A soil texture detection method and device based on machine learning and image processing

Abstract

The present invention discloses a method and device for detecting soil texture based on machine learning and image processing. The method includes the following steps: collecting images of soil samples; obtaining image feature data on the surface of the soil samples; randomly dividing the image feature data of the soil samples into a training set and a test set at a ratio of 8:2; performing standardization processing on the image feature data of the training set and the test set; using a genetic algorithm to perform feature selection on the image feature data to obtain optimal features; using a BP neural network, with the accurate value of the soil texture data as the target, training the training set to obtain a regression model for soil texture detection; calculating the soil texture data of the soil to be detected through the regression model. This method extracts features from the information on the soil surface based on image processing and uses a genetic algorithm to screen the features, and establishes a regression model for soil texture detection based on machine learning, realizing the rapid detection of soil texture data with high detection efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural soil detection, and particularly relates to a soil texture detection method and device based on machine learning and image processing. Background Art

[0002] Soil texture is one of the important physical properties of soil. The soil texture comprehensively reflects the physical condition of the soil and is the most basic trait of the soil. There is a close relationship between soil texture and water, fertilizer, air, heat in the soil and the soil's fertilizer supply capacity. Soil is the growth environment of plant roots. Soil texture mainly affects the root distribution of crops and affects the growth and development of plants.

[0003] Soil texture is classified according to the content of particle sizes in the soil. According to the American soil texture classification method, the triangular coordinate diagram method can be used. The three sides of the equilateral triangle respectively represent the contents (%) of clay (<0.002 mm), silt (0.002 - 0.05 mm) and sand (0.05 - 2 mm). Among them, sandy soil has good permeability, many large pores and few small pores. The pores are often filled with air, making it prone to drought and having poor fertilizer retention; clayey soil has small pores between soil particles, poor ventilation and water permeability, is prone to waterlogging, and has poor tillage properties.

[0004] In the prior art, when detecting soil texture, the hydrometer method is usually adopted. During the detection, after the preset sedimentation time arrives, technicians need to timely read and record the readings on the hydrometer. There are multiple groups of preset sedimentation times, and technicians need to wait to record the corresponding hydrometer readings; during this process, the experimental waiting time is long, time-consuming and laborious. When the number of sample soil types is large, it is even more complicated and extremely prone to errors. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above-mentioned existing problems and provide a soil texture detection method based on machine learning and image processing. This method extracts features of the information on the soil surface based on image processing, and establishes a soil texture detection model based on machine learning to achieve rapid detection of the contents of various soil textures with high detection efficiency.

[0006] Another purpose of the present invention is to provide a soil texture detection device based on machine learning and image processing.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A soil texture detection method based on machine learning and image processing includes the following steps:

[0009] (1) Use an image acquisition device to acquire images of soil samples with known soil texture data;

[0010] (2) The image processing module processes the image to obtain the image feature data on the surface of the soil sample;

[0011] (3) Randomly divide the image feature data of the soil sample into a training set and a test set at a ratio of 8:2;

[0012] (4) The feature engineering data processing module standardizes the image feature data of the training set and the test set so that different image feature data have the same scale;

[0013] (5) Use the genetic algorithm to perform feature selection on the image feature data, screen out the better features and remove the worse features, so as to achieve the optimal result and obtain the optimal features;

[0014] (6) The regression model module uses the BP neural network, takes the accurate value of the soil texture data of the soil sample as the target, trains the training set to obtain the regression model for soil texture detection, and evaluates the regression model through the test set, and judges the quality of the regression model based on the obtained results;

[0015] (7) The image acquisition device is used to acquire images of the soil to be detected, the image processing module processes the images to obtain the image feature data of the soil to be detected, and the soil texture data of the soil to be detected is calculated through the regression model.

[0016] A preferred solution of the present invention, wherein in step (1), the soil sample is placed in a detection box, and an industrial camera is used to acquire images of the soil sample.

