] Simply put, GLCM records how frequently "adjacent relationships of a certain grayscale combination" occur within this object.

[0055] GLCM contrast in this embodiment It measures the distance of the numerical distribution in GLCM from the diagonal, that is, the degree of drastic change in local grayscale.

[0056] Its physical meaning is that if the object's texture is very smooth (like the tiny clusters of Michelia champaca flowers), and the gray levels of adjacent pixels are similar, the values ​​in GLCM are mainly concentrated along the diagonal. ≈ (The grayscale difference is 0 or very small) Lowest.

[0057] If the object has a coarse texture (such as the large leaves of a prickly vine, where veins, shadows, and leaf surfaces alternate), there are many adjacent pairs with large gray-level differences. These values ​​are distributed far from the diagonal of the matrix and are multiplied by a large weight. back, It's high.

[0058] The data in Table 3 below shows that the GLCM contrast ratio of white Mikania micrantha is 12.8±3.5, while that of Aralia elata is 22.4±6.5, which reflects this difference.

[0059] In this embodiment, homogeneity measures the degree to which the numerical distribution in GLCM is concentrated near the diagonal, reflecting the uniformity of the texture.

[0060] Its physical meaning lies in being the opposite of contrast. Smooth areas ( ≈ Components with a grayscale difference close to 0 are fully preserved (with a weight close to 1), resulting in high homogeneity.

[0061] Rough area ( and The components with large differences are significantly suppressed, resulting in low homogeneity.

[0062] The data in Table 3 below show that the homogeneity of white flowers of Mikania micrantha in GLCM is 0.86±0.04, and that of Pteris vittata is 0.61±0.10.

[0063] In this embodiment, the formula for calculating the Fujimoto texture index combines two features in opposite directions into a ratio, which can magnify the differences between different textures many times over.

[0064] Mikania micrantha (small flower clusters): Low contrast ÷ High homogeneity → Very small LTI (<18) Spiny vine, dragon's beard vine (large, rough leaves): High contrast ÷ Low homogeneity → Very high LTI (>25) Compared to using contrast or homogeneity alone, the Fujimoto texture index calculation formula has a larger numerical range, making it easier to set the distinction threshold and more robustly separate the two types of targets in the classifier.

[0065] Step S60: Extract micro-topographic factors based on digital elevation model, analyze the relationship between flowering lag and micro-topography, establish a missed detection risk prediction model, and guide differentiated supplementary flight. The micro-topographic factors include slope aspect, slope, altitude and topographic humidity index.

[0066] For missed detection patches marked on the ground (including late-flowering ravines and non-flowering Mikania micrantha), this embodiment extracts the mean WFI values ​​for three periods: November, December, and January of the following year. Based on a 5 m resolution Digital Elevation Model (DEM), aspect, slope, altitude, and Topographic Wetness Index (TWI) are extracted, and the relationship between the number of days of late-flowering and micro-topographic factors is analyzed through regression analysis.

[0067] A digital elevation model (DEM) is a digital model that uses a numerical array to store ground elevation information. It only records the elevation of exposed ground surfaces and does not include the height of features such as trees and buildings.

[0068] The 5-meter resolution used in this embodiment means that the DEM is composed of countless 5-meter × 5-meter grids, with each grid cell storing the elevation value of that location. This resolution data is typically generated through techniques such as stereo relativistic or radar interferometry, and is not obtained directly from UAV imagery.

[0069] It should be noted that for the topography (slope aspect, valley morphology) of the analysis area, a 5-meter resolution is sufficiently fine, capable of clearly depicting micro-topographic features such as narrow valleys and slope transitions. It has a higher resolution than commonly used 30-meter (such as ASTER GDEM) or 12.5-meter (such as ALOS PALSAR) resolutions, and can capture the "micro-topographic driven" patterns that this invention focuses on.

[0070] The core is to calculate two parameters: Based on this DEM, the two key terrain factors used in this paper can be calculated using software such as ArcGIS and QGIS: Slope: This indicates the direction the grid faces, used to distinguish between north slopes, south slopes, etc.

[0071] Slope: The steepness of the ground.

[0072] Altitude: Values ​​directly taken from the DEM.

[0073] Topographic humidity index ( ): A composite parameter used to characterize the potential for surface water accumulation. Its basic formula is:

[0074] in It is the upstream catchment area per unit contour line length, representing how much water from higher elevations flows to this location; It refers to a local slope (in radians). Simply put, The higher the value, the more likely it is to be the bottom of a gully or a low-lying catchment area, which is more humid and cold year-round.

[0075] Step S60 specifically includes the following steps, the purpose of which is to extract quantitative patterns between "missed patches and micro-topography" from the data.

[0076] Step S601: Mark the missed plaques.

[0077] Through ground-based manual surveys, 42 locations were confirmed where Mikania micrantha actually existed but were missed in the main voyage UAV imagery, covering a total area of ​​98.6 hectares.

[0078] Step S602: Extract multi-temporal phenological characteristics of missed plaques.

