CN113627392A

CN113627392A - Rape waterlogging hyperspectral image detection method based on Spark platform and image acquisition device thereof

Info

Publication number: CN113627392A
Application number: CN202111029184.0A
Authority: CN
Inventors: 夏吉安; 曹宏鑫; 张文宇; 张伟欣; 葛道阔
Original assignee: Jiangsu Academy of Agricultural Sciences; Nanjing Vocational University of Industry Technology NUIT
Current assignee: Jiangsu Academy of Agricultural Sciences; Nanjing Vocational University of Industry Technology NUIT
Priority date: 2021-09-02
Filing date: 2021-09-02
Publication date: 2021-11-09
Anticipated expiration: 2041-09-02
Also published as: CN113627392B

Abstract

The invention discloses a rape waterlogging hyperspectral image detection method and an image acquisition device thereof based on Spark platforms, which relate to the technical field of information processing and comprise the steps of image acquisition, image format conversion, image storage and management … … to obtain the classification detection analysis result of image data, and the like, the invention carries out the distributed parallel classification analysis of the image data under the stress of crop disasters by a Hadoop framework to accelerate the modeling and prediction speed of a classification algorithm, the greater the image data quantity, the more obvious the advantage is, the algorithm is suitable for the parallel operation under the Hadoop framework to avoid the damage and the loss of the image data, and a plurality of mooring platforms are arranged in the field, corresponding image collectors are arranged on each mooring platform according to the needs to carry out the acquisition of crop spectral images, thereby solving the defects of inconvenient image acquisition and low image quality acquired by unmanned planes in the traditional mode, the mooring platform is flexible in installation mode, can be arranged at the corresponding position in the field as required, is low in cost, and improves the collection efficiency.

Description

Rape waterlogging hyperspectral image detection method based on Spark platform and image acquisition device thereof

Technical Field

The invention relates to the technical field of information processing, in particular to a rape waterlogging hyperspectral image detection method based on a Spark platform and an image acquisition device thereof.

Background

Along with the rapid development of agricultural informatization and agricultural big data, agriculture in China is in a transition stage from traditional agriculture to precision agriculture. The spectroscopy technology and the big data technology provide an effective solution for accurate agricultural detection and provide scientific basis for agricultural research.

Present crops blade spectral information gathers, basically all gather through artifical handheld spectral scanner or use unmanned aerial vehicle spectral scanner, ground environment is complicated in the farm, and ground unevenness also increases the degree of difficulty of researcher collection work, the collection efficiency has been reduced, and the position height that unmanned aerial vehicle was located is generally very high (in tens of meters even in the air of hundreds of meters) and stability relatively poor, the data quality of collection is also relatively poor, the post processing degree of difficulty is big.

In the current crop non-stress spectral image detection and analysis, crop disaster information collected by a spectrum and machine vision system is usually stored in a computer, and traditional analysis software such as Matlab and SPSS is used for modeling analysis processing. The accumulated crop spectral image information will be more and the data volume will be larger and larger as time goes on. When the amount of spectral image data is too large, the analysis processing speed is slower and slower when the traditional analysis processing method is used, the modeling complexity is higher and higher, and even the analysis can not be normally carried out.

The hyperspectral image technology plays an important role in modern agriculture and food processing industry. In the application of agricultural hyperspectral imaging, the advantages of hyperspectral maps in one can be fully utilized, and the growth vigor of crops, particularly the aspects of crop growth vigor assessment, disaster monitoring, agricultural management and the like can be accurately monitored through a large amount of spectral image information contained in hyperspectral images. In addition, the hyperspectral images can improve and develop references to biological specimens of interest and build a library of computer-aided analysis images.

Spark is a large data distributed processing and analyzing framework with a mature system, can provide a scalable and reliable data analysis processing and storage system, and forms a large server cluster. On the other hand, the Spark framework can also support effective parallel computation in a distributed mode, and the processing and analysis speed of the hyperspectral image can be greatly improved.

Disclosure of Invention

The invention provides a rape waterlogging hyperspectral image detection method based on a Spark platform and an image acquisition device thereof, and aims to solve the technical problems.

The technical solution for realizing the purpose of the invention is as follows:

s1, placing a plurality of image collectors 3 on a plurality of mooring platforms 2 arranged in the field by using an unmanned aerial vehicle 1, fixing the image collectors 3 by means of the mooring platforms 2, collecting images of rape waterlogging stress leaves in the field by using a plurality of camera modules 34 arranged on the side surfaces of the image collectors 3, transmitting spectral data to an analysis platform of a management center through a wireless transceiver module arranged in the image collectors 3, and recovering the image collectors 3 to an equipment warehouse of the management center for storage by using the unmanned aerial vehicle 1 after collection work is finished;

s2, for miningCorrecting and preprocessing the collected image, converting the color channel of the image, and converting RGB into La^*b^*Converting the RGB image into a three-dimensional matrix, and adopting a principal component analysis method and extracting image characteristic information to reduce data dimensionality and reduce data calculation amount;

s3, performing Distributed storage and management on the image matrix by using an HDFS (Hadoop Distributed File System) node, wherein a Master node is used as a Namenode node to manage the whole HDFS File System and the directory tree, and one copy is stored on the other 5 Datano nodes;

s4, constructing a parallel neural network algorithm on the main node;

s5, distributing and scheduling the jobs by utilizing an HDFS distributed framework through a Hadoop cooperative work mechanism, reading data blocks by nodes and finishing job tasks and calculation;

s6, the master node sorts and combines the calculation results, and finally gathers the calculation results to obtain the neural network parallel image classification results;

s7, in order to avoid over (under) fitting of a clustering algorithm, dividing image data into a training set and a prediction set, wherein the data volume of the training set accounts for 75%, establishing a neural network training model, and performing parameter tuning;

and S8, inputting the residual 25% of data sets after the algorithm training is finished, performing algorithm prediction, and performing model performance evaluation to obtain a classification detection analysis result of the image data.