[0017] Preferably, in step (2), the image feature data includes color features and texture features; the image processing module extracts the color features and texture features of the image, and the extracted color features include the mean values of the RGB color channels, the mean values of the HSV color channels, and the mean values of the LAB color channels; the extracted color feature texture features include the gray-level co-occurrence matrix and the gray-level binary pattern; for the image feature data of all the acquired images, a soil sample feature library is constructed.

[0018] Preferably, in step (3), the image feature data of the soil sample is randomly divided into a training set and a test set at a ratio of 8:2 from the soil sample feature library.

[0019] Preferably, in step (4), Z-score standardization is used to standardize the image feature data. After processing, the feature data conforms to a normal distribution, and the formula is as follows:

[0020]

[0021] In the formula: x *The image feature data after standardization; x is the actual value of each feature parameter; μ is the average value of a certain feature variable; σ is the standard deviation of a certain feature variable.

[0022] Preferably, in step (5), the genetic algorithm is used to perform feature selection on the image feature data, screen out better features, and remove poorer features, so as to achieve the optimal result. The specific steps to obtain the optimal features are as follows: Take the soil texture data as the optimization target, take the standardized image feature data as genes, and perform binary coding. In the binary coding, the total number of "0" and "1" is the same as the total number of features included in a single sample. Among them, the number of genes "1" is the number of features to be selected, and the selected features are used to generate a feature subset; then randomly generate an initial population; the fitness function in the genetic algorithm is used to judge the quality of individuals in the population, and the correlation coefficient R in the BP neural network 2 is used as the fitness function; during the operation of the genetic algorithm, when the fitness value result of the optimal feature combination no longer rises, the operation is terminated and the optimal features are output; otherwise, the genes continue to be selected, crossed, and mutated.

[0023] Preferably, in step (6), the training model of the BP neural network is set with a total of 4 layers, namely the input layer, the first hidden layer, the second hidden layer, and the output layer. Among them, the first hidden layer is set with 200 nodes, the second hidden layer is set with 100 nodes, and the output result of the output layer is 3 variables. The first variable is the content of clay in the soil sample, the second variable is the content of silt in the soil sample, and the third variable is the content of sand in the soil sample.

[0024] Preferably, in step (6), the random gradient descent method is used to optimize and solve the weights of the BP neural network, and the number of iterations is 1000 times.

[0025] A soil texture detection device based on machine learning and image processing includes an image acquisition device for collecting soil sample images, an image processing module for processing the images to obtain image feature data, a feature engineering data processing module for performing standardization processing on the image feature data, a feature selection module for performing feature selection on the image feature data using the genetic algorithm, and a regression model module for establishing a regression model; wherein, the image acquisition device is connected to the image processing module, the image processing module is connected to the feature engineering data processing module, the feature engineering data processing module is connected to the feature selection module, and the feature selection module is connected to the regression model module.

[0026] Preferably, the image acquisition device includes a detection box, an industrial camera, and an LED light source. Among them, the detection box provides a dark environment for detection. The industrial camera is used to collect images of soil samples in the detection box and send the collected images to the image processing module. The LED light source provides light for the detection environment.

[0027] Preferably, the image processing module extracts the color features and texture features of the image through opencv-python.

[0028] The present invention has the following beneficial effects compared with the prior art:

[0029] 1. The soil texture detection method based on machine learning and image processing in the present invention collects images of soil samples with known soil texture data, extracts image feature data from the images, standardizes the image feature data, and then performs feature selection on the image feature data through a genetic algorithm to select better features. A regression model for soil texture detection is trained using a BP neural network. A detection model for soil texture detection prediction can provide a detection model for subsequent soil texture detection. By collecting images of the soil to be detected through an image acquisition device, extracting image feature data from the images through an image processing module, and using the extracted image feature data as variables for input, soil texture data can be obtained, reducing the time spent on training; achieving rapid detection of the contents of various soil textures, with high detection efficiency, and greatly reducing manpower and material resources.

[0030] 2. The soil texture detection method based on machine learning and image processing in the present invention uses a genetic algorithm to perform feature selection on image feature data, screen out better features to make the result optimal, and generate a feature subset of the selected better features. A regression model for soil texture detection is trained using a BP neural network for the feature subset. Through the genetic algorithm to optimize the feature selection of the extracted features, select the feature variables useful for the training model, and improve the prediction accuracy of soil texture. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic structural diagram of a specific implementation manner of a soil texture detection device based on machine learning and image processing in the present invention.