[0079] For these missed plaques, we returned to the image and extracted their mean WFI values ​​at different times: Mid-November (Main voyage time): The average WFI value in the missed detection area was only 0.38 (which did not reach the judgment threshold of 0.45, so it was missed).

[0080] From late December to early January of the following year (supplementary flight period): the average WFI value of 35 locations (83.3%) rebounded to 0.51, indicating that the flowers in these areas finally bloomed and reached the standard that could be monitored.

[0081] Therefore, it is possible to calculate the number of days the flowering period lags behind the normal area (which was in full bloom in mid-November) for each missed patch.

[0082] Step S603: Extract the terrain features of the missed patches.

[0083] The boundaries of 42 missed patches were superimposed onto a 5-meter resolution DEM, and the average aspect, average slope, average elevation, and average TWI of each patch were extracted.

[0084] Step S604, Multiple Regression Analysis: Using topographic features as the independent variable and the number of days of delayed flowering as the dependent variable, perform multiple regression analysis to establish a model for predicting the risk of missed detection, and see to what extent topography can predict the number of days of delayed flowering.

[0085] This step involves using "topographic features" as the independent variable X and "flowering lag days" as the dependent variable Y to establish a statistical model and see to what extent topography can predict the lag days.

[0086] The key result presented in this embodiment is: Slope aspect (northward) and TWI together explain 62.4% of the variation in flowering lag days (R²=0.624), which is very explanatory.

[0087] Quantitative analysis: The flowering period on the north slope is delayed by an average of 18.5 days; for every unit increase in TWI, the flowering period is delayed by an average of about 4.2 days.

[0088] In this embodiment, step S60 reveals a clear causal chain, which fully explains the cause of systematic missed detection in Mikania micrantha monitoring: micro-topography → microclimate → phenology → image features.

[0089] Topography determines hydrothermal conditions: north-facing slopes and valleys with high TWI values ​​share the characteristics of weak sunlight, high humidity, and low temperature.

[0090] Hydrothermal conditions determine phenology: the cold and damp environment significantly delays the growth, development, and flowering of Mikania micrantha. While Mikania micrantha in other places is already in full bloom by mid-November, the plants in these valleys may still be in the bud stage or have just started to bloom sporadically, without yet forming a bright white flower cover that is recognizable in images.

[0091] Phenological lag leads to spectral failure: Due to the low proportion of white flowers, the mean WFI of these areas in the main flight images in mid-November was only 0.38, which could not pass the threshold of 0.45 and was ultimately judged as background by the algorithm, resulting in missed detection.

[0092] Step S60 not only explains "why the delay occurred," but also provides a predictive model—in future surveys, based on the slope aspect and TWI map of the study area, it can predict in advance which areas are most likely to experience delayed flowering, thus allowing for targeted drone re-flight. This is the key scientific basis for how this embodiment can ultimately reduce the re-flight coverage to 15%-20%.

[0093] Step S70: Repeat steps S30 to S50 on the supplementary flight image, and fuse the results with the original recognition results to obtain the final Mikania distribution map.

[0094] It should be noted that in this embodiment, the "original recognition result" refers to the white-flowered patch of Mikania micrantha identified after removing the misjudged source of similar woody vines in step S50, and the "result" refers to the white-flowered patch of Mikania micrantha identified after repeating steps S30 to S50 on the supplementary flight image and removing the misjudged source of similar woody vines.

[0095] In this embodiment, after step S50 or step S70, the following step is also included: step S80, which uses an object-oriented classification method to extract the spectral features, texture features and geometric features of each object, uses training samples to construct a random forest classifier, and outputs four categories: white flowers of Mikania micrantha, other white patches, non-white flower vegetation and non-vegetation.

[0096] The step S80, which employs an object-oriented classification method to extract the spectral features, texture features, and geometric features of each object, specifically includes: Step S801: Using eCognition multi-scale segmentation, adjacent pixels or small objects are merged into large objects from top to bottom until a preset heterogeneity threshold is met. In step S802, after segmentation, eCognition calculates attributes for each object and extracts the spectral features, texture features, and geometric features of each object.

[0097] The step S80, which involves constructing a random forest classifier using training samples and outputting four categories—white Mikania micrantha flowers, other white patches, non-white flower vegetation, and non-vegetation—specifically includes: Step S803: Use eCognition to segment the entire drone image into irregularly shaped "image objects"; Step S804: Label the category of each training object to form a training set; Step S805: Extract the feature vector of each object: spectral features, texture features, and geometric features, and attach a category label for each feature; Step S806: Input the feature vectors and labels of all training objects into the random forest algorithm to train the random forest model.

[0098] This embodiment employs eCognition multi-scale segmentation to extract spectral features (mean values ​​of each band, WFI, NDVI, red-edge / near-infrared ratio), texture features (contrast, homogeneity, and entropy based on the Gray-Level Co-occurrence Matrix (GLCM), and geometric features (area and rectangularity) for each object. A Random Forest classifier is constructed using training samples, outputting four categories: white Mikania micrantha flowers, other white patches, non-white flower vegetation, and non-vegetation. The accuracy is evaluated using 81 independent validation samples.