In the data analysis process, the characteristic wave band analysis is carried out by collecting the rape hyperspectral image stressed by waterlogging, the spectral characteristic information is extracted by adopting a multivariate scattering correction, projection and regression method, the image characteristic information is extracted by using an image enhancement and principal component analysis technology, and the characteristic analysis hyperspectral image information is constructed. The distributed computing platform based on the Spark framework is used, the spectrum image matrix is stored in an HDFS mode, the parallel multilayer perceptron algorithm based on the Spark framework is used for carrying out classification detection and identification on the characteristic information of the hyperspectral image, and finally the weighted harmonic mean is used for computing the classification detection result, so that the crop waterlogging image detection and the distributed storage of crop image data are realized.

The invention also discloses an image acquisition device for acquiring images of field rape waterlogging stress blades in the rape waterlogging hyperspectral image detection method based on the Spark platform, which comprises an unmanned aerial vehicle, a mooring platform and an image collector, wherein the unmanned aerial vehicle is used for transferring the image collector, and the bottom of the unmanned aerial vehicle is provided with an electromagnet;

the image collector is used for collecting images of rape waterlogging stress leaves in the field, a cylindrical magnetic metal block matched with the electromagnet is arranged at the top of the image collector, and a movable female connector is arranged at the bottom of the image collector;

the mooring platform is provided with a plurality of locking assemblies and a sub-connector, the locking assemblies are used for fixing the image collector, the sub-connector is matched with the female connector, and the mooring platform is used for temporarily placing the image collector.

Preferably, the image collector includes quick-witted case, cylindric magnetic metal piece, female connector and a plurality of module of making a video recording, the machine case is cylindrical, and cylindric magnetic metal piece is installed in the top of quick-witted case, and a plurality of modules of making a video recording are installed in the side of quick-witted case and are circumference and arrange, and the bottom of machine case is equipped with female connector matched with spout one, and female connector sliding connection is in spout one, and when the bottom of quick-witted case does not have the barrier, female connector is located the bottom and extends to the quick-witted case outside for quick-witted case under the effect of dead weight.

Preferably, the mooring platform comprises a base, a shell, a column, a locking assembly, a plugging assembly and an adsorption assembly, the base is arranged at the upper end of the shell, a through hole for connecting equipment is arranged at the geometric center of the base, the shell is arranged on a fixture arranged in the field by means of an upright post, the locking component is arranged in the shell and is used for fixing the image collector placed on the base, the plugging component is arranged at the upper end of the locking component and can plug the through hole, the sub-connector is arranged in the plugging component, the suction force generated by the adsorption component can absorb the sub-connector in the plugging component and is connected with the female connector so as to connect the antenna connector of the wireless transceiver module in the image collector with an external antenna on the mooring platform, so that the signal transmission capability of the image collector and the management center is enhanced.

Preferably, the locking component comprises a clamping block, a connecting rod, a connecting seat and an electric push rod, the base is provided with a sliding hole matched with the clamping block, the clamping block is three and is respectively connected in the three sliding holes in a sliding manner, the lower end of the clamping block is rotatably connected with the connecting rod, the other end of the connecting rod is rotatably connected to the connecting seat, the lower end of the connecting seat is connected to the output end of the electric push rod, and the electric push rod is installed on a base arranged inside the shell.

Preferably, the plugging component comprises a plug and a flange step, the lower end of the plug is connected with the connecting seat, the flange step is fixed on the periphery of the plug, and a second sliding groove for the connector of the plug to slide up and down is formed in the upper end of the plug.

Preferably, the sub-connector comprises a lower connecting block, an upper connecting block, a ring-shaped magnet and a metal rod, the upper connecting block is fixed on the lower connecting block, the ring-shaped magnet is embedded in the periphery of the upper end of the lower connecting block, an installation groove is formed in the upper end of the upper connecting block, the metal rod is installed in the installation groove, and a waterproof rubber plug with a cross-shaped notch is installed at an opening in the upper end of the installation groove;

female connector includes slide and adapter sleeve, the bottom of machine case is equipped with and supplies the gliding spout three of slide, and the bottom of slide is equipped with and supplies the spout four that the connecting block passed, the adapter sleeve cooperates and installs in spout four with the metal pole.

Preferably, the side of slide is equipped with the ring channel, adsorption component includes arc seat and arc magnet, the arc seat have three and with three clamp tight piece one-to-one, the upper end of arc seat with the help of traveller sliding connection in the guide way of base lower extreme, the arc seat still with the help of the connecting plate clamp the piece and be connected, arc magnet involves on the arc seat, a complete ring magnet can be constituteed to three arc magnet.

Compared with the prior art, the invention has the advantages that:

(1) the distributed parallel classification analysis of the image data under the crop disaster stress is carried out through a Hadoop framework, the modeling and predicting speed of a classification algorithm is accelerated, and the advantages are more obvious when the image data volume is larger than that of a single machine mode; writing a neural network algorithm by using a Scala language, wherein the algorithm is suitable for parallelization operation under a Hadoop framework; and the HDFS of the Hadoop frame is adopted for distributed storage of the image data, so that damage and loss of the image data are avoided.