[0032] Figure 2 It is a schematic flow diagram of extracting image feature data of soil samples in the present invention.

[0033] Figure 3 It is a schematic flow diagram of genetic algorithm feature selection in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] To enable those skilled in the art to well understand the technical solution of the present invention, the present invention will be further described below in conjunction with embodiments and drawings, but the implementation manners of the present invention are not limited thereto.

[0035] Embodiment 1

[0036] Refer to Figures 1 - 3 , this embodiment discloses a soil texture detection method based on machine learning and image processing, including the following steps:

[0037] (1) Image acquisition device is used to acquire images of soil sample 4 with known soil texture data;

[0038] (2) The image processing module processes the images to obtain image feature data on the surface of soil sample 4;

[0039] (3) Randomly divide the image feature data of the soil sample into a training set and a test set in a ratio of 8:2;

[0040] (4) The feature engineering data processing module standardizes the image feature data of the training set and the test set so that different image feature data have the same scale;

[0041] (5) Use the genetic algorithm to perform feature selection on the image feature data, screen out better features, remove poor features, so as to achieve the optimal result and obtain the optimal features;

[0042] (6) The regression model module uses the BP neural network, takes the accurate value of the soil texture data of the soil sample as the target, trains the training set of the optimal features to obtain a regression model for soil texture detection, and evaluates the regression model through the test set of the optimal features, and judges the quality of the regression model based on the obtained results;

[0043] (7) The image acquisition device is used to acquire images of the soil to be detected, the image processing module processes the images to obtain image feature data of the soil to be detected, and the soil texture data of the soil to be detected is calculated through the regression model.

[0044] In step (1), the soil sample 4 is a soil sample in the soil sample database, and the soil texture data of this soil sample is known, that is, the content of clay particles, the content of silt particles, and the content of sand particles of the known soil sample are known.

[0045] Refer to Figure 1 , in step (1), the soil sample 4 is placed in the detection box 3, and the industrial camera 1 is used to acquire images of the soil sample 4.

[0046] Refer to Figure 2, in step (2), the image feature data includes color features and texture features; the specific steps for the image processing module to process the image and obtain the image feature data on the surface of the soil sample 4 are as follows: the image processing module extracts the color features and texture features of the image. The extracted color features include the mean values of the RGB color channels, the mean values of the HSV color channels, and the mean values of the LAB color channels; the extracted texture features of the color features include the gray-level co-occurrence matrix and the gray-level binary pattern; for the image feature data of all the collected images, a soil sample feature library is constructed.

[0047] See Figure 2 , in this embodiment, extracting the features of one image is a set of image feature data, and there are a total of 45 features in a set of image feature data. The soil sample feature library contains multiple sets of image feature data. There are multiple soil samples, and the soil texture data of each sample is different. One image can be collected for each soil sample, or multiple images can be collected. After standardizing the features, optimize them using the genetic algorithm, select the soil texture feature variables that have the greatest impact on the results, and finally obtain 19 feature variables. The 45 feature parameters here respectively refer to the mean value of the R channel, the mean value of the G channel, the mean value of the B channel, the mean value of the H channel, the mean value of the S channel, the mean value of the V channel, the mean value of the L channel, the mean value of the A channel, the mean value of the B channel, the mean value of the gray-level co-occurrence matrix, the variance of the gray-level co-occurrence matrix, the contrast of the gray-level co-occurrence matrix, the entropy of the gray-level co-occurrence matrix, the energy of the gray-level co-occurrence matrix, the inverse variance of the gray-level co-occurrence matrix, the correlation of the gray-level co-occurrence matrix, the standard deviation of the gray-level co-occurrence matrix, the homogeneity of the gray-level co-occurrence matrix, the difference of the gray-level co-occurrence matrix, the pixel amounts of the gray-level binary pattern (LBP) grays 1 to 26; the finally obtained 19 features include the mean value of the G channel, the mean value of the R channel, the mean value of the H channel, the mean value of the L channel, the mean value of the gray-level co-occurrence matrix, the variance of the gray-level co-occurrence matrix, the inverse variance of the gray-level co-occurrence matrix, the contrast of the gray-level co-occurrence matrix, and the pixel amounts of the LBP gray values 2, 3, 5, 8, 11, 13, 15, 16, 17, 24, 25.