[0099] In this embodiment, eCognition multi-scale segmentation uses the Fractal Network Evolution (FNEA) algorithm, which merges adjacent pixels or small objects into larger objects from the bottom up until a preset "heterogeneity" threshold is met. There are three key parameters: scale parameter, band weight, and homogeneity criterion.

[0100] A larger scale parameter value allows for larger and fewer objects to be merged. Band weights assign different levels of importance to each band. For example, the weights of near-infrared and red edges can be appropriately increased to make the segmentation boundaries more consistent with vegetation differences. The homogeneity criterion is composed of a weighted average of spectral heterogeneity and shape heterogeneity. Shape heterogeneity is further determined by compactness and smoothness. Properly setting shape weights ensures that object edges are neither too fragmented nor overly smooth.

[0101] After segmentation, the entire image is divided into irregularly shaped "image objects," each of which is relatively homogeneous internally. The "object-oriented classification" in this embodiment is based on extracting features from these objects.

[0102] After segmentation, eCognition automatically calculates a large number of attributes for each object. This example uses three main types of features: spectral features, texture features, and geometric features, and their extraction methods are as follows: (1) Spectral characteristics: Spectral features reflect the overall level of radiation / reflection in each band within an object, and the calculation method involves statistically analyzing the values ​​of all pixels contained in the object.

[0103] A. Mean values ​​across different wavelengths, such as Mean Blue, Mean Green, Mean NIR, etc. These are calculated by summing the reflectance values ​​of all pixels within the object in a specific wavelength band and then dividing by the total number of pixels. It represents the overall brightness and color of the object.

[0104] B. White Flower Index (WFI) and Normalized Difference Vegetation Index (NDVI). These two indices need to be created as custom features. There are two equivalent implementation paths: The first method is based on the object's mean, which involves directly substituting the mean values ​​of each band calculated above into the formula.

[0105] WFI = (Mean Blue + Mean Green + Mean Red) / 3 NDVI = (Mean NIR - Mean Red) / (Mean NIR + Mean Red) This calculation method is the most direct and efficient.

[0106] The first method is pixel-by-pixel calculation followed by aggregation: first, calculate the exponential layer (such as an NDVI image) for each pixel, then average the exponential pixel values ​​within the object. The results from both methods are usually highly consistent.

[0107] C. Red edge / Near infrared ratio: The calculation method is to directly divide the mean value of the red-edge band of the object by the mean value of the near-infrared band.

[0108] Red Edge / NIR Ratio = Mean Red Edge / Mean NIR This feature is sensitive to chlorophyll content and vegetation leaf area index. In this embodiment, it is used to help distinguish true white flowers from confused leaves that still have chlorophyll activity.

[0109] (2) Texture features (based on Gray-Level Co-occurrence Matrix (GLCM)) Texture features reflect the spatial arrangement and structural coarseness of gray levels within an object, rather than the spectral values ​​themselves. eCognition's texture analysis is entirely based on the gray-level co-occurrence matrix.

[0110] The calculation of texture features involves two steps: The first step is to generate the GLCM matrix.

[0111] Select a band (in this embodiment, the near-infrared or the brighter green band) and merge the gray values ​​of all pixels into a certain number of levels (e.g., 64 levels).

[0112] For each pair of adjacent pixels within an object, the frequency of their grayscale value combinations in a specified direction (0°, 45°, 90°, 135°) and at a distance (usually 1 pixel) is counted, forming a matrix. This ultimately generates an averaged GLCM that represents the overall texture orientation of the object.

[0113] The second step is to extract second-order statistics from GLCM.

[0114] This embodiment uses three of the most commonly used GLCM measures: contrast ratio ( Homogeneity ) and entropy.

[0115] (1) Contrast ratio ): This measures the severity of local grayscale changes. It's calculated by weighting the elements furthest from the diagonal in a statistical matrix. Larger values ​​indicate coarser textures (like the large leaves of a prickly vine or a gnarled vine), while smaller values ​​indicate smoother textures (like the delicate clusters of flowers in a mitten chrysanthemum). The formula is roughly Σ(ij)² × P(i,j).

[0116] (2) Homogeneity : Measures the uniformity of texture, and is inversely related to contrast.

[0117] The main statistics are for elements near the diagonal in GLCM. The closer the value is to 1, the more uniform and smooth the texture. The formula is roughly ΣP(i,j) / [1+(ij)²].

[0118] (3) Entropy: It measures the disorder or complexity of the texture. Entropy is highest when all element values ​​in the GLCM are relatively average. The texture of Mikania micrantha is relatively regular and its entropy is usually low, while the entropy of mixed vegetation with complex structure is high.

[0119] (3) Geometric features: area and rectangularity. Geometric features are based entirely on the shape of the object and are independent of the internal spectrum. The calculation depends on the object's vector boundaries and raster cell composition.