(2) According to the invention, a plurality of mooring platforms are arranged in the field, and corresponding image collectors are placed on the mooring platforms as required to collect the images of the rape waterlogging stress leaves in the field, so that the defects of inconvenience in image collection and low image quality collected by an unmanned aerial vehicle in the traditional mode are overcome, the mounting mode of the mooring platforms is flexible, the mooring platforms can be arranged at corresponding positions in the field as required, the cost is low, the transfer of the image collectors is completed by the unmanned aerial vehicle, the unmanned aerial vehicle is not limited by the region and space, and the image collection efficiency is greatly improved.

(3) According to the mooring platform, even if the position of the unmanned aerial vehicle is not accurate enough in the process of placing the image collector, the locking and positioning of the image collector can be rapidly completed through the arranged locking assembly, the adsorption assembly can clamp and fix the female connector falling from the bottom of the machine box while locking, the auxiliary connector is driven to move upwards by the aid of the arc-shaped magnet and finally connected with the female connector, so that the antenna connector of the wireless transceiver module in the image collector can be connected with the external antenna on the mooring platform, the signal transmission capacity is improved, and the mooring platform is suitable for long-distance signal transmission. Meanwhile, the plugging assembly can plug the through hole when the mooring platform is in no-load state, and related parts inside the shell are protected.

Drawings

FIG. 1 is a hyperspectral image processing flow.

FIG. 2 is a hyperspectral image feature band extraction.

FIG. 3 is a process flow of big data platform hyperspectral image processing.

FIG. 4 is a big data platform workflow.

Fig. 5 is a schematic diagram of an image acquisition process in the present invention.

Fig. 6 and 7 are schematic structural views of the mooring platform and the image collector in different viewing angles.

Fig. 8 and 9 are schematic structural views of the mooring platform part in different viewing angles.

Fig. 10 is a cross-sectional view of fig. 8.

Fig. 11 is a schematic structural view of the plugging element and the sub-connector of the present invention.

Fig. 12 is a cross-sectional view of fig. 11.

Fig. 13 is a sectional view of an image collector in the present invention.

Fig. 14 is a schematic connection diagram of the female connector and the male connector.

Wherein: 1 unmanned aerial vehicle, 2 mooring platform, 21 base, 211 through-hole, 22 casing, 23 stand column, 24 locking subassembly, 241 presss from both sides tight piece, 242 connecting rods, 243 connecting seat, 244 electric push rod, 245 slide opening, 246 base, 25 shutoff subassembly, 251 end cap, 252 flange step, 253 spout two, 26 absorption subassembly, 261 arc seat, 262 arc magnet, 263 connecting plate, 264 guide way, 27 stand column, 28 external antenna, 3 image collector, 31 quick-witted case, 311 spout one, 32 cylindric magnetic metal piece, 33 female connector, 331 slide, 332 adapter sleeve, 333 spout three, 334 spout four, 335 annular groove, 34 the module of making a video recording, 35 sub-connector, 351 lower connecting block, 352 upper connecting block, 353 annular magnet, 354 mounting groove, 355 metal pole, waterproof plug.

Detailed Description

The invention will be further illustrated with reference to the following specific examples. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the present invention. Further, it should be understood that various changes and modifications of the present invention may be made by those skilled in the art after reading the teachings of the present invention, and such equivalents may fall within the scope of the invention as defined in the appended claims.

As shown in fig. 1 to 4, the embodiment discloses a rape waterlogging hyperspectral image detection method based on a Spark platform, which specifically includes:

s2, correcting and preprocessing the collected image, converting the color channel of the image, and converting RGB into La^*b^*The RGB image is converted into a three-dimensional matrix, and the principal component analysis method and the image characteristic information are extracted, so that the data dimensionality is reduced, and the data calculation amount is reduced.

Before image acquisition, the following work is required:

s01, establishing a corresponding relation between the hyperspectral image based on rape waterlogging and different waterlogging stress grades, and predicting the waterlogging stress grade of rape.

S02, establishing a rape waterlogging damage hyperspectral image prediction model based on a Spark parallel computing framework, and obtaining parameters of the prediction model.

Before forecasting rape waterlogging stress, collecting hyperspectral images of rape leaves stressed by waterlogging in the flowering phase, wherein the waterlogging stress lasts for 3 days and 6 days respectively.

Collecting hyperspectral image information corresponding to different levels of rape waterlogging;

selecting spectral image information of an ROI (region of interest) area, and establishing a spectral matrix and an image matrix.

In the hyperspectral image acquisition process, key factors such as the scanning line number, the integral time, the frame rate, the moving speed of a linear platform, the acquisition wave band and the like of a hyperspectral camera are determined.

Specifically, the hyperspectral image acquisition comprises the scanning line number of 800 lines, the distance between the linear platform and the camera lens is 35cm, and the included angle between the light source and the sample is 30 degrees. The camera integration time is 12.01ms, the frame rate is 31.23Hz, the scanning speed of the hyperspectral camera is 0.23cm/s, and the acquisition waveband is 400-1100 nm. Specifically, the extraction of spectral feature information in the hyperspectral image is carried out by adopting a mode of combining projection and stepwise regression.

Smoothing, denoising and feature extraction are carried out on the collected hyperspectral images, and hyperspectral images of 6 characteristic wave bands are selected to establish corresponding spectra and image analysis matrixes.

And specifically, preprocessing modes such as convolution smoothing and integration are used for determining corresponding red edges and green peaks of the spectrum, and a wave band with good predictability is determined to be used as a spectrum prediction model parameter.