[0048] In step (3), randomly divide the image feature data of the soil samples in the soil sample feature library into a training set and a test set at a ratio of 8:2. Specifically, randomly divide multiple sets of image feature data in the soil sample feature library into a training set and a test set at a ratio of 8:2. If 10 images of different soil samples are collected and 10 sets of image feature data are extracted, 8 sets of image feature data are used as the training set and 2 sets of image feature data are used as the test set.

[0049] In step (4), use Z-score standardization to standardize the image feature data. After processing, the feature data conforms to a normal distribution, and its formula is as follows:

[0050]

[0051] where: x * is the standardized image feature data; x is the actual value of each feature parameter; μ is the average value of a certain feature variable; σ is the standard deviation of a certain feature variable.

[0052] See Figure 3 , in step (5), the genetic algorithm is used to perform feature selection on the image feature data, screening out better features and removing worse features. The specific steps to achieve the optimal are as follows: taking the soil texture data as the optimization target, taking the standardized image feature data as genes, and performing binary encoding. In the binary encoding, the total number of "0" and "1" is the same as the total number of features included in a single sample. Among them, the number of genes "1" is the number of features to be selected, and the selected features are used to generate a feature subset; then an initial population is randomly generated; the fitness function in the genetic algorithm is used to judge the quality of individuals in the population, and the correlation coefficient R 2 in the BP neural network is used as the fitness function; during the operation of the genetic algorithm, when the fitness value result of the optimal feature combination no longer rises, the operation is terminated and the optimal features are output; otherwise, the genes continue to be selected, crossed, and mutated.

[0053] In the above generation of the initial population, the initialized population tries to ensure gene diversity as much as possible. Here, 100 randomly generated individuals are selected as the initialized population.

[0054] See Figure 3 , after the setting of the initial population is completed, a fitness value evaluation test is carried out. Calculate the fitness of the new individuals generated by the exchange. The fitness can be used to measure the quality of individuals in the population. The fitness here is the criterion value of the feature combination. The specific fitness function is as follows. Use the BP neural network to train and test the extracted feature subset. Among them, the training set is used to construct the regression model for soil texture detection, and the test set is used to verify the performance of the regression model. When predicting the test set, the closer the correlation coefficient R 2 is to 1, the better the individual. After the fitness value evaluation test, it is judged whether the termination condition is met. Here, it is set that when the fitness value result of the optimal feature combination selected by the genetic algorithm no longer rises, the operation is terminated and the features with the best results are output; for the features that do not meet the termination condition, gene operations are carried out. The specific gene operations are as follows:

[0055] (a) Selection; Selection is to select excellent individuals from the exchanged population so that they have more opportunities to be the parents and reproduce offspring for the next generation. The principle of selection is that individuals with strong adaptability (that is, R 2Individuals closer to 1 have a greater probability of contributing to the next generation, and selection implements Darwin's principle of survival of the fittest. The selection operator used here is the tournament selection method, whose core idea is: in each evolution process, a certain number of individuals are randomly selected from the parent population, and the individual with the highest fitness is selected from these individuals for genetic operations. This process is repeated until the size of the offspring population is the same as that of the parent population. Through this process, the diversity of the population is greatly maintained;

[0056] (b) Setting of the crossover operator; The crossover operator is the main operation method for generating new population individuals. The crossover operator endows the genetic algorithm with global search ability; specifically, for each pair of parents selected by the crossover probability (P c ), the dissimilar partial genes are exchanged to generate new individuals. The new generation of individuals obtained inherit the excellent characteristics of their parents. Usually, the value of the crossover probability is between 0.5 and 0.9. The crossover probability (P c ) here is set to 0.5;

[0057] (c) Mutation; Mutation randomly changes a certain gene value in the chromosome with a certain probability (Pm). Mutation provides opportunities for the generation of new individuals. Usually, the value is between 0.001 and 0.01. The mutation probability (Pm) here is set to 0.05.