[0120] 1) Area The area is calculated by first counting the total number of pixels contained in the object, and then multiplying it by the ground area of ​​a single pixel (in this embodiment, GSD is 5cm, and one pixel corresponds to 0.05m × 0.05m = 0.0025m²). This can filter out very small noise points or retain contiguous patches of a certain size.

[0121] 2) Rectangular Fit The method for calculating rectangularity is to first divide the area of ​​the object by the area of ​​its smallest bounding rectangle, and obtain a ratio between 0 and 1.

[0122] Mikania micrantha often spreads irregularly along forest edges and tree canopies, exhibiting a low degree of rectangularity; in contrast, man-made features such as roads and regular farmland have a high degree of rectangularity. This characteristic helps to eliminate false targets with overly regular shapes.

[0123] It should be noted that the training samples in this embodiment can be directly derived from ground-based manual surveys: 136 10 m × 10 m quadrats were established using stratified random sampling for the corresponding month or the following month of the year, taking into account different altitudes, slope aspects, and vegetation types. The measured data included Mikania micrantha cover, flowering status, white flower density, associated plant species and cover, and topographic factors. The quadrats were randomly divided into 55 training samples and 81 independent validation samples. These 136 quadrats are 10 m × 10 m ground-based measured quadrats, each with accurate records of land cover types. 55 of these were used as training samples, and the other 81 were reserved for independent validation, with no overlap between them.

[0124] The specific correspondence is as follows: Ground quadrat record → Determine the true category of the area (white-flowered Mikania micrantha, other white patches, non-white-flowered vegetation, non-vegetation). In UAV imagery, image objects corresponding to the spatial location of the sample plot are assigned the same category labels as those on the ground. Simply put, it involves mapping "ground truth" onto "image objects" to obtain labeled training data.

[0125] In this embodiment, the steps for constructing a random forest classifier using training samples are as follows: Step 1: Multi-scale segmentation to generate image objects.

[0126] eCognition is used to segment the entire drone image into irregularly shaped "image objects" instead of isolated pixels. The ground area covered by each object is typically controlled to be a few square meters.

[0127] Step 2: Label the training objects with categories.

[0128] The ground categories of the 55 training quadrats are assigned one by one to the image objects falling within the quadrats. For example, if a quadrat is confirmed to contain white Mikania micrantha, then the image object(s) covering it will be labeled as "white Mikania micrantha".

[0129] This results in 55 sample plots contributing several labeled objects, forming the training set.

[0130] Step 3: Extract the feature vector of each object.

[0131] Spectral characteristics: mean values ​​of each band, white flower index (WFI), NDVI, red edge / near-infrared ratio.

[0132] Texture features: Contrast, homogeneity, and entropy based on the Gray-Level Co-occurrence Matrix (GLCM).

[0133] Geometric features: area and rectangularity.

[0134] Each object will receive a high-dimensional feature vector, such as [Mean_Blue, Mean_Green, WFI, NDVI, Contrast, Homogeneity, Area, ...], along with its class label.

[0135] Step 4: Train the random forest model.

[0136] The feature vectors and labels of all training objects are input into the random forest algorithm. The core mechanism of random forest is: Bootstrap sampling: N subsets are drawn with replacement from the original training set, and a decision tree is trained for each subset.

[0137] Random feature selection: When splitting a node, each decision tree randomly selects a subset of all features (e.g., sqrt(total number of features)) as candidates to increase the diversity between trees.

[0138] Voting decision: After all decision trees have grown, for a new object, each tree casts its own judgment on the category, and the category with the most votes is the final prediction result.

[0139] After training, the model learns how to distinguish four types of targets based on feature vectors.

[0140] In this embodiment, the random forest classifier does not rely on a single hard rule like "if WFI > 0.45, it's a white flower," but rather on a complex nonlinear combination of features through voting by hundreds or thousands of decision trees. A simplified decision tree might look like this: [Decision Tree Example] If WFI > 0.48: If GLCM_contrast < 15: If NDVI > 0.5: Zewei, chamomile, white flowers otherwise: Then other white patches otherwise: If the red edge / near-infrared ratio is >0.6: Non-white-flowered vegetation otherwise: Then other white patches otherwise: If NDVI > 0.5: Non-white-flowered vegetation otherwise: Non-vegetation A real random forest consists of hundreds of such trees, each with a different structure, and the final classification result is determined by voting from all the trees. This mechanism allows the classifier to integrate multi-dimensional features and avoid the instability of a single threshold.

[0141] The logic for distinguishing the four categories is roughly as follows: 1. Non-vegetation They can be easily separated using features such as NDVI and WFI. Non-vegetation (bare soil, water bodies) typically have an NDVI far below 0.45 and their spectral characteristics are distinctly different from those of vegetation, so they are excluded earliest.