And filtering and denoising the spectrum by using a convolution smoothing method to obtain a smoother spectrum curve. Savitzky-Golay is a widely used filtering and smoothing technique in absorption and reflection spectroscopy^[89-90]。

Specifically, the collected spectral data is smoothed and denoised by adopting 5-point convolution smoothing, and the weight factors of corresponding windows are (-3, 12, 17, 12 and 3)

The method comprises the steps of calculating the average spectrum of all samples by using multivariate scattering correction to serve as a reference spectrum, carrying out unary regression operation on the spectrum data of each sample and the reference sample to obtain the linear translation amount (regression constant) and the offset amount (regression coefficient) of each sample spectrum, subtracting the linear translation amount from each sample spectrum and dividing the linear translation amount by the regression coefficient, and modifying the baseline translation amount and the offset amount of each spectrum under the reference spectrum to improve the signal-to-noise ratio of the spectrum.

The specific algorithm steps are as follows:

step 1: calculating the average spectrum:

step 2: unary linear regression:

and step 3: scattering the spectrum:

in the above formula, R is the sample average spectrum matrix, n is the number of samples,

is an average spectrum, m_iAnd b_iThe relative offset coefficient and the translation amount after linear regression.

The derivative spectrum adopts differential derivation, which can improve the resolution of spectrogram, display subtle changes on spectrogram, reduce spectral interference, and eliminate baseline drift and band overlapping.

Specifically, the first-order reflectance spectrum derivation formula: lambda is the reflection value in the original spectrum,

F′(λ)＝[F(λ+Δλ)-F(λ-Δλ)]/2Δλ

the second derivative spectrum is obtained by derivation of the first spectrum.

F″(λ)＝[F′(λ+Δλ)-F′(λ)]/2Δλ

1) Recording the randomly selected initial iteration vector as X_k(0)The number of vectors to be extracted is N, namely the characteristic number after dimensionality reduction;

2) before the first iteration, n is equal to (1), randomly selecting any column j of the spectrum matrix, and assigning the j-th column in the spectrum matrix X of the correction set to X_k(0)Wherein: j is 1, 2, 3 … J-1, J is the number of bands.

3) Let S be the number of bands that are not selected, S is represented as:

S＝{j，1≤j≤J，j∈{k(0)，k(1)，…k(n-1)}}

4) to X_jOrthogonal to X in subspace_k(n-1)The vector projection of (a) is calculated:

wherein: j belongs to S, and P is a projection operator, namely the projection of the initial wave band orthogonal to the other wave bands.

5) In the formula (4), k (n) is represented by:

6) taking the maximum projection value as the initial value of the next iteration:

X_j＝PX_j

7) and if N is equal to N +1 and N is less than N, returning to the step 3), and continuing the loop calculation.

8) All band combinations obtained after dimensionality reduction are represented as W, and then W can be represented as:

W＝{x_k(n)；n＝0，1，…，n-1}

and finally, cycling the W set corresponding to each k (0) and N once, and then performing multiple linear regression analysis (MLR) to obtain the predicted standard deviation (RMSEV) of the verification set, wherein the k (0) and N corresponding to the minimum RMSEV value are the selected characteristic wave band and the selected characteristic value.

And performing feature selection by combining projection and stepwise regression, and selecting hyperspectral images with wave bands of 529nm, 641nm, 698nm, 749nm, 856nm and 979nm to extract spectra and image information for further information processing.

Image enhancement:

extracting RGB image information from the collected stain hyperspectral image, converting the RGB image into an HSI channel image by using color channel conversion, extracting H (hue), I (brightness) and S (saturation) information from an HIS image, and modifying I, S channel histogram information to enhance the HSI image.

And converting the HIS channel image into an RGB image to obtain an enhanced color RGB image, and enhancing the reliability and resolution of image data.

The principal components of the hyperspectral images are extracted, the data volume of the hyperspectral images can be further reduced, and an image analysis model of the characteristic index is established. And standardizing RGB image data, calculating a covariance matrix of the image matrix, solving an eigenvalue and an eigenvector of the covariance matrix of the image matrix, turning the vector of the image in the horizontal direction and turning the vector from top to bottom, and extracting the main components of RGB image information.

The hyperspectral image is assumed to be an n × m dimensional matrix:

wherein:

solving the normalized matrix of the image:

(1) the data standard deviation σ is:

where n is the number of samples, μ is the mean of the total sample data, x_iIs the value of the ith element.

The normalization matrix is:

and (3) solving a correlation coefficient matrix:

calculating eigenvalue and eigenvector (lambda) of correlation coefficient matrix₁，λ₂，…，λ_n) And a feature vector a_i＝(a_i1，a_i2，…，a_in)

Ax＝λx

Solving the cumulative variance contribution rate:

the image is converted into a three-dimensional matrix, and the image characteristic information is extracted by using a principal component analysis method, so that the data volume is reduced.

Calculating the mean and variance of the image matrix; standardizing the matrix and calculating a covariance matrix; solving an eigenvalue and an eigenvector of the covariance matrix; and calculating the cumulative variance contribution rate, wherein the contribution rate is more than 70%, and the cumulative variance contribution rate can be set as a principal component.

The Hadoop platform uses a standalon mode, wherein a Master Node is responsible for resource allocation and job scheduling, a Node1-Node5 is used as a Node to execute a calculation task, and a Master Node is also used as a Node to perform task processing.

The multilayer perceptron algorithm comprises a three-layer structure, wherein the first layer is an input layer, and the number of nodes of the input layer is the characteristic number of a sample; the second layer is a hidden layer, and the number of nodes is manually set; the third layer is an output layer, and the number of nodes is the characteristic quantity of the sample target. In the forward propagation process of the signal, the input layer is used as the input of the hidden layer node to calculate the output of the hidden layer node, and meanwhile, the output of the hidden layer node is used as the input of the output layer to calculate the input of the output layer.

