CN117152645A - Wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics - Google Patents
Wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics Download PDFInfo
- Publication number
- CN117152645A CN117152645A CN202311306532.3A CN202311306532A CN117152645A CN 117152645 A CN117152645 A CN 117152645A CN 202311306532 A CN202311306532 A CN 202311306532A CN 117152645 A CN117152645 A CN 117152645A
- Authority
- CN
- China
- Prior art keywords
- index
- wheat
- feature
- feature set
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/17—Terrestrial scenes taken from planes or by drones
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/54—Extraction of image or video features relating to texture
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/188—Vegetation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Multimedia (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The application provides a wheat rust monitoring method based on the multispectral image depth characteristics of an unmanned aerial vehicle, which comprises the steps of collecting measured data on the ground; acquiring remote sensing image data of wheat; calculating a spectrum index to obtain a spectrum index image feature set; carrying out correlation feature sequencing and first feature screening on the spectrum index image feature set to obtain an optimal spectrum feature set; after carrying out texture feature calculation on the wheat remote sensing image data set and the images in the optimal spectrum feature set, combining the wheat remote sensing image data set and the images in the optimal spectrum feature set, and carrying out feature sequencing and second feature screening to obtain an optimal combined feature set; extracting depth features corresponding to the wheat remote sensing image data set and the images in the optimal combination feature set, combining the extracted depth features with the optimal combination feature set, performing feature ordering and third feature screening, and constructing a plurality of correlation-regression models by using the screened optimal depth feature set; and selecting a correlation-regression model with highest precision to carry out inversion mapping on the wheat region to be detected, and obtaining a wheat rust monitoring result.
Description
Technical Field
The application belongs to the technical field of wheat rust monitoring, and particularly discloses a wheat rust monitoring method based on multispectral image depth characteristics of an unmanned aerial vehicle.
Description of the background
Wheat is one of the main grain crops in China, has important roles in national grain safety and national economy development, and wheat rust is one of the main diseases and insect pests affecting the wheat production, and has the characteristics of wide influence range, strong evolution, high transmission speed and the like, and the crop yield can be reduced by more than 50% when serious. The traditional wheat rust monitoring method mainly depends on manual field investigation, and the method is time-consuming and labor-consuming, and is difficult to quickly and accurately acquire the disaster-affected area range and the disaster-affected degree. Therefore, a rapid, large-scale, nondestructive wheat rust monitoring method capable of guaranteeing the identification accuracy is needed, so that relevant measures can be timely taken, and crop losses are reduced. The remote sensing technology can realize the monitoring and identification of crop growth parameters and physiological parameters by analyzing electromagnetic wave information reflected or radiated by crops, and has become one of the main modes for monitoring crop growth vigor, plant diseases and insect pests and predicting yield gradually. The satellite remote sensing can acquire image data of a large-range area, so that inversion of crop growth parameters and monitoring of plant diseases and insect pests are realized, but problems of long access period, low resolution, mixed pixels and the like are caused, and timely and effective monitoring results of plant diseases and insect pests are difficult to acquire. Aiming at the problems, a novel wheat rust monitoring method based on the multispectral image depth characteristics of the unmanned aerial vehicle is researched and designed, and the method is necessary to overcome the problems existing in the existing wheat rust monitoring.
Disclosure of Invention
In order to solve the problems that in the prior art, a great deal of manpower and time are required for monitoring the wheat rust through manual field investigation, the disaster area range and the disaster degree are difficult to acquire rapidly and accurately, a long access period is required for monitoring the wheat rust through satellite remote sensing, the resolution is low, and a monitoring result cannot be acquired timely, the application provides a wheat rust monitoring method based on the multispectral image depth characteristics of an unmanned plane.
The application provides a wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics, which comprises the following steps:
s1, collecting measured data on the ground, randomly selecting wheat at different positions in a research area as sampling parties, collecting corresponding coordinates of each sampling party, and respectively calculating the leaf disease ratio of the wheat in each sampling party;
s2, acquiring multispectral images of different wave bands in a research area through a multispectral unmanned aerial vehicle, and acquiring a wheat remote sensing image dataset;
s3, splicing and preprocessing the multispectral images in the wheat remote sensing image dataset acquired in the step S2 to obtain a wheat spectral image, and calculating the spectral index of the wheat spectral image to obtain a sampling side spectral index image feature set;
s4, carrying out correlation feature sequencing and first feature screening on the spectrum index image feature set obtained in the step S3 to obtain an optimal spectrum feature set;
s5, performing texture feature calculation on the wheat remote sensing image dataset obtained in the step S2 and the image in the optimal spectrum feature set obtained in the step S4 to obtain a texture feature set, combining the texture feature set and the optimal spectrum feature set to generate a first candidate feature set, and performing feature sequencing and second feature screening on the first candidate feature set to obtain an optimal combined feature set;
s6, carrying out depth feature extraction on the wheat remote sensing image data set obtained in the step S2 and the images in the optimal combination feature set obtained in the step S5 through a pre-trained deep learning network to obtain a depth feature set, combining the depth feature set and the optimal combination feature set to generate a second candidate feature set, carrying out feature sequencing and third feature screening on the second candidate feature set to obtain an optimal depth feature set, and constructing a plurality of correlation-regression models by utilizing the optimal depth feature set;
s7, comparing the accuracy of the correlation-regression models constructed in the step S6 through the disease leaf ratio of the wheat in each sampling party obtained in the step S1, selecting the correlation-regression model with the highest accuracy, and carrying out inversion drawing on the wheat region to be detected to obtain a wheat rust monitoring result.