[0142] 2. White flowers of Mikania micrantha vs. other white patches This is the core challenge of classification. The two are highly similar in the visible light spectrum (WFI), and random forests rely heavily on discriminative features: Red edge / near-infrared ratio: Other white patches (such as the back of white catalpa leaves) have a characteristic increase in red edge and near-infrared reflection due to the presence of chlorophyll; white flowers of Mikania micrantha are petals and do not have this effect.

[0143] GLCM texture: Mikania micrantha has small flower clusters with uniform texture (high homogeneity and low contrast); White Crane Vine and other plants have larger leaves or obvious leaf vein shadows and coarser texture.

[0144] Geometric characteristics: Mikania micrantha often spreads in a scattered pattern, while other white ground features may have different shapes.

[0145] The model will take into account these subtle differences when making a judgment.

[0146] 3. Non-white-flowered vegetation This is a broad category of green vegetation, including non-flowering species such as Mikania micrantha, Pteris vittata, and Dianthus chinensis. They have low WFI, high NDVI, and their texture varies depending on the species. Random forests use the mean values ​​and texture features of each band to separate them from the first two categories, and can further distinguish them within each category (this embodiment specifically addresses vine confusion using the LTI index, but the baseline method only distinguishes four categories).

[0147] After the model was trained, predictions were made for the image objects corresponding to 81 independent validation samples. The prediction results were compared with the ground truth, and the user accuracy, mapping accuracy and overall accuracy were calculated to obtain the confusion matrix in Table 1 and the overall accuracy of 90.0%.

[0148] The following section presents the results and analysis of the method for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging.

[0149] 1. Study Area and Data 1.1 Overview of the Study Area The Dapeng Peninsula-Tiantoushan Nature Reserve in Shenzhen (22°27′–22°39′N, 114°17′–114°22′E), covering a total area of ​​166.23 km², has a subtropical marine monsoon climate with an average annual temperature of 22.4 ℃ and annual precipitation of 1948.4 mm. Areas severely infested by *Mikania micrantha*, including low-altitude secondary evergreen broad-leaved forests, orchards, forest edges of plantations, and forest gaps, were selected, with a historically affected area exceeding 400 hectares. Based on UAV multispectral imagery from 2023 to 2025 and 136 ground plots, the confusion mechanism and missed detection patterns were quantitatively analyzed, a quantitative identification standard was proposed, and a robust identification technology system was constructed.

[0150] 1.2 UAV Multispectral Data The DJI P4 Multispectral drone, integrating one visible light RGB (Red, Green, Blue) camera and five multispectral cameras, acquired reflectance data in five bands: Blue (center wavelength 450 nm), Green (center wavelength 560 nm), Red (center wavelength 650 nm), Red Edge (center wavelength 730 nm), and Near-Infrared (NIR, center wavelength 840 nm). The drone was equipped with a Real-Time Kinematic (RTK) module and a light intensity sensor. The flight altitude was 50-80 m, the Ground Sampling Distance (GSD) was 5 cm, with 80% forward overlap and 70% lateral overlap. Data was collected in November 2023, November 2024, and November 2025, and generated as an orthophoto image of the ground reflectance using DJI Terra radiometric correction and orthophoto stitching.

[0151] 1.3 Manual survey quadrats For the corresponding month or the following month of the year, 136 10 m × 10 m quadrats were established using stratified random sampling, taking into account different altitudes, slope aspects, and vegetation types. The coverage, flowering status, white flower density, associated plant species and coverage, and topographic factors of *Mikania micrantha* were measured. The quadrats were randomly divided into 55 training samples and 81 independent validation samples.

[0152] 2. Results and Analysis 2.1 Easily Confused Plant Species and Confusion Mechanisms Long-term field verification and spectral comparison identified 15 easily confused plant species, constituting three complex background types: (1) White confusion sources (6 kinds): White Crane Vine (silver-white silky hairs on the back of the leaves), White Catalpa (white and reflective on the back of the leaves), Fine Round Vine (pinkish-white on the back of the leaves), White-flowered Bidens pilosa, Five-clawed Golden Dragon, False Eagle Claw (yellowish-white flowers). The first three are white confusion caused by the reflection of the back of the leaves or the structure of the young branches and leaves, while the latter three are confusion caused by the same spectrum of true flowers or flowers with similar colors.

[0153] (2) Sources of misidentification of similar species (5 species): Prickly vine, dragon's beard vine, pink-leaved bauhinia, uncaria rhynchophylla, and wild kudzu. All are woody vines that entwine tree crowns over large areas, forming dark green patches in the images, which are easily confused with unflowered Mikania micrantha.

[0154] (3) Habitat-light composite interference sources (4 types): coastal thick vine, narrow sword bean, rootless vine and valley fern, affected by sea surface reflection or weak light environment, aggravate misjudgment.

[0155] 2.2 Accuracy and Confusion Source Statistics A total of 525 white-flowered patches were extracted during the three main cruises, and the evaluation results of 81 validation quadrats are shown in Table 1. The overall user accuracy over the three years was 90.1%, the mapping accuracy was 91.3%, and the overall accuracy was 90.0%. The lowest accuracy (86.5%) was achieved in 2024 due to the warm winter causing confusion and earlier flowering. The highest accuracy (98.7%) was achieved in 2025, as the cruises best matched the peak flowering period.