On the basis of multivariate scattering correction, projection, stepwise regression and image enhancement, the spectrum and image matrix analysis of the rape waterlogging hyperspectral image are combined for innovation and optimization, the characteristic spectrum image of the hyperspectral image is combined, the noise information is removed through the decomposition and screening of the spectrum and the image matrix, and proper parameters are selected to prevent overfitting of training. And (3) realizing a classifier model based on a multilayer perceptron by utilizing a Spark framework, and training and predicting a parallel model.

The specific modeling process is as follows:

let X ∈ R as input space (feature space)ⁿThe output space is y { +1, -1}, and X ∈ X is the feature vector of the instance, so that the following function is applied from the input space to the output space:

f(x)＝sign(ωx+b)

wherein: omega epsilon to RⁿAs a weight vector, b ∈ R is a bias, ω x is an inner product of ω and x, sign is a sign function:

the loss function is defined as:

the distance from one point in the feature space to the hyperplane is set as follows:

the distances from all error points to the hyperplane are:

the loss function is:

wherein, the meaning of omega, x, b and y is the formula, and M is the set of classification error points.

And solving the minimum value of the loss function by adopting a random gradient descent mode, and accelerating the function convergence.

Randomly selecting a point of misclassification (x)_i，y_i) Update (ω, b):

ω←ω+ηx_iy_i

b←b+ηy_i

where η is the step size (learning rate), the loss function L (ω, b) is reduced by iterative calculation until it becomes 0.

In this problem, assuming that the initial values of (ω, b) are both 0, the point of miscut (x) is defined_i，y_i) By:

ω←ω+ηx_iy_i

b←b+ηy_i

updating the parameters (ω, b) and finally learning:

wherein alpha is_i＝n_iη

The parameters of the multi-layer perceptron include weights and biases of the input layer and the hidden layer, so that

Is the connection weight between the ith cell of the kth layer and the ith cell of the (k + 1) th layer, b_kIs a k-layer bias term, and after the signal propagates forward, the output function of the output layer node is:

the middle layer uses Sigmoid function as activation function. By introducing nonlinear factors into the neurons, the neural network can arbitrarily approximate a nonlinear function. The output neuron calculates the output value of the whole neural network according to the input value and the activation function.

And (3) the back propagation of the error is carried out, the partial derivative of the objective function to the weight of each neuron is calculated layer by layer from the output layer, the weight and the threshold value of each layer are adjusted by using a gradient descent function according to the calculated partial derivative, and the weight value is updated until the final output value of the modified network is close to the expected value. Since the Sigmoid function is easy to have gradient vanishing when the Sigmoid function is reversely propagated, the Softmax function is used as a gradient descending function of a multi-classification problem when the Sigmoid function is reversely propagated.

Inputting a spectral image characteristic data set, wherein the step length is alpha, the prior (training) sample size is beta, and the iteration number is tau.

Initializing network parameters

For t∈{1，2，…n}

Feed forward: the output of each neuron is computed at each connected layer.

Calculating an error according to the loss function to obtain a gradient function of the output layer;

inversely calculating the degree and gradient of error contribution of each neuron in the upper layer, calculating connection parameters and updating weights,

and (4) when the algorithm reaches a reverse input layer, iteratively updating the parameters and training the multilayer perceptron.

And S3, uploading the spectral image data to a Hadoop computing platform, managing the whole HDFS file system and the directory tree by adopting an HDFS storage form and taking a Master node as a Namenode node, and storing a copy of an image data set on other 5 Datano nodes.

And (3) completing a neural network algorithm program by a Hadoop platform client (JobClient), and compiling and operating the program. Hadoop JobClient sends a task request to JobTracker, applies for free available computing resources (Job), and the JobTracker returns an available Job ID to the JobClient.

And after obtaining the Job ID, the JobClient copies the resource file required by running Job into a file system HDFS, and the copied file information is stored in a folder created by JobTracker.

After the resources are completely prepared, the submission task is started, and JobClient submits a Job task (Job) to JobTracker.

After receiving Job, JobTracker initializes Job, puts it into the Job queue, and waits for the Job scheduler to schedule it.

After initialization is completed, the JobTracker acquires input splits from the HDFS, creates a Map task for each partition according to a scheduling algorithm and partition information, and allocates the Map task to the TaskTracker to complete execution.

And the TaskTracker sends a heartbeat packet to the JobTracker according to a fixed interval time to inform the JobTracker of the running state, meanwhile, the heartbeat packet contains other information such as the progress of the completion of the current Map task and the like, and when the JobTracker receives the completion information of the operation task, the current operation state is set to be 'completed'. If the JobTracker does not receive the heartbeat packet returned by the TaskTracker within the fixed interval time, the JobTracker judges that the TaskTracker fails, and meanwhile, the job tasks of the TaskTracker are dispatched to other TaskTracker to be continuously executed.

The TaskTracker reads the job resource file from the HDFS. And after the TaskTracker obtains the tasks distributed by the JobTracker, acquiring Job resources from the HDFS.

After the resources are obtained, the TaskTracker starts the JVM subprocess and starts to run the task.

Job is divided into data blocks of the same size (64 MB by default) by the block size of the HDFS, and the user Job programs corresponding to the data blocks are executed.

S4, the Master node (Master) finds and prepares available Map nodes for the operation block, and transmits the data block to the Map nodes. And meanwhile, the master node also finds and allocates available Reduce nodes for the data blocks and transmits the data blocks to the Reduce nodes.