According to the wheat rust monitoring method based on the multispectral image depth characteristics of the unmanned aerial vehicle according to some embodiments of the application, in the step S1, through a centimeter-level differential positioning system provided by a thousand-finding position, sampling sides are randomly selected in a research area, are all 10cm multiplied by 10cm sampling sides, central position coordinates of the sampling sides are marked and recorded, the number of diseased leaves and the total number of wheat leaves in the sampling sides are obtained, the disease She Bilv RD of the wheat in each sampling side is calculated, and the disease leaf ratio RD is shown in formula (1):
RD＝d/s (1)
wherein RD is disease She Bilv, d is the number of diseased leaves in the sampling side, and s is the total number of wheat leaves in the sampling side.
According to some embodiments of the present application, in the step S2, when multispectral images of different wavebands in a research area are acquired by the multispectral unmanned aerial vehicle, the multispectral unmanned aerial vehicle flies in sunny weather, and the local time is 11:00-13: and 00, finishing the flight, wherein the spatial resolution of the multispectral image is not lower than 1.6cm per pixel, the course overlap ratio of the multispectral unmanned aerial vehicle is not lower than 75%, and the side overlap ratio of the multispectral unmanned aerial vehicle is not lower than 75%.
According to some embodiments of the application, in the step S3, the preprocessing includes: geographic registration, geometric correction, reflectivity correction, radiation correction, and atmospheric correction;
the spectral indexes comprise a calculated spectral index and an empirical spectral index, the calculated spectral index comprises a difference spectral index, a ratio spectral index and a normalized spectral index, and the calculation of the difference spectral index is shown in a formula (2):
DSI i,j ＝R i -R j (2)
wherein DSI i,j The spectrum index of the difference between the i wave band and the j wave band is represented, i represents any wave band of blue, green, red edge or near infrared, j represents any wave band of blue, green, red edge or near infrared, R i Representing the reflectivity of any one of blue, green, red edge or near infrared, R j Representing the reflectance of any one of the blue, green, red-edge or near-infrared bands;
the calculation of the ratio spectrum index is shown in a formula (3):
wherein RSI is i,j The ratio spectrum index of the i wave band and the j wave band is represented;
the normalized spectral index is calculated as shown in formula (4):
wherein, NDSI i,j Normalized spectrum finger representing i-band and j-bandA number;
the empirical spectrum index comprises anthocyanin reflection index, optimized soil adjustment vegetation index, triangular vegetation index, wide dynamic range vegetation index, soil adjustment vegetation index, red edge chlorophyll index, modified triangular vegetation index, modified nonlinear vegetation index, green leaf index and green chlorophyll vegetation index;
the anthocyanin reflection index is calculated as shown in a formula (5):
wherein ARI represents anthocyanin reflectance index, R G Representing the reflectance of the green band, R RE Indicating the red-side band reflectivity;
the calculation of the optimized soil adjustment vegetation index is shown in a formula (6):
wherein OSAVI represents optimizing soil adjustment vegetation index, R NIR Indicating the reflectivity of the near infrared band;
the calculation of the triangular vegetation index is shown in formula (7):
wherein TVI represents a triangular vegetation index;
the calculation of the wide dynamic range vegetation index is shown in formula (8):
wherein WDRVI represents a wide dynamic range vegetation index, R R Indicating the red band reflectivity;
the calculation of the soil conditioning vegetation index is shown in formula (9):
wherein SAVI represents a soil conditioning vegetation index;
the calculation of the chlorophyll index of the red edge is shown in the formula (10):
wherein, CIre represents the red edge chlorophyll index;
the calculation of the modified triangular vegetation index is shown in formula (11):
wherein, MTVI 2 Representing a modified triangular vegetation index;
the calculation of the modified nonlinear vegetation index is shown in formula (12):
wherein MNLI represents the modified nonlinear vegetation index;
the green leaf index is calculated as shown in equation (13):
wherein GLI represents green leaf index, R B Representing blue band reflectivity;
the green chlorophyll vegetation index is calculated as shown in formula (14):
wherein CI is G And represents the green chlorophyll vegetation index.
According to the wheat rust monitoring method based on the depth characteristics of the multispectral images of the unmanned aerial vehicle, in the step S4, the relevance ranking comprises pearson correlation coefficient, gray relevance and projection variable importance; the first feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of correlation feature sequencing, calculating AIC values of the regression fitting models by adopting a red pool information criterion, and outputting a spectrum index contained in the regression fitting model with the minimum AIC value as an optimal spectrum feature set, wherein the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
According to some embodiments of the present application, in the step S5, texture feature extraction is performed on the multispectral image in the wheat remote sensing image dataset obtained in the step S2 and the multispectral image and the spectral index image in the optimal spectral feature set obtained in the step S4 by using a probability statistics method of a gray level co-occurrence matrix, where the texture features include mean, cooperativity, dissimilarity, information entropy, second moment, correlation, contrast and variance.