[0156] Table 1. Classification Confusion Matrix of White Flower Index Method, 2023-2025 (Three-Year Summary)

[0157] The source composition of 100 confusing patches (Table 2) shows that the highest proportion of reflective hairs in *Symplocos buergeriana* (28.0%) was from *Symplocos nigra* (20.0%), followed by young leaves of *Pueraria lobata* (14.0%), reflective underside leaves of *Impatiens balsamina* (11.0%), and *Bidens pilosa* (9.0%), which together accounted for 82.0% of the total. Among the white confusing sources, *Symplocos nigra*, *Symplocos rubra*, and *Symplocos serrata* (with reflective underside leaves) and *Bidens pilosa* (with true flowers) together constituted important sources of false alarms.

[0158] Table 2 Composition of sources of confusing plaques (summary for three years from 2023 to 2025, n=100)

[0159] 2.3 Spectral and Texture Differences and Quantitative Differentiation The spectral and textural characteristics of the main confusing species are shown in Table 3. *Bidens pilosa* and *Gynostemma pentaphyllum* are almost indistinguishable from white-flowered *Mikania micrantha* in the visible light band (difference <0.02). However, the red-edge reflectance (<0.28) and near-infrared reflectance (<0.42) of the undersides of leaves of *Symplocos buergeriana* and *Symplocos nigra* are significantly lower than those of *Mikania micrantha* (p<0.05), and can be effectively separated by FSI. In terms of texture, *Mikania micrantha* has small flower clusters, with a GLCM homogeneity of 0.86 and a contrast ratio of 12.8; the reflective areas on the undersides of leaves of *Symplocos buergeriana* and *Symplocos buergeriana* overlap with the shadows of the veins, and the texture roughness is between that of true flowers and large-leaved vines; the leaves of *Symplocos nigra* and *Symplocos natans* are large with overlapping reflective and shadow areas, and the contrast ratios reach 22.4 and 20.7 respectively, showing extremely significant differences (p<0.01), and LTI can effectively distinguish these similar species.

[0160] Table 3 Comparison of spectral and texture characteristics of major impurities

[0161] Independent validation, combining the decision rules of FSI and LTI, showed that confusion caused by white leaves with reflective undersides, such as *Gnaphalium affine*, *Cathaya argyrophylla*, and *Gnaphalium affine*, as well as misclassification of similar species, such as *Gnaphalium affine* and *Gnaphalium affine*, was effectively suppressed, reducing the confusion misclassification rate from 11.3% to 4.8%. However, confusion with true flowers of similar species, such as *Bidens pilosa*, remains the ultimate bottleneck.

[0162] 2.4 Missed Micro-topographical Features and Their Impact During Non-Flowering Period Over three years, 42 missed patches were identified, covering a total area of ​​98.6 hm², accounting for approximately 7.3% of the actual distribution area. The missed patches consisted of two parts: first, delayed flowering on shady slopes of valleys, concentrated on north-facing slopes at altitudes of 80–280 m and in narrow valleys, with a mean TWI value of 8.7, significantly higher than the 6.2 of normally detected patches (p<0.01); second, *Mikania micrantha* outside its flowering period, which was completely missed due to the lack of white flower characteristics. During the main flight in mid-November, the mean WFI of the missed areas was only 0.38±0.05; during the supplementary flights from late December to early January of the following year, 35 sites (83.3%) had entered their peak flowering period, and the WFI rebounded to 0.51±0.06. Multiple regression analysis showed that slope aspect (north-facing) and TWI could explain 62.4% of the lag days in flowering (R²=0.624). The average lag in flowering on north-facing slopes was 18.5 days, and each additional unit of TWI resulted in a lag of approximately 4.2 days.

[0163] 2.5 Evaluation of Improvement Effectiveness Table 4 shows the accuracy changes of different improvement measures. Using only the white flower threshold method, the overall accuracy was 87.8%, the misclassification rate was 11.3%, and the missed detection rate was 8.2%. After introducing FSI, white confusion such as dead leaves and leaf underside reflection was suppressed, and the accuracy increased to 90.1%. After superimposing LTI, misclassification of similar species such as large-leaved vines was effectively achieved, with an accuracy of 92.0%. Differential re-flight based on TWI and slope aspect models reduced the missed detection area in gullies by 83.1%, and the overall accuracy increased to 94.5%. Combining all measures with manual review, the overall accuracy reached 98.1%, and the annual comprehensive accuracy remained stable at 95.3%–98.5%.