The main node starts each Map node to execute a program, allocates a Map task to a TaskTracker containing a data block processed by the Map, and copies a JAR packet during program operation to the TaskTracker for operation, so that each Map node reads local data as much as possible to calculate.

And S5, each Map node processes the read data blocks, combines and sorts the data, and stores the data on the local machine. And meanwhile, the master node is informed of the completion of the calculation task and the storage position of the intermediate result data of the master node, and after all Map nodes such as the master node and the like are calculated, the Reduce node is started to run. And the Reduce node acquires the position information of the intermediate result data from the main node and reads the data.

S6, the Reduce node sorts and combines the calculation results, and finally an output file is obtained through collection, namely the sensor classification result of the rape waterlogging hyperspectral image.

S7, using 75% of the spectral image matrix as input, operating the parallel multilayer perceptron algorithm, performing algorithm training, establishing a perceptron algorithm model, adjusting training parameters, and preventing the algorithm from over (under) fitting.

And S8, performing algorithm prediction by taking the residual 25% of spectral image data as input, and detecting the rape stain spectral image.

Training parameters are continuously adjusted in the training process, and the training effect is optimal when the training parameters are 600 input layers, 1 middle layer and 3 output layers.

After obtaining the spectrum classification result and the image classification result (classification detection analysis result of the computer image data), calculating the detected classification result by adopting the weighted harmonic mean, and calculating the final hyperspectral image classification detection result of the oil vegetable.

Specific harmonic weighted average:

wherein R is_nIs a harmonic mean, x_iIs the ith classification result, and n is the number of classification results.

As shown in fig. 5 to 14, the embodiment also discloses an image collecting device for collecting images of field rape waterlogging stress blades in the rape waterlogging hyperspectral image detection method based on the Spark platform, which comprises an unmanned aerial vehicle 1, a mooring platform 2 and an image collector 3, wherein the unmanned aerial vehicle 1 is used for transferring the image collector 3, and the bottom of the unmanned aerial vehicle is provided with an electromagnet;

the image collector 3 is used for collecting images of field rape waterlogging stress blades of field crops, the top of the image collector 3 is provided with a cylindrical magnetic metal block 32 matched with an electromagnet, the bottom of the image collector 3 is provided with a movable female connector 33, the image collector 3 comprises a case 31, the cylindrical magnetic metal block 32, the female connector 33 and a plurality of camera modules 34, the case 31 is cylindrical, the cylindrical magnetic metal block 32 is installed at the top of the case 31, the camera modules 34 are installed on the side surface of the case 31 and are arranged in the circumferential direction, the bottom of the case 31 is provided with a first sliding groove 311 matched with the female connector 33, the female connector 33 is connected in the first sliding groove 311 in a sliding manner, and when no obstacle exists at the bottom of the case 31, the female connector 33 is located at the lowest end relative to the case 31 under the action of self weight and extends to the outer side of the case 31;

the mooring platforms 2 are arranged in the field, specifically, one mooring platform 2 is arranged at each of four corners and the center of the field to realize the overall coverage of each field, the mooring platforms 2 are provided with locking assemblies 24 for fixing the image collectors 3 and sub-connectors 35 matched with the female connectors 33, the mooring platforms 2 are used for temporarily placing the image collectors 3, each mooring platform 2 comprises a base 21, a shell 22, a stand 2327, a locking assembly 24, a plugging assembly 25 and an adsorption assembly 26, the base 21 is arranged at the upper end of the shell 22, a through hole 221 for connecting equipment is formed in the geometric center of the base, the shell 22 is arranged on a fixture arranged in the field by means of the stand 2327, the locking assembly 24 is arranged in the shell 22 and used for fixing the image collectors 3 placed on the base 21, the plugging assembly 25 is arranged at the upper end of the locking assembly 24 and can plug the through hole 221, the sub-connector 35 is installed in the plugging component 25, the adsorption component 26 is connected to the lower end of the base 21 in a sliding manner and is connected with the locking component 24, the sub-connector 35 in the plugging component 25 can be adsorbed by the suction force generated by the adsorption component 26 and is connected with the female connector 33 so as to connect the antenna connector of the wireless transceiver module in the image collector 3 with the external antenna 28 on the mooring platform 2, and therefore the signal transmission capability of the image collector 3 and the management center is enhanced;

the locking assembly 24 comprises clamping blocks 241, connecting rods 242, a connecting seat 243 and an electric push rod 244, wherein the base 21 is provided with three sliding holes 245 matched with the clamping blocks 241, the clamping blocks 241 are respectively connected in the three sliding holes 245 in a sliding manner, the lower end of the clamping block 241 is rotatably connected with the connecting rods 242, the other end of the connecting rod 242 is rotatably connected to the connecting seat 243, the lower end of the connecting seat 243 is connected to the output end of the electric push rod 244, the electric push rod 244 is installed on a base 246 arranged in the shell 22, and a power module is arranged in the shell 22 and used for supplying power to the electric push rod 244;

the plugging component 25 comprises a plug 251 and a flange step 252, the lower end of the plug 251 is connected with the connecting seat 243, the flange step 252 is fixed on the periphery of the plug 251, and the upper end of the plug 251 is provided with a second sliding groove 253 for the sub-connector 35 to slide up and down.