According to some embodiments of the present application, in the step S5, the relevance ranking includes pearson correlation coefficient, gray relevance and projection variable importance; the second feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of correlation feature sequencing, calculating AIC values of the regression fitting models by adopting a red pool information criterion, and outputting spectral indexes and texture index features contained in the regression fitting model with the minimum AIC values as an optimal combined feature set, wherein the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
According to some embodiments of the present application, the deep learning network in the step S6 is a res net50 network, a skip link is added between convolution layers of the VGG19 network to obtain the res net50 network, the res net50 network includes 49 convolution layers and 1 full-link layer, and the step of extracting the deep feature of the image through the pre-trained res net50 network includes the following steps:
step a, extracting the wheat remote sensing image dataset obtained in the step S2 and an image with 15×15 pixels with a sampling area as a center in the optimal combined feature set obtained in the step S5, and expanding the influence into 224×224 pixels to obtain a first expanded image;
step b, expanding the first expanded image from 1 dimension to 3 dimensions by constructing a 3-layer convolution kernel taking a Gaussian kernel as a kernel function to obtain a second expanded image;
and c, inputting the second expanded image into the ResNet50 network to obtain 2048 depth features.
According to the wheat rust monitoring method based on the depth features of the multispectral images of the unmanned aerial vehicle, in the step S6, the relevance ranking comprises pearson relevance coefficients, gray relevance degrees and projection variable importance, the third feature screening comprises the steps of constructing a plurality of regression fit models based on a machine learning model according to various results of the relevance feature ranking, calculating AIC values of the regression fit models by adopting a red pool information criterion, and outputting spectral indexes, texture indexes and depth features contained in the regression fit model with the minimum AIC values as an optimal depth feature set, wherein the machine learning model comprises a partial least squares regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
According to the wheat rust monitoring method based on the depth features of the multispectral image of the unmanned aerial vehicle according to some embodiments of the present application, in the step S6, the plurality of correlation-regression models constructed by using the optimal depth feature set include a pearson correlation coefficient-partial least squares regression model, a gray correlation-partial least squares regression model, a variable projection importance-partial least squares regression model, a pearson correlation coefficient-back propagation neural network model, a gray correlation-back propagation neural network model, a variable projection importance-back propagation neural network model, a pearson correlation coefficient-random forest model, a gray correlation-random forest model, a variable projection importance-random forest model, a pearson correlation coefficient-extreme learning model, a gray correlation-extreme learning model, and a variable projection importance-extreme learning model.
According to the wheat rust monitoring method based on the multispectral image depth features of the unmanned aerial vehicle, the feature set for monitoring the wheat rust is constructed by utilizing the method of combining the spectral index, the texture features and the depth features, potential effective depth feature information is mined by utilizing a deep learning network, the accuracy of monitoring the wheat rust by the multispectral image of the unmanned aerial vehicle is improved, the problems of feature redundancy and overfitting which possibly occur are avoided as far as possible by using a plurality of feature screening methods, the model accuracy is improved, the complexity of a monitoring model is reduced, and the operation efficiency of the monitoring model is improved.
Drawings
Fig. 1 is a flow chart of a wheat rust monitoring method based on the depth characteristics of multispectral images of an unmanned aerial vehicle.
Detailed Description
Embodiments of the present application are described in further detail below with reference to the accompanying drawings and examples. The following examples are illustrative of the application but are not intended to limit the scope of the application.
Example 1
The embodiment provides a wheat rust monitoring method based on the multispectral image depth characteristics of an unmanned aerial vehicle, which is shown in fig. 1 and comprises the following steps:
s1, collecting measured data on the ground, randomly selecting wheat at different positions in a research area as sampling parties, collecting corresponding coordinates of each sampling party, and respectively calculating the leaf disease ratio of the wheat in each sampling party;
specifically, through the centimeter level differential positioning system provided by the thousand-point position, randomly selecting sampling sides in a research area, marking central position coordinates of the sampling sides and recording, obtaining diseased leaves and total number of wheat leaves in the sampling sides, and respectively calculating diseases She Bilv RD of the wheat in each sampling side, wherein the disease leaf ratio RD is shown in a formula (1):
RD＝d/s (1)
wherein RD is disease She Bilv, d is the number of diseased leaves in the sampling side, and s is the total number of wheat leaves in the sampling side.
S2, acquiring multispectral images of different wave bands in a research area through a multispectral unmanned aerial vehicle, and acquiring a wheat remote sensing image dataset;
specifically, when obtaining the multispectral image of different wave bands in the research area through multispectral unmanned aerial vehicle, in order to guarantee spectral quality, multispectral unmanned aerial vehicle flies in sunny weather, at local time 11:00-13:00 completes the flight, the spatial resolution of the multispectral image is not lower than 1.6cm per pixel, and in order to ensure the splicing quality of the images, the course overlapping rate of the multispectral unmanned aerial vehicle is not lower than 75%, and the side overlapping rate of the multispectral unmanned aerial vehicle is not lower than 75%.