[0164] Table 4. Changes in classification accuracy under different improvement measures

[0165] 3. Discussion Studies show that the white flower index method achieves a single-year user accuracy of 86.5%–88.7%, with the bottleneck being the complex background of winter forests in South China. A total of 15 easily confused plant species were identified. Among them, plants with reflective leaf backs, such as *Gnaphalium affine*, *Cathaya argyrophylla*, and *Gnaphalium affine*, together with true-flowered plants like *Bidens pilosa*, constitute white-flowered confusion. *Gnaphalium affine* and *Gnaphalium affine* are the main subjects of misclassification for similar species. The proposed FSI and LTI indices quantify spectral and textural differences into operable thresholds, requiring only conventional multispectral sensors, significantly reducing the confusion rate under complex background interference. Missed detections are driven by micro-topography; a re-flying model based on slope aspect and TWI can compress the re-flying range to 15%–20%, making it economical and efficient. However, for *Mikania micrantha* outside its flowering period, further exploration is needed to develop an independent identification method combining morphology and texture.

[0166] The proposed three-tiered investigation system of "rapid screening of white flowers - detailed classification of multiple features - ground verification" can reduce manual verification to 10%-15%, and reliably cope with complex backgrounds under the premise of controllable costs, providing an operational technical solution for the annual prevention and control verification of Mikania micrantha in South China.

[0167] 4. Conclusion (1) The traditional white flower index method is constrained by complex backgrounds, and the accuracy in a single year is often less than 90%. The three-year summary user accuracy is 90.1%, the confusion misclassification rate is 11.3%, and the missed detection rate is 8.2%. A total of 15 easily confused plants were identified, which constitute three major categories of background interference: white confusion, misjudgment of similar species, and missed detection in valleys / non-flowering periods.

[0168] (2) The constructed white flower authenticity index FSI and vine texture index LTI can robustly distinguish between white confusion represented by white crane vine, white catalpa and fine round vine and misjudgment of similar species represented by prickly fruit vine and dragon beard vine, reducing the confusion misclassification rate from 11.3% to 4.8%.

[0169] (3) The missed detection in the gully is driven by micro-topography. The prediction model based on slope aspect and TWI (R²=0.624) can guide differentiated re-flight, reducing the missed detection area by 83.1%.

[0170] (4) After combining FSI, LTI and multi-temporal supplementary flights, the automatic classification accuracy reached 94.5%, and after manual verification, it reached 97.1%. The final three-level survey system provides a reliable technical solution for the accurate identification of Mikania micrantha based on UAV multispectral data.

[0171] The beneficial effects of this invention, based on the multispectral detection of Mikania micrantha by unmanned aerial vehicles (UAVs), for accurate identification and remedial detection of missed cases are: This invention effectively distinguishes white confusion sources from similar species by constructing the White Flower Authenticity Index (FSI) and the Vine Texture Index (LTI). By analyzing the quantitative relationship between micro-topographic factors and flowering lag, it guides differentiated replanting, significantly reducing the false recognition rate and false negative rate of Mikania micrantha, and ultimately achieving high-precision and stable identification of Mikania micrantha.

[0172] To achieve the above objectives, this invention also proposes a system for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging, such as... Figure 5As shown, the system includes a processor 1001, a CPU, a network interface 1004, a user interface 1003, a memory 1005, a communication bus 1002, and a UAV-based multispectral MIRV (Mikania micrantha) accurate identification and missed detection supplementation program stored on the processor. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or stable non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0173] Those skilled in the art will understand that Figure 5 The system structure shown does not constitute a limitation on the system and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0174] like Figure 5 As shown, the memory 1005, which serves as a computer storage medium, may include an operating device, a network communication module, a user interface module, and a program for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging.

[0175] exist Figure 5 In the system shown, the network interface 1004 is mainly used to connect to the network server and communicate with the network server; the user interface 1003 is mainly used to interact with the user terminal and receive user input commands; and the processor 1001 can be used to call the UAV multispectral-based Mikania micrantha accurate identification and missed detection supplementation program stored in the memory 1005.

[0176] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a program for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging. When the UAV multispectral imaging program is run by a processor, it executes the steps of the method described above, which will not be repeated here.

[0177] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for accurate identification and supplementary detection of Mikania micrantha based on UAV multispectral imaging, characterized in that, The method includes the following steps: Step S10: Obtain UAV multispectral images of the study area or monitoring area during the peak flowering period of Mikania micrantha, wherein the images contain reflectance data of at least five bands: blue, green, red, red edge, and near-infrared. Step S20: Preprocess the UAV multispectral image to generate an orthophoto image of the ground surface reflectance; Step S30: Extract vegetation areas based on normalized difference vegetation index, and use the constructed white flower index to perform preliminary screening on the extracted vegetation areas to obtain suspected white flower patches of Mikania micrantha. Step S40: Based on the pre-constructed white flower authenticity index, the texture features of suspected white flower patches of Mikania micrantha are finely distinguished to eliminate white sources of confusion that are reflected on the underside of leaves. Step S50: Based on the pre-constructed vine texture index, perform fine feature differentiation on the finely differentiated patches and remove misidentified sources of similar woody vines; Step S60: Extract micro-topographic factors based on digital elevation model, analyze the relationship between flowering lag and micro-topography, establish a missed detection risk prediction model, and guide differentiated supplementary flight. The micro-topographic factors include slope aspect, slope, altitude and topographic humidity index. Step S70: Repeat steps S30 to S50 on the supplementary flight image, and fuse the results with the original recognition results to obtain the final Mikania distribution map.