In this embodiment, the sub-connector 35 includes a lower connecting block 351, an upper connecting block 352, a ring magnet 353 and a metal rod 355, the upper connecting block 352 is fixed on the lower connecting block 351, the ring magnet 353 is embedded in the periphery of the upper end of the lower connecting block 351, an installation groove 354 is arranged at the upper end of the upper connecting block 352, the metal rod 355 is installed in the installation groove 354, and a waterproof rubber plug 356 with a cross-shaped notch is installed at an opening at the upper end of the installation groove 354, so that a good waterproof effect is achieved;

the female connector 33 includes a sliding base 331 and a connecting sleeve 332, the bottom of the case 31 is provided with a third sliding slot 333 for the sliding base 331 to slide, the bottom of the sliding base 331 is provided with a fourth sliding slot 334 for the upper connecting block 352 to pass through, the connecting sleeve 332 is matched with a metal rod 355 and is installed in the fourth sliding slot 334, the metal rod 355 can be inserted into the connecting sleeve 332, the connecting sleeve 332 is connected to an antenna connector of a wireless transceiver module in the image collector 3, and the metal rod 355 is connected to the external antenna 28 on the mooring platform 2.

In this embodiment, the side of the sliding base 331 is provided with an annular groove 335, the adsorption assembly 26 includes three arc-shaped seats 261 and arc-shaped magnets 262, the arc-shaped seats 261 have three and are in one-to-one correspondence with the three clamping blocks 241, the upper ends of the arc-shaped seats 261 are slidably connected in the guide grooves 264 at the lower end of the base 21 by means of sliding columns, the arc-shaped seats 261 are further connected with the clamping blocks 241 by means of connecting plates 263, the arc-shaped magnets 262 are involved in the arc-shaped seats 261, and when the three arc-shaped seats 261 are folded inwards under the action of the clamping blocks 241, the three arc-shaped magnets 262 can form a complete annular magnet, thereby generating an attraction force on the annular magnet 353 on the lower connecting block 351, and driving the sub-connecting head 35 to move upwards and finally connected with the female connecting head 33.

The image acquisition process in this embodiment is as follows:

firstly, the unmanned aerial vehicle 1 takes out the image collectors 3 from the equipment warehouse one by one, and sends each image collector 3 to the corresponding mooring platform 2 according to an instruction sent by a management center, after the unmanned aerial vehicle 1 reaches the position right above the mooring platform 2, the unmanned aerial vehicle 1 descends to the height, finally, the image collectors 3 are placed at the upper end of the mooring platform 2 and located between the clamping blocks 241, and then the unmanned aerial vehicle 1 flies away; when the image collector 3 is placed, the position accuracy of the image collector 3 may not be high, at this time, a pressure sensor arranged on the base 21 senses the device and starts the electric push rod 244 by means of a controller inside the housing 22, the output end of the electric push rod 244 contracts, the three clamping blocks 241 gradually contract inwards until the image collector 3 is locked and fixed, and at this time, the female connector 33 is located at the through hole 221 on the base 21 and is located at the lowest end under the action of self weight; in the process, the arc-shaped seat 261 is communicated with the arc-shaped magnet 262 and is also folded inwards along with the arc-shaped magnet 262 until the three arc-shaped magnets 262 synthesize an annular magnet, the female connector 33 can be clamped and positioned by means of the annular magnet, on the other hand, the sub-connector 35 positioned in the plug 251 is adsorbed by utilizing the magnetic attraction force, and the connection with the female connector 33 is completed, so that the antenna connector of the wireless transceiver module in the image collector 3 is connected with the external antenna 28 on the mooring platform 2, and the quick installation of the image collector 3 and the mooring platform 2 can be completed. Thus, the second and third … … Nth image collectors 3 are installed, and then the spectral information of the crops in the peripheral field is collected by the camera module 34 and is remotely sent to the analysis platform of the management center through the wireless transceiver module.

When the image collector 3 needs to be recovered, the electric push rod 244 extends out and resets to unlock the image collector 3, the unmanned aerial vehicle 1 falls right above the image collector 3, the electromagnet at the bottom of the unmanned aerial vehicle is started to suck the image collector 3 and transfer the image collector into an equipment warehouse for storage, and the electric push rod 244 continues to extend out until the plug 251 reaches the highest point and blocks the through hole 221.

It will be appreciated by those skilled in the art that the invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The embodiments disclosed above are therefore to be considered in all respects as illustrative and not restrictive. All changes which come within the scope of or equivalence to the invention are intended to be embraced therein.

Claims

1. A rape waterlogging hyperspectral image detection method based on a Spark platform is characterized by comprising the following steps:

s2, correcting and preprocessing the collected image, converting the color channel of the image, and converting RGB into La^*b^*Converting the RGB image into a three-dimensional matrix, and adopting a principal component analysis method and extracting image characteristic information to reduce data dimensionality and reduce data calculation amount;

extracting RGB image information from the collected stain hyperspectral image, converting the RGB image into an HSI channel image by using color channel conversion, extracting hue, brightness and saturation information from the HIS image, and modifying the brightness and saturation channel histogram information to enhance the HSI image;

s3, performing distributed storage and management on the image matrix by using the HDFS, wherein a Master node is used as a Namenode node to manage the whole HDFS file system and the directory tree, and one copy is stored on other 5 Datanode nodes;

s4, constructing a parallel neural network algorithm on the main node;

s7, dividing the image data into a training set and a prediction set, wherein the data volume of the training set accounts for 75%, establishing a neural network training model, and performing parameter tuning;

s8, inputting the remaining 25% of data sets after the algorithm training is finished, performing algorithm prediction, and performing model performance evaluation to obtain a classification detection analysis result of the image data;

after the classification detection analysis result of the image data is obtained, calculating the detected classification result by adopting a weighted harmonic mean, and calculating the final hyperspectral image classification detection result of the rape, wherein the specific harmonic weighted mean is as follows:

wherein R is_nIs a harmonic mean, x_iThe number of classification results is the ith classification result, and n is the number of classification results;

before image acquisition, the following work is required:

s01, establishing a corresponding relation between the hyperspectral image based on rape waterlogging and different waterlogging stress grades, and predicting the level of the rape waterlogging stress;

s02, establishing a rape waterlogging damage hyperspectral image prediction model based on a Spark parallel computing framework to obtain parameters of the prediction model;

before forecasting rape waterlogging stress, hyperspectral images of rape leaves stressed by waterlogging in the flowering season are collected, the waterlogging stress lasts for 3 days and 6 days respectively, and hyperspectral image information corresponding to different waterlogging grades of rape is collected.