S3, splicing and preprocessing the multispectral images in the wheat remote sensing image dataset acquired in the step S2 to obtain a wheat spectral image, and calculating the spectral index of the wheat spectral image to obtain a sampling side spectral index image feature set;
the pretreatment comprises the following steps: geographic registration, geometric correction, reflectivity correction, radiation correction, and atmospheric correction;
the unmanned aerial vehicle image is spliced through Pix4 Dapp er software; the geographic registration is completed through ground control points paved in the region to be detected before flying; preprocessing such as geometric correction, reflectivity correction, radiation correction, atmospheric correction and the like is completed in ENVI software;
the spectral indexes comprise a calculated spectral index and an empirical spectral index, the calculated spectral index comprises a difference spectral index, a ratio spectral index and a normalized spectral index, and the calculation of the difference spectral index is shown in a formula (2):
DSI i,j ＝R i -R j (2)
wherein DSI i,j Representing i-band and j-bandThe difference spectrum index of the segments, i represents any one wave band of blue, green, red edge or near infrared, j represents any one wave band of blue, green, red edge or near infrared, R i Representing the reflectivity of any one of blue, green, red edge or near infrared, R j Representing the reflectance of any one of the blue, green, red-edge or near-infrared bands;
the calculation of the ratio spectrum index is shown in the formula (3):
wherein RSI is i,j The ratio spectrum index of the i wave band and the j wave band is represented;
the normalized spectral index is calculated as shown in equation (4):
wherein, NDSI i,j The normalized spectrum indexes of the i wave band and the j wave band are represented;
the empirical spectral index includes anthocyanin reflectance index, optimized soil adjustment vegetation index, triangular vegetation index, wide dynamic range vegetation index, soil adjustment vegetation index, red edge chlorophyll index, modified triangular vegetation index, modified nonlinear vegetation index, green leaf index, and green chlorophyll vegetation index;
the anthocyanin reflectance index is calculated as shown in formula (5):
wherein ARI represents anthocyanin reflectance index, R G Representing the reflectance of the green band, R RE Indicating the red-side band reflectivity;
the calculation of the optimized soil adjustment vegetation index is shown as a formula (6):
wherein OSAVI represents optimizing soil adjustment vegetation index, R NIR Indicating the reflectivity of the near infrared band;
the calculation of the triangular vegetation index is shown in formula (7):
wherein TVI represents a triangular vegetation index;
the calculation of the wide dynamic range vegetation index is shown in formula (8):
wherein WDRVI represents a wide dynamic range vegetation index, R R Indicating the red band reflectivity;
the calculation of the soil conditioning vegetation index is shown in formula (9):
wherein SAVI represents a soil conditioning vegetation index;
the calculation of the chlorophyll index of the red edge is shown in the formula (10):
wherein, CIre represents the red edge chlorophyll index;
the calculation of the modified triangular vegetation index is shown in formula (11):
wherein, MTVI 2 Representing a modified triangular vegetation index;
the calculation of the modified nonlinear vegetation index is shown in formula (12):
wherein MNLI represents the modified nonlinear vegetation index;
the green leaf index is calculated as shown in equation (13):
wherein GLI represents green leaf index, R B Representing blue band reflectivity;
the green chlorophyll vegetation index is calculated as shown in formula (14):
wherein CI is G And represents the green chlorophyll vegetation index.
S4, carrying out correlation feature sequencing and first feature screening on the spectrum index image feature set obtained in the step S3 to obtain an optimal spectrum feature set;
specifically, the relevance ranking comprises pearson relevance coefficient, gray relevance and projection variable importance; the first feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of correlation feature sequencing, calculating AIC values of the regression fitting models by adopting a red pool information criterion, outputting a spectrum index contained in the regression fitting model with the minimum AIC value as an optimal spectrum feature set, and the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
S5, performing texture feature calculation on the wheat remote sensing image dataset obtained in the step S2 and the image in the optimal spectrum feature set obtained in the step S4 to obtain a texture feature set, combining the texture feature set and the optimal spectrum feature set to generate a first candidate feature set, and performing feature sequencing and second feature screening on the first candidate feature set to obtain an optimal combined feature set;
specifically, the multi-spectral image in the wheat remote sensing image data set obtained in the step S2 and the multi-spectral image and the spectrum index image in the optimal spectrum feature set obtained in the step S4 are subjected to texture feature extraction by a probability statistical method of a gray level co-occurrence matrix, wherein the texture features comprise mean value, cooperativity, dissimilarity, information entropy, second moment, correlation, contrast and variance;
the relevance ranking comprises pearson relevance coefficient, gray relevance and projection variable importance; the second feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of correlation feature sequencing, calculating AIC values of the regression models by adopting a red pool information criterion, and outputting spectral indexes and texture index features contained in the regression fitting model with the minimum AIC values as an optimal combined feature set, wherein the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
S6, carrying out depth feature extraction on the wheat remote sensing image data set obtained in the step S2 and the image in the optimal combination feature set obtained in the step S5 through a pre-trained deep learning network to obtain a depth feature set, combining the depth feature set and the optimal combination feature set to generate a second candidate feature set, carrying out feature sequencing and third feature screening on the second candidate feature set to obtain an optimal depth feature set, and constructing a plurality of correlation-regression models by utilizing the optimal depth feature set;