[0058] In step (6), a regression model for detecting soil texture is established using a BP neural network. In the detection, the optimal feature parameters input are used as labels, and the accurate value of the soil texture data obtained by sending soil sample 4 to the laboratory for detection is used as the target for model training. The training model of the BP neural network is set with 4 layers, namely the input layer, the first hidden layer, the second hidden layer, and the output layer. Among them, the first hidden layer is set with 200 nodes, the second hidden layer is set with 100 nodes, and the output result of the output layer is 3 variables. The first variable is the clay content of soil sample 4, the second variable is the silt content of soil sample 4, and the third variable is the sand content of soil sample 4; the input layer is used to input image feature data. The random gradient descent method is used to optimize and solve the weights of the BP neural network, and the number of iterations is set to 1000 times.

[0059] As shown in Table 1, MSE in Table 1 is a more convenient method for measuring the "average error". MSE can evaluate the degree of data change. The smaller the value of MSE, the better the accuracy of the prediction model in describing experimental data. Through comparative analysis, it can be seen from the effects of using genetic algorithm for feature selection combined with BP neural network and simply using BP neural network that the recognition efficiency and stability of the BP neural network combination optimized by genetic algorithm feature selection in this embodiment have been improved.

[0060] Table 1 Soil Texture Prediction and Evaluation Table

[0061]

[0062] After the above model is established, the image of the soil to be detected is collected, and specific features are extracted as model variables and input into the model, so that the soil texture data of the soil to be detected can be detected. The soil texture detection method based on machine learning and image processing in this embodiment has a simple structure, small volume, convenient operation, time-saving and labor-saving, high online detection effect, and fast recognition speed.

[0063] Embodiment 2

[0064] See Figure 1 , this embodiment discloses a soil texture detection device based on machine learning and image processing, including an image acquisition device for collecting images of soil samples 4, an image processing module for processing images to obtain image feature data, a feature engineering data processing module for standardizing image feature data, a feature selection module for performing feature selection on image feature data using a genetic algorithm, and a regression model module for establishing a regression model; wherein, the image acquisition device is connected to the image processing module, the image processing module is connected to the feature engineering data processing module, the feature engineering data processing module is connected to the feature selection module, and the feature selection module is connected to the regression model module.

[0065] The feature engineering data processing module, the feature selection module, and the regression model module are all APPs or running programs on a computer.

[0066] See Figure 1 , the image acquisition device includes a detection box 3, an industrial camera 1, and an LED light source 2. Among them, the detection box 3 provides a dark environment for detection, the industrial camera 1 is used to collect images of the soil sample 4 in the detection box 3 and send the collected images to the image processing module, and the LED light source 2 provides light for the detection environment.

[0067] See Figure 1 , the LED light source 2 is an LED light strip arranged on the inner wall of the detection box 3. The detection box 3 is provided with a detection platform for placing the soil sample 4. The industrial camera 1 is installed on the top of the detection box 3. The industrial camera 1 is connected to a computer 5 through a data cable for image transmission; the computer 5 is provided with an image acquisition module and an image processing module. The image acquisition module is used to control the industrial camera 1 to take pictures and collect images of the soil sample 4, and the image processing module extracts the color features and texture features of the image through opencv-python.

[0068] The above is a preferred embodiment of the present invention. However, the embodiments of the present invention are not limited to the above content. Any other changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A soil texture detection method based on machine learning and image processing, characterized in that: The following steps are involved: (1) Capturing images of soil samples with known soil texture data using an image acquisition device; (2) The image processing module processes the image to obtain image feature data of the soil sample surface; (3) The image feature data of soil samples were randomly divided into training set and test set in a ratio of 8:2; (4) The feature engineering data processing module standardizes the image feature data of the training set and the test set so that different image feature data have the same scale; (5) Use genetic algorithms to select features from image feature data, filter out better features, and remove worse features, so as to achieve the best results and obtain the best features; (6) The regression model module uses the BP neural network to train the training set with the accurate value of the soil texture data of the soil sample as the target to obtain the regression model for soil texture detection, and then evaluates the regression model through the test set. The results obtained are used to judge the quality of the regression model. (7) The image acquisition device acquires images of the soil to be tested, and the image processing module processes the images to obtain image feature data of the soil to be tested, and calculates the soil texture data of the soil to be tested through the regression model; In step (5), a genetic algorithm is used to select features from the image feature data, screen out better features, and remove worse features, thereby achieving the optimal result. The specific steps for obtaining the optimal features are: using soil texture data as the optimization target, using the standardized image feature data as genes, and performing binary encoding. In the binary encoding, the total number of "0" and "1" is the same as the total number of features contained in a single sample, where the number of gene "1" is the number of features to be selected, and the selected features are used to generate a feature subset; Then the initial population is randomly generated; the fitness function in the genetic algorithm is used to judge the quality of individuals in the population, and the relationship coefficient R in the BP neural network is used 2 As the fitness function; in the operation of the genetic algorithm, when the fitness value of the optimal feature combination no longer increases, the operation is terminated and the optimal feature is output; otherwise, the gene continues to be selected, crossed and mutated.

2. The soil texture detection method based on machine learning and image processing according to claim 1, characterized in that: In step (1), a soil sample is placed in a detection box, and an industrial camera is used to capture images of the soil sample.

3. The soil texture detection method based on machine learning and image processing according to claim 1, characterized in that: In step (2), the image feature data includes color features and texture features; the image processing module extracts color features and texture features from the image, and the extracted color features include RGB color channel mean, HSV color channel mean, and LAB color channel mean; the extracted color features and texture features include grayscale co-occurrence matrix and grayscale binary pattern; for the image feature data of all collected images, a soil sample feature library is constructed.

4. The soil texture detection method based on machine learning and image processing according to claim 2, characterized in that: In step (3), the image feature data of the soil samples are randomly divided into a training set and a test set in a ratio of 8:2 from the soil sample feature library.

5. The soil texture detection method based on machine learning and image processing according to claim 1, characterized in that: In step (4), the image feature data is normalized using Z-score normalization. After processing, the feature data conforms to the normal distribution, and the formula is as follows: Where: is the standardized image feature data; is the actual value of each characteristic parameter; is the average value of a characteristic variable; is the standard deviation of a characteristic variable.

6. The soil texture detection method based on machine learning and image processing according to claim 1, characterized in that: In step (6), the BP neural network training model has four layers, namely the input layer, the first hidden layer, the second hidden layer and the output layer. The first hidden layer has 200 nodes, the second hidden layer has 100 nodes, and the output result of the output layer is three variables. The first variable is the clay content of the soil sample, the second variable is the silt content of the soil sample, and the third variable is the sand content of the soil sample.

7. The soil texture detection method based on machine learning and image processing according to claim 1, characterized in that: In step (6), the stochastic gradient descent method is used to optimize the weights of the BP neural network, and the number of iterations is 1000.

8. A soil texture detection device based on machine learning and image processing, characterized in that: The soil texture detection device is used to implement the soil texture detection method according to any one of claims 1 to 7, and the soil texture detection device includes an image acquisition device for acquiring soil sample images, an image processing module for processing images to obtain image feature data, a feature engineering data processing module for standardizing image feature data, a feature selection module for selecting features of image feature data using a genetic algorithm, and a regression model module for establishing a regression model; wherein the image acquisition device is connected to the image processing module, the image processing module is connected to the feature engineering data processing module, the feature engineering data processing module is connected to the feature selection module, and the feature selection module is connected to the regression model module.

9. The soil texture detection device based on machine learning and image processing according to claim 8, characterized in that: The image acquisition device includes a detection box, an industrial camera and an LED light source, wherein the detection box provides a dark environment for detection, the industrial camera is used to capture images of soil samples in the detection box and send the captured images to the image processing module, and the LED light source provides light for the detection environment.

Images