2. The method for accurate identification and missed detection supplementation of Mikania micrantha based on UAV multispectral imaging according to claim 1, characterized in that, Step S30 includes: Step S301, construct the White Flower Index (WFI), wherein the formula for calculating the White Flower Index (WFI) is: WFI = (Blue + Green + Red) / 3 Wherein, Blue refers to the reflectance value of the blue light band, Green refers to the reflectance value of the green light band, and Red refers to the reflectance value of the red light band; Step S302: Vegetation areas are extracted using the Normalized Difference Vegetation Index (NDVI). The formula for calculating the NDVI is as follows: NDVI = (NIR - Red) / (NIR + Red) NIR refers to the reflectance in the near-infrared band, and Red refers to the reflectance in the red band. Step S303: Patches with a white flower index (WFI) greater than 0.45 and a three-band standard deviation less than 0.05 are designated as white flower patches of Mikania micrantha.

3. The method for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging according to claim 2, characterized in that, Step S40 includes: Step S401: To address white confusion and misjudgment of similar species, a white flower authenticity index (FSI) is constructed based on spectral and textural differences to quantify the strength of the red edge effect. The calculation formula is as follows: FSI = (NIR – RedEdge) / (NIR + RedEdge) Wherein, NIR is the reflectance in the near-infrared band, and RedEdge is the reflectance in the red-edge band.

4. The method for accurate identification and missed detection supplementation of Mikania micrantha based on UAV multispectral imaging according to claim 3, characterized in that, Step S50 includes: Step S501, construct the vine texture index LTI, the calculation formula is as follows: LTI = GLCM contrast ratio / homogeneity GLCM contrast is a measure of the distance of the numerical distribution in the gray-level co-occurrence matrix from the diagonal in image processing, i.e., the drasticness of local gray-level changes; GLCM contrast The calculation formula is: in, It refers to weight; the greater the difference in grayscale, the higher the weight. It is a grayscale pair The probability of occurrence; Homogeneity measures the degree to which the numerical distribution of gray-level co-occurrence matrices in image processing is concentrated near the diagonal, reflecting the uniformity of texture. Its calculation formula is: Among them, the denominator The function is that the greater the grayscale difference, the more... The smaller the weight.

5. The method for accurate identification and missed detection supplementation of Mikania micrantha based on UAV multispectral imaging according to claim 4, characterized in that, Step S60 includes: Step S601: Mark the missed plaques; Step S602: Extract the multi-temporal phenological characteristics of the missed detection patches; Step S603: Extract the terrain features of the missed patches; Step S604: Using terrain features as the independent variable and the number of days of delayed flowering as the dependent variable, perform multiple regression analysis to establish a model for predicting the risk of missed detection, and see to what extent terrain can predict the number of days of delayed flowering.

6. The method for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging according to any one of claims 1 to 5, characterized in that, After step S50 or step S70, the method further includes step S80, which uses an object-oriented classification method to extract the spectral features, texture features and geometric features of each object, uses training samples to construct a random forest classifier, and outputs four categories: white flowers of Mikania micrantha, other white patches, non-white flower vegetation and non-vegetation.

7. The method for accurate identification and missed detection supplementation of Mikania micrantha based on UAV multispectral imaging according to claim 6, characterized in that, The step S80, which employs an object-oriented classification method to extract the spectral features, texture features, and geometric features of each object, includes the following steps: Step S801: Using eCognition multi-scale segmentation, adjacent pixels or small objects are merged into large objects from top to bottom until a preset heterogeneity threshold is met. In step S802, after segmentation, eCognition calculates attributes for each object and extracts the spectral features, texture features, and geometric features of each object.

8. The method for accurate identification and missed detection supplementation of Mikania micrantha based on UAV multispectral imaging according to claim 7, characterized in that, The step S80, which involves constructing a random forest classifier using training samples and outputting four categories—white Mikania micrantha flowers, other white patches, non-white flower vegetation, and non-vegetation—includes the following: Step S803: Use eCognition to segment the entire drone image into irregularly shaped "image objects"; Step S804: Label the category of each training object to form a training set; Step S805: Extract the feature vector of each object: spectral features, texture features, and geometric features, and attach a category label for each feature; Step S806: Input the feature vectors and labels of all training objects into the random forest algorithm to train the random forest model.

9. A system for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging, characterized in that, The system includes a memory, a processor, and a UAV-based multispectral Mikania micrantha precise identification and missed detection supplementation program stored in the memory. The UAV-based multispectral Mikania micrantha precise identification and missed detection supplementation program is executed by the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for accurate identification and missed detection of Mikania micrantha based on UAV multispectral imaging. When the UAV multispectral imaging program is run by a processor, it executes the steps of the method as described in any one of claims 1 to 8.

Images