specifically, the deep learning network is a ResNet50 network, a jump link is added between convolution layers of the VGG19 network to obtain the ResNet50 network, the ResNet50 network comprises 49 convolution layers and 1 full connection layer, and the depth feature of the image is extracted through the pretrained ResNet50 network comprises the following steps:
step a, extracting a wheat remote sensing image data set obtained in the step S2 and an image with 15×15 pixels with a sampling area as a center in the optimal combined feature set obtained in the step S5, and expanding the influence into 224×224 pixels to obtain a first expanded image;
step b, expanding the first expanded image from 1 dimension to 3 dimensions by constructing a 3-layer convolution kernel taking a Gaussian kernel as a kernel function to obtain a second expanded image;
inputting the second expanded image into a ResNet50 network to obtain 2048 depth features;
specifically, the relevance sorting comprises pearson relevance coefficient, gray relevance and projection variable importance, the third feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of the relevance feature sorting, calculating AIC values of the regression models by adopting a red pool information criterion, and outputting a spectrum index, a texture index and depth features contained in the regression fitting model with the minimum AIC value as an optimal depth feature set, wherein the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model;
the multiple correlation-regression models constructed by utilizing the optimal depth feature set comprise a pearson correlation coefficient-partial least square regression model, a gray correlation degree-partial least square regression model, a variable projection importance-partial least square regression model, a pearson correlation coefficient-backward propagation neural network model, a gray correlation degree-backward propagation neural network model, a variable projection importance-backward propagation neural network model, a pearson correlation coefficient-random forest model, a gray correlation degree-random forest model, a variable projection importance-random forest model, a pearson correlation coefficient-extreme learning model, a gray correlation degree-extreme learning model and a variable projection importance-extreme learning model;
the partial least square regression model is a multivariate regression analysis method integrating principal component analysis, linear correlation analysis and multiple linear regression, and can simultaneously realize data dimension reduction, regression modeling and correlation analysis among analysis data; the backward propagation neural network model is a multilayer feedforward neural network, has the characteristics of forward propagation and error reverse propagation of signals, and is mainly a model trained by an error reverse propagation algorithm, and comprises 1 input layer, 10 hidden layers and 1 output layer, wherein the learning rate is set to be 0.01; the random forest model is a multi-factor machine learning algorithm, a central thinking is that a sampling sample which is randomly put back from a training set by utilizing a Bootstrap resampling algorithm is utilized, decision tree modeling is carried out on each randomly sampled sample set, the decision trees are formed into a random forest, the voting decision result of a plurality of decision trees is taken as a final prediction result, and the number of the decision trees set by the random forest model adopted by the embodiment is 500; the extreme learning machine model is an improved algorithm based on a single-layer feedforward neural network theory, has the advantages of high learning rate, strong generalization capability, high training precision and the like, and the number of hidden layer nodes of the extreme learning machine model adopted in the embodiment is set to be 50.
S7, comparing the precision of the correlation-regression models constructed in the step S6 through the leaf ratio of the wheat in each sampling party obtained in the step S1, and selecting the correlation-regression model with the highest precision to carry out inversion drawing on the wheat region to be detected to obtain a wheat rust monitoring result.
Example 2
In the embodiment, a certain wheat rust disease test field area of Dakudzuvine city in Henan province is selected, the Dakudzuvine city is located in the middle of Henan province, belongs to a warm-zone quaternary wind area, has average daily irradiation of 2280 hours, average annual temperature of 15.0 ℃ and annual precipitation of 579mm, and is suitable for wheat growth. The length of the test field is 115m in the north and south, the east and west directions are respectively 93.6m, and wheat in the region has rust diseases with different degrees.
The first specific operation of the step of this embodiment is as follows:
and randomly selecting 87 ground sampling sides in a research area, marking the sampling side area by using a unit frame with the size of 10cm multiplied by 10cm, respectively recording the number of diseased leaves and the total number of wheat leaves in each sampling side, marking the central position coordinate of the sampling side by using a centimeter-level differential positioning system provided by thousand-hunting positions, and recording numbers.
The specific operation of the second step of this embodiment is as follows:
the multispectral unmanned aerial vehicle used in the example is a Xinjiang fairy 4 multispectral version unmanned aerial vehicle Phantom4-M, the Xinjiang fairy 4 multispectral version unmanned aerial vehicle Phantom4-M integrates a visible light camera and five multispectral cameras, the effective pixels of the visible light camera and the multispectral cameras are 200 ten thousands, and the focal length is 5.74mm. The test is carried out by DJITerra software for route planning, the planned flying height is 30m, the heading and the side overlap rate are respectively set to 80 percent and 75 percent, and the spatial resolution of a single image is 1.6cm per pixel.
The specific operation of the third step of this embodiment is as follows:
137X 5 images are collected through the unmanned aerial vehicle, spliced through the Pix4 Dapper software, and then preprocessed through the ENVI software. And extracting a sampling side spectrum value in the multispectral image by using ENVI software according to the sampling side coordinates, repeating each sampling side for more than three times, and selecting an average value of repeatedly extracted spectrum values for multiple times as the sampling side spectrum value.