2. An image acquisition device adopted in the Spark platform based rape waterlogging hyperspectral image detection method is characterized by comprising an unmanned aerial vehicle (1), a mooring platform (2) and an image collector (3), wherein the unmanned aerial vehicle (1) is used for transferring the image collector (3) and the bottom of the unmanned aerial vehicle is provided with an electromagnet;

the image collector (3) is used for collecting images of the rape waterlogging stress leaves in the field, the top of the image collector is provided with a cylindrical magnetic metal block (32) matched with the electromagnet, and the bottom of the image collector (3) is provided with a movable female connector (33);

mooring platform (2) have a plurality ofly and install in the field, mooring platform (2) are gone up and are installed locking Assembly (24) that are used for fixed image collector (3) and with female connector (33) matched with sub-connector (35), mooring platform (2) are used for temporarily placing image collector (3).

3. The image acquisition device according to claim 2, wherein the image acquisition device (3) comprises a case (31), a cylindrical magnetic metal block (32), a female connector (33) and a plurality of camera modules (34), the case (31) is cylindrical, the cylindrical magnetic metal block (32) is installed at the top of the case (31), the camera modules (34) are installed at the side of the case (31) and are arranged circumferentially, a first chute (311) matched with the female connector (33) is arranged at the bottom of the case (31), the female connector (33) is slidably connected in the first chute (311), and when no obstacle exists at the bottom of the case (31), the female connector (33) is located at the lowest end relative to the case (31) under the action of self weight and extends to the outer side of the case (31).

4. The image acquisition device according to claim 2, wherein the mooring platform (2) comprises a base (21), a housing (22), columns (23) (27), a locking assembly (24), a blocking assembly (25) and a suction assembly (26), the base (21) is installed at the upper end of the housing (22) and is provided with a through hole (221) for equipment connection at the geometric center thereof, the housing (22) is installed on a fixture arranged in the field by means of the columns (23) (27), the locking assembly (24) is installed inside the housing (22) and is used for fixing the image collector (3) placed on the base (21), the blocking assembly (25) is installed at the upper end of the locking assembly (24) and can block the through hole (221), the sub-connector (35) is installed in the blocking assembly (25), the suction assembly (26) is slidably connected to the lower end of the base (21) and is connected with the locking assembly (24), the suction force generated by the adsorption component (26) can adsorb the sub-connector (35) in the plugging component (25) and is connected with the female connector (33) so as to connect the antenna connector of the wireless transceiver module in the image collector (3) with the external antenna (28) on the mooring platform (2).

5. The image capturing device as claimed in claim 3, wherein the locking assembly (24) includes a clamping block (241), a connecting rod (242), a connecting seat (243), and an electric push rod (244), the base (21) is provided with three sliding holes (245) matching with the clamping block (241), the clamping block (241) has three sliding holes (245) and is slidably connected to the three sliding holes (245), the lower end of the clamping block (241) is rotatably connected to the connecting rod (242), the other end of the connecting rod (242) is rotatably connected to the connecting seat (243), the lower end of the connecting seat (243) is connected to an output end of the electric push rod (244), and the electric push rod (244) is mounted on a base (246) disposed inside the housing (22).

6. The image acquisition device according to claim 4, wherein the blocking assembly (25) comprises a plug (251) and a flange step (252), the lower end of the plug (251) is connected with the connecting seat (243), the flange step (252) is fixed on the periphery of the plug (251), and a second sliding groove (253) for the sub-connecting head (35) to slide up and down is formed in the upper end of the plug (251).

7. The image acquisition device according to claim 6, wherein the sub-connector (35) comprises a lower connecting block (351), an upper connecting block (352), a ring-shaped magnet (353) and a metal rod (355), the upper connecting block (352) is fixed on the lower connecting block (351), the ring-shaped magnet (353) is embedded on the periphery of the upper end of the lower connecting block (351), an installation groove (354) is formed in the upper end of the upper connecting block (352), the metal rod (355) is installed in the installation groove (354), and a waterproof rubber plug (356) with a cross-shaped notch is installed at an opening of the upper end of the installation groove (354);

female connector (33) are including slide (331) and adapter sleeve (332), the bottom of machine case (31) is equipped with and supplies slide (331) gliding spout three (333), and the bottom of slide (331) is equipped with and supplies the spout four (334) that connecting block (352) passed, adapter sleeve (332) and metal pole (355) cooperation are installed in spout four (334).

8. The image capturing device according to claim 7, characterized in that the side of the sliding base (331) is provided with an annular groove (335), the suction assembly (26) comprises an arc-shaped base (261) and arc-shaped magnets (262), the arc-shaped base (261) has three and one-to-one correspondence with the three clamping blocks (241), the upper end of the arc-shaped base (261) is slidably connected in a guide groove (264) at the lower end of the base (21) by means of a sliding column, the arc-shaped base (261) is further connected with the clamping blocks (241) by means of a connecting plate (263), the arc-shaped magnets (262) are involved in the arc-shaped base (261), and the three arc-shaped magnets (262) can form a complete ring magnet.