The specific operation of the step four of the application is as follows:
calculating to obtain a spectrum index of a wheat spectrum image, obtaining a spectrum index image feature set, carrying out feature sequencing and first feature screening on the spectrum index image feature set, carrying out correlation sequencing through pearson correlation coefficient, gray correlation degree and projection variable importance, constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of the correlation feature sequencing, adopting a red pool information criterion, and outputting a spectrum index contained in a regression fitting model with the minimum AIC value as an optimal spectrum feature set by calculating AIC values of the plurality of regression models; performing texture feature calculation on the wheat remote sensing image dataset and images in the optimal spectral feature set to obtain a texture feature set, combining the texture feature set and the optimal spectral feature set to generate a first candidate feature set, performing feature ordering and second feature screening on the first candidate feature set, performing relevance ordering through a pearson correlation coefficient, gray relevance and projection variable importance, constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of the relevance feature ordering, and calculating A of the regression models by adopting a red pool information criterionThe IC value outputs the spectrum index and texture index features contained in the regression fit model with the minimum AIC value as an optimal combination feature set; deep feature extraction is carried out on a feature image corresponding to a wheat remote sensing image data set and an optimal combination feature set through a pre-trained deep learning network to obtain a deep feature set, the deep feature set and the optimal combination feature set are combined to generate a second candidate feature set, feature ordering and third feature screening are carried out on the second candidate feature set, a plurality of regression fitting models are built based on a machine learning model according to various results of correlation feature ordering, an AIC value of the regression fitting models is calculated through an AIC information rule, a spectrum index, a texture index and a depth feature contained in the regression fitting model with the minimum AIC value are output as an optimal deep feature set, a plurality of correlation-regression models are built through the optimal deep feature set, and an AIC value and a judgment coefficient R of the red pool information rule of the correlation-regression models are calculated 2 Value and root mean square error RMSE, determination coefficient R 2 The higher the value and the lower the root mean square error RMSE, the better the model effect, and the specific results are shown in table 1.
Table 1 monitoring model effect calculation results
The fifth specific operation of the present embodiment is as follows:
based on the comparison, the decision coefficient R of the extreme learning machine-variable projection importance model 2 The value is highest and the root mean square error RMSE is lowest, so the precision of the extreme learning machine-variable projection importance model is highest, and the correlation-regression model is used for carrying out inversion drawing on the wheat region to be detected to obtain the wheat rust monitoring result.
The embodiments of the application have been presented for purposes of illustration and description, and are not intended to be exhaustive or limited to the application in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments were chosen and described in order to best explain the principles of the application and the practical application, and to enable others of ordinary skill in the art to understand the application for various embodiments with various modifications as are suited to the particular use contemplated.
Claims (10)
1. A wheat rust monitoring method based on the multispectral image depth characteristics of an unmanned aerial vehicle is characterized by comprising the following steps:
s1, collecting measured data on the ground, randomly selecting wheat at different positions in a research area as sampling parties, collecting corresponding coordinates of each sampling party, and respectively calculating the leaf disease ratio of the wheat in each sampling party;
s2, acquiring multispectral images of different wave bands in a research area through a multispectral unmanned aerial vehicle, and acquiring a wheat remote sensing image dataset;
s3, splicing and preprocessing the multispectral images in the wheat remote sensing image dataset acquired in the step S2 to obtain a wheat spectral image, and calculating the spectral index of the wheat spectral image to obtain a sampling side spectral index image feature set;
s4, carrying out correlation feature sequencing and first feature screening on the spectrum index image feature set obtained in the step S3 to obtain an optimal spectrum feature set;
s5, performing texture feature calculation on the wheat remote sensing image dataset obtained in the step S2 and the image in the optimal spectrum feature set obtained in the step S4 to obtain a texture feature set, combining the texture feature set and the optimal spectrum feature set to generate a first candidate feature set, and performing feature sequencing and second feature screening on the first candidate feature set to obtain an optimal combined feature set;
s6, carrying out depth feature extraction on the wheat remote sensing image data set obtained in the step S2 and the images in the optimal combination feature set obtained in the step S5 through a pre-trained deep learning network to obtain a depth feature set, combining the depth feature set and the optimal combination feature set to generate a second candidate feature set, carrying out feature sequencing and third feature screening on the second candidate feature set to obtain an optimal depth feature set, and constructing a plurality of correlation-regression models by utilizing the optimal depth feature set;
s7, comparing the accuracy of the correlation-regression models constructed in the step S6 through the disease leaf ratio of the wheat in each sampling party obtained in the step S1, selecting the correlation-regression model with the highest accuracy, and carrying out inversion drawing on the wheat region to be detected to obtain a wheat rust monitoring result.
2. The wheat rust monitoring method based on the multispectral image depth characteristics of the unmanned aerial vehicle according to claim 1, wherein in the step S1, a sampling party is randomly selected in a research area through a centimeter-level differential positioning system provided by a thousand-finding position, the sampling party is a 10cm×10cm sampling party, the central position coordinates of the sampling party are marked and recorded, the number of diseased leaves and the total number of wheat leaves in the sampling party are obtained, the disease She Bilv RD of the wheat in each sampling party is calculated, and the disease leaf ratio RD is shown in a formula (1):
RD＝d/s (1)
wherein RD is disease She Bilv, d is the number of diseased leaves in the sampling side, and s is the total number of wheat leaves in the sampling side.
3. The wheat rust monitoring method based on the depth features of the multispectral images of the unmanned aerial vehicle according to claim 2, wherein in the step S2, when multispectral images of different wave bands in the research area are acquired through the multispectral unmanned aerial vehicle, the multispectral unmanned aerial vehicle flies in sunny weather, and the local time is 11:00-13: and 00, finishing the flight, wherein the spatial resolution of the multispectral image is not lower than 1.6cm per pixel, the course overlap ratio of the multispectral unmanned aerial vehicle is not lower than 75%, and the side overlap ratio of the multispectral unmanned aerial vehicle is not lower than 75%.
4. A wheat rust monitoring method based on the depth features of the multispectral image of the unmanned aerial vehicle according to claim 3, wherein in the step S3, the preprocessing includes: geographic registration, geometric correction, reflectivity correction, radiation correction, and atmospheric correction;
the spectral indexes comprise a calculated spectral index and an empirical spectral index, the calculated spectral index comprises a difference spectral index, a ratio spectral index and a normalized spectral index, and the calculation of the difference spectral index is shown in a formula (2):
DSI i,j ＝R i -R j (2)
wherein DSI i,j The spectrum index of the difference between the i wave band and the j wave band is represented, i represents any wave band of blue, green, red edge or near infrared, j represents any wave band of blue, green, red edge or near infrared, R i Representing the reflectivity of any one of blue, green, red edge or near infrared, R j Representing the reflectance of any one of the blue, green, red-edge or near-infrared bands;
the calculation of the ratio spectrum index is shown in a formula (3):
wherein RSI is i,j The ratio spectrum index of the i wave band and the j wave band is represented;
the normalized spectral index is calculated as shown in formula (4):
wherein, NDSI i,j The normalized spectrum indexes of the i wave band and the j wave band are represented;
the empirical spectrum index comprises anthocyanin reflection index, optimized soil adjustment vegetation index, triangular vegetation index, wide dynamic range vegetation index, soil adjustment vegetation index, red edge chlorophyll index, modified triangular vegetation index, modified nonlinear vegetation index, green leaf index and green chlorophyll vegetation index;
the anthocyanin reflection index is calculated as shown in a formula (5):
wherein ARI represents anthocyanin reflectance index, R G Representing the reflectance of the green band, R RE Indicating the red-side band reflectivity;
the calculation of the optimized soil adjustment vegetation index is shown in a formula (6):
wherein OSAVI represents optimizing soil adjustment vegetation index, R NIR Indicating the reflectivity of the near infrared band;
the calculation of the triangular vegetation index is shown in formula (7):
wherein TVI represents a triangular vegetation index;
the calculation of the wide dynamic range vegetation index is shown in formula (8):
wherein WDRVI represents a wide dynamic range vegetation index, R R Indicating the red band reflectivity;
the calculation of the soil conditioning vegetation index is shown in formula (9):
wherein SAVI represents a soil conditioning vegetation index;
the calculation of the chlorophyll index of the red edge is shown in the formula (10):
wherein, CIre represents the red edge chlorophyll index;
the calculation of the modified triangular vegetation index is shown in formula (11):
wherein, MTVI 2 Representing a modified triangular vegetation index;
the calculation of the modified nonlinear vegetation index is shown in formula (12):
wherein MNLI represents the modified nonlinear vegetation index;
the green leaf index is calculated as shown in equation (13):
wherein GLI represents green leaf index, R B Representing blue band reflectivity;
the green chlorophyll vegetation index is calculated as shown in formula (14):
wherein CI is G And represents the green chlorophyll vegetation index.
5. The method for monitoring wheat rust based on the depth features of the multispectral image of the unmanned aerial vehicle according to claim 4, wherein in the step S4, the relevance ranking includes pearson correlation coefficient, gray relevance and projection variable importance; the first feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of correlation feature sequencing, calculating AIC values of the regression fitting models by adopting a red pool information criterion, and outputting a spectrum index contained in the regression fitting model with the minimum AIC value as an optimal spectrum feature set, wherein the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
6. The wheat rust monitoring method based on the depth features of the multispectral images of the unmanned aerial vehicle according to claim 5, wherein in the step S5, texture feature extraction is performed on the multispectral images in the data set of the wheat remote sensing images obtained in the step S2 and the multispectral images and the spectral index images in the optimal spectral feature set obtained in the step S4 by a probability statistics method of a gray level co-occurrence matrix, and the texture features comprise mean value, cooperativity, dissimilarity, information entropy, second moment, correlation, contrast and variance.
7. The method for monitoring wheat rust based on the depth features of the multispectral image of the unmanned aerial vehicle according to claim 6, wherein in the step S5, the correlation ranking includes pearson correlation coefficient, gray correlation degree and projection variable importance; the second feature screening comprises the steps of constructing a plurality of regression fitting models based on a machine learning model according to a plurality of results of correlation feature sequencing, calculating AIC values of the regression fitting models by adopting a red pool information criterion, and outputting spectral indexes and texture index features contained in the regression fitting model with the minimum AIC values as an optimal combined feature set, wherein the machine learning model comprises a partial least square regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
8. The wheat rust monitoring method based on the multispectral image depth features of the unmanned aerial vehicle according to claim 7, wherein the deep learning network in the step S6 is a res net50 network, a skip link is added between convolution layers of the VGG19 network to obtain the res net50 network, the res net50 network comprises 49 convolution layers and 1 full-connection layer, and the depth features of the images are extracted through the pre-trained res net50 network comprises the following steps:
step a, extracting the wheat remote sensing image dataset obtained in the step S2 and an image with 15×15 pixels with a sampling area as a center in the optimal combined feature set obtained in the step S5, and expanding the influence into 224×224 pixels to obtain a first expanded image;
step b, expanding the first expanded image from 1 dimension to 3 dimensions by constructing a 3-layer convolution kernel taking a Gaussian kernel as a kernel function to obtain a second expanded image;
and c, inputting the second expanded image into the ResNet50 network to obtain 2048 depth features.
9. The method for monitoring wheat rust based on the depth features of the multispectral image of the unmanned plane according to claim 8, wherein in the step S6, the relevance ranking comprises pearson relevance coefficient, gray relevance and projection variable importance, the third feature screening comprises constructing a plurality of regression fit models based on a machine learning model according to various results of the relevance feature ranking, calculating AIC values of the regression fit models by using a red pool information criterion, and outputting spectral indexes, texture indexes and depth features contained in the regression fit model with the minimum AIC values as an optimal depth feature set, and the machine learning model comprises a partial least squares regression model, a backward propagation neural network model, a random forest model and an extreme learning machine model.
10. The method according to claim 9, wherein in the step S6, the plurality of correlation-regression models constructed by using the optimal depth feature set include a pearson correlation coefficient-partial least squares regression model, a gray correlation degree-partial least squares regression model, a variable projection importance-partial least squares regression model, a pearson correlation coefficient-backward propagation neural network model, a gray correlation degree-backward propagation neural network model, a variable projection importance-backward propagation neural network model, a pearson correlation coefficient-random forest model, a gray correlation degree-random forest model, a variable projection importance-random forest model, a pearson correlation coefficient-extreme learning model, a gray correlation degree-extreme learning model, and a variable projection importance-extreme learning model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311306532.3A CN117152645A (en) | 2023-10-10 | 2023-10-10 | Wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311306532.3A CN117152645A (en) | 2023-10-10 | 2023-10-10 | Wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117152645A true CN117152645A (en) | 2023-12-01 |
Family
ID=88900967
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311306532.3A Pending CN117152645A (en) | 2023-10-10 | 2023-10-10 | Wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117152645A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117451639A (en) * | 2023-12-21 | 2024-01-26 | 内蒙古工业大学 | A method for retrieving chlorophyll concentration in water based on remote sensing data |
| CN118566141A (en) * | 2024-06-22 | 2024-08-30 | 中国农业大学 | A method for detecting oat seed varieties based on multispectral technology |
-
2023
- 2023-10-10 CN CN202311306532.3A patent/CN117152645A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117451639A (en) * | 2023-12-21 | 2024-01-26 | 内蒙古工业大学 | A method for retrieving chlorophyll concentration in water based on remote sensing data |
| CN118566141A (en) * | 2024-06-22 | 2024-08-30 | 中国农业大学 | A method for detecting oat seed varieties based on multispectral technology |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111461052B (en) | Wheat lodging area identification method for multiple growth periods based on transfer learning | |
| Zhang et al. | UAV hyperspectral remote sensing image classification: A systematic review | |
| US20230316555A1 (en) | System and Method for Image-Based Remote Sensing of Crop Plants | |
| CN113553697B (en) | Long-time-sequence multi-source data-based vegetation disturbance analysis method for coal mining | |
| CN110287944A (en) | Crop pest monitoring method based on multispectral remote sensing imagery based on deep learning | |
| CN110889394A (en) | Rice lodging recognition method based on deep learning UNet network | |
| CN111161362A (en) | Tea tree growth state spectral image identification method | |
| CN116883853B (en) | Remote sensing classification method of crop spatiotemporal information based on transfer learning | |
| CN117152645A (en) | Wheat rust monitoring method based on unmanned aerial vehicle multispectral image depth characteristics | |
| CN120147724A (en) | Intelligent decision-making method and system for precision fertilization based on multi-source remote sensing | |
| CN116503590B (en) | A crop segmentation method for multispectral UAV remote sensing images | |
| CN115170987A (en) | Method for detecting diseases of grapes based on image segmentation and registration fusion | |
| Zhang et al. | Lightweight deep learning models for high-precision rice seedling segmentation from UAV-based multispectral images | |
| CN109063660A (en) | A Crop Recognition Method Based on Multispectral Satellite Imagery | |
| CN117197668A (en) | Prediction method and system for crop lodging level based on deep learning | |
| CN116630827B (en) | A method for estimating cotton yield using cotton boll index obtained from UAV remote sensing imagery. | |
| CN117197656A (en) | Multi-mode pasture image feature extraction and recognition system | |
| CN114429471A (en) | Crop disaster identification and classification method and device based on deep learning | |
| Johri et al. | Role of satellites in agriculture | |
| Wang et al. | A grid-level segmentation model based on encoder-decoder structure with multi-source features for crop lodging detection | |
| CN114863273B (en) | A method for processing and predicting multi-source remote sensing yield data in the early season of crops | |
| CN119295910B (en) | Crop lodging area grading identification method and system | |
| Bhavana et al. | Crop field boundary detection and classification using machine learning | |
| AHM et al. | A deep convolutional neural network based image processing framework for monitoring the growth of soybean crops | |
| Bagherian et al. | Phenotyping agronomic traits of peanuts using UAV-based hyperspectral imaging and deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |