Sweet potato plot area division and background removal method based on unmanned aerial vehicle multispectral remote sensing

Abstract

The invention discloses a sweet potato plot area division and background removal method based on unmanned aerial vehicle multispectral remote sensing. According to the method, the random forest model is adopted for training, so that remarkable advantages are brought. Firstly, the comprehensiveness of data preparation and the accuracy of feature selection jointly improve the prediction accuracy of the model. By integrating multiple characteristics such as multispectral remote sensing data, vegetation indexes and topographic parameters, the model can comprehensively consider multiple factors influencing the growth of sweet potatoes, so that the results of cell division and background removal are predicted more accurately. The feature importance evaluation further screens out the feature which has the largest influence on the model, interference caused by irrelevant factors is reduced, the model is more focused on key information, and the prediction accuracy and reliability are improved. An integrated learning mechanism of multiple decision trees is combined with randomly selected features and samples, so that an overfitting problem is effectively avoided, and stable performance of the model on unknown data is ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of sweet potato planting, and specifically relates to a sweet potato plot area division and background removal method based on unmanned aerial vehicle multispectral remote sensing. Background Art

[0002] Sweet potato, also known as sweet potato or sweet potato, is a plant of the genus Ipomoea in the family Convolvulaceae. It is native to South America and was later introduced to my country and widely planted. Sweet potato is rich in starch, dietary fiber, vitamins, minerals and various amino acids, with high nutritional value. It has the effects of nourishing the body, strengthening the spleen and stomach, and preventing constipation. In my country, sweet potato is both a food crop and an important economic crop. The edible part of sweet potato is the underground tuber, which is bright in color and has a soft and glutinous taste. It can be boiled, steamed, baked and eaten in many ways, and is deeply loved by people. In modern agricultural planting management, monitoring and evaluating the growth of sweet potatoes is a key link in increasing yields and optimizing fertilization plans. In recent years, the development of drone technology has provided new solutions for agricultural remote sensing monitoring. Drones have the characteristics of flexible operation, low cost and high resolution, and can carry multispectral cameras to collect farmland images. However, how to accurately divide crop plots from these high-resolution images, remove background interference, and how to effectively analyze the differences in spectral characteristics between different treatment plots are still challenges facing the current agricultural remote sensing field.

[0003] However, existing image processing technologies often rely on manual operations when performing cell division and background removal, which is not only inefficient but also prone to introducing errors. At the same time, there is a lack of effective statistical analysis methods for the extracted spectral data to evaluate the impact of different agricultural management measures on crop growth. Summary of the invention

[0004] The purpose of the present invention is to provide a method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing in order to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: a method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing, the method comprising the following steps:

[0006] S1: According to the needs of the sweet potato fertilizer test, design the field experiment plan and determine the number, area, and arrangement parameters of the plots; arrange necessary markers in the field, such as white or black PVC boards, QR codes, and special reflective coatings, to facilitate drone image recognition;

[0007] S2: Use a drone equipped with a multispectral camera to fly in the field to collect multispectral images of sweet potato fields; set reasonable flight parameters, such as flight altitude and heading overlap rate, to ensure image quality and coverage; when collecting multispectral images, synchronously record flight track information;

[0008] S3: Use Pix4D or other image processing software to process the multispectral images collected by the drone. First, perform radiation correction to eliminate the influence of lens distortion and flight attitude on the image. Then perform image stitching to generate orthophotos. In this process, geometric correction is also required to ensure that the plane position accuracy of the image is at the centimeter level to meet the needs of subsequent analysis.

[0009] S4: Use the random forest model machine learning algorithm to automatically segment the drone imagery; train the model to identify the boundaries of plots in the field and automatically extract the image data of each plot;

[0010] S5: Using NDVI, EVI or other vegetation indices, perform threshold segmentation to remove soil and weed background pixels; extract the average value of the remaining vegetation pixels in each plot to obtain pure sweet potato canopy spectral feature data;

[0011] S6: Based on the results of cell division and background removal, extract the multispectral reflectance data of each cell; use statistical analysis methods to analyze the differences in spectral characteristics of different cells

[0012] S7: Combine spectral reflectance characteristics with machine learning algorithms to achieve parameter prediction of sweet potato growth conditions. Design a dynamic monitoring solution for the plot based on the time series change analysis of plot data.

[0013] S8: Store the processed multispectral images, cell division results, and background removal data in a database; establish a data management system to facilitate users to query, analyze, and download data.

[0014] In a preferred embodiment, in step S1, first, the number, area and arrangement of the experimental plots need to be determined according to the purpose and requirements of the sweet potato fertilizer test; usually, the number of plots should be sufficient for statistical analysis, with an area of ​​10-20 square meters; the arrangement usually adopts a random block design or a Latin square design; in the field, white or black PVC boards are arranged at regular intervals as reflective controls, and QR codes or special reflective paints are used to mark the boundaries of the plots so that the drone can accurately distinguish each plot during image recognition.

[0015] In a preferred embodiment, in step S2, before the UAV flies, it is necessary to set appropriate flight parameters according to the geographical location, terrain and weather conditions of the sweet potato field; the flight altitude is usually set to 50-100 meters to ensure sufficient image resolution; the heading overlap rate is set to 60%-80%, and the lateral overlap rate is set to 70%-80% to ensure the integrity and quality of image stitching; during the flight, the multispectral camera carried by the UAV will synchronously record image data and flight track information to ensure that the actual situation of the field can be accurately restored during subsequent data processing.

[0016] In a preferred embodiment, in step S3, the radiation correction includes:

[0017] Black and white board calibration: Place a full white board and a full black board with standard reflectivity at the same time at the shooting site; shoot the black and white boards before and after the drone flies to obtain their images;

[0018] Image acquisition: During the flight of the drone, multispectral image data is collected, and the camera exposure time and ISO parameters are recorded;

[0019] Radiometric calibration: Use black and white image to establish the conversion relationship between the image DN value and the actual reflectivity or radiant brightness; this is usually done using the following formula: L = a\timesDN+bL = a×DN+b, where L is the radiant brightness or reflectivity, DN is the digital number of the image, and a and b are calibration coefficients calculated using the black and white image;

[0020] Apply correction: Apply the above transformation relationship to all images and perform radiometric correction on each pixel to eliminate inconsistencies in camera response and atmospheric effects;

[0021] Geometric correction includes:

[0022] Control point selection: Select several control points on the ground that are easy to identify in the image. These points should be evenly distributed throughout the study area and have precise geographic coordinates.

[0023] Control point marking: accurately marking the location of these control points in the image, usually by clicking on the exact location of the control points on the image;

[0024] Geometric model selection: Select an appropriate geometric correction model, such as a polynomial model, an affine transformation model, or a projective transformation model, to describe the relationship between image pixels and the actual position on the ground;

[0025] Correction parameter calculation: Calculate the parameters of the geometric correction model using the ground coordinates and image coordinates of the control points;

[0026] Coordinate transformation: Apply the calculated correction parameters to resample the image to map each pixel to its correct ground position; this process usually involves the following steps:

[0027] Reprojection: Convert the pixel coordinates of the original image to coordinates in the new coordinate system;

[0028] Resampling: In the new coordinate system, the pixel values ​​are calculated by interpolation method to generate a corrected image.

[0029] In a preferred embodiment, in step S4, a random forest model is used for training, and the specific steps are:

[0030] Model training: Data preparation: Input features include multispectral remote sensing data from drones, sweet potato vegetation index, and terrain parameters; Data partitioning: Divide the dataset into training and validation sets with a ratio of 8:2 or 7:3;

[0031] Feature selection: Based on the feature importance assessment, the features with greater impact on sweet potato plot area division and background removal were selected;

[0032] Model construction: Build multiple decision trees. Each tree randomly selects some features and samples for training. Each tree is fully grown without pruning.

[0033] Parameter optimization: Optimize the parameters of random forest, including the number of trees, tree depth, and number of feature selections, through grid search or random search methods.

[0034] In a preferred embodiment, in step S4, the feature importance evaluation formula is:

[0035] Where:

[0036] Importance(f) indicates the importance of feature f;

[0037] N represents the number of decision trees;

[0038] ΔIi(f) represents the reduction in impurity of feature f in the i-th tree; the reduction in impurity is calculated using the Gini impurity and information gain indicators;

[0039] Grid search finds the optimal parameter settings by traversing all possible parameter combinations. The calculation formula is:

[0040] BestParams=argmaxθ∈ΘPerformance(Model(θ,X train ,y train ),X val ,y val )

[0041] In the formula

[0042] BestParamsBestParams represents the optimal parameter combination;

[0043] Θ represents the set of all possible parameter combinations;

[0044] Model(θ,X train ,y train ) means using parameter θ in the training set (X train ,y train )The model trained on

[0045] Performance represents the model performance evaluation function, including accuracy, recall, and F1 value;

[0046] (X val ,y val ) represents the validation set.

[0047] In a preferred embodiment, in step S5, threshold segmentation is performed, and a suitable threshold needs to be determined first; this threshold is determined by visual interpretation or statistical methods; a group of typical sweet potato vegetation pixels and a non-vegetation pixel are selected, their NDVI or EVI values ​​are calculated, and then the difference between them is used as the threshold; once the threshold is determined, the NDVI or EVI values ​​of all pixels are compared; those pixels greater than the threshold are considered to be vegetation pixels, and those less than the threshold are considered to be background pixels; in this way, the sweet potato vegetation is separated from other background elements.

[0048] In a preferred embodiment, in step S6, the specific process includes:

[0049] Data extraction: Extract multispectral reflectance data for each cell from the processed drone image; assuming there are n cells, each cell has spectral data of m bands;

[0050] Data preprocessing: To ensure data quality, perform the following preprocessing steps:

[0051] Outlier removal: Identify and remove abnormal data points due to noise or other factors;

[0052] Data normalization: Spectral data were normalized to the same range using the min-max normalization method.

[0053] Organize data structure: Organize the extracted spectral data into a matrix, where rows represent plots and columns represent spectral reflectance values ​​of different bands; ensure that the data of each plot corresponds to the corresponding fertilization treatment label;

[0054] Perform one-way ANOVA:

[0055] Hypothesis Testing:

[0056] Null hypothesis: The mean phase of the spectral reflectance of plots in all fertilization treatments;

[0057] Alternative hypothesis: The mean spectral reflectance of plots in at least one fertilization treatment is different from that of the other treatments;

[0058] Calculate statistics:

[0059] Calculate the within-group mean, between-group mean, and total mean;

[0060] Calculate within-group variance and between-group variance;

[0061] Calculate the F value, which is the ratio of the between-group variance to the within-group variance, F = MSB / MSE, where MSB is the between-group mean square and MSE is the within-group mean square;

[0062] Determine the significance level: Choose a significance level, usually 0.05;

[0063] Find the F distribution table: Find the critical value in the F distribution table based on the degrees of freedom and significance level;

[0064] Make a decision: If the calculated F value is greater than the critical value, reject the null hypothesis and think that different fertilization treatments have a significant effect on spectral reflectance; if the F value is not greater than the critical value, fail to reject the null hypothesis and think that different fertilization treatments have no significant effect on spectral reflectance;

[0065] Interpretation of results: If the null hypothesis is rejected, further multiple comparison analysis is performed to determine which specific treatments have significant differences; record and report the ANOVA F value, P value, mean and standard deviation statistical results of each group;

[0066] Where n represents the number of cells; m represents the number of spectral bands; MSB represents the mean square between groups; MSE represents the mean square within groups; F value represents the ratio of the variance between groups to the variance within groups; P value represents the probability value, which indicates the probability of observing the current result or a more extreme result.

[0067] In a preferred embodiment, in step S7, regular drone flights: according to a predetermined plan, regular drone flights are carried out on sweet potato fields to collect the latest multispectral image data; data processing: pre-processing of the collected data, including radiation correction, geometric correction, and vegetation index calculation; model updating: based on new data, the prediction model is continuously updated and adjusted to improve the accuracy and reliability of the model.

[0068] In a preferred embodiment, in step S8, the data management system includes the following aspects: user authority management: defining different user roles and corresponding authority levels to ensure that only authorized personnel can access sensitive data; regularly backing up important data in the database to prevent data loss in the event of an accident; and recording all data operation logs to facilitate tracking and analysis of potential security issues.

[0069] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0070] 1. In the present invention, the use of random forest model for training brings significant advantages. First, the comprehensiveness of data preparation and the accuracy of feature selection jointly improve the prediction accuracy of the model. By integrating multiple features such as multispectral remote sensing data, vegetation index and terrain parameters, the model can comprehensively consider the various factors affecting sweet potato growth, thereby more accurately predicting the results of cell division and background removal. Feature importance assessment further screens out the features that have the greatest impact on the model, reduces the interference caused by irrelevant factors, makes the model more focused on key information, and improves the accuracy and reliability of prediction.

[0071] 2. In the present invention, the construction of the random forest model and the parameter optimization process enhance the generalization ability and processing efficiency of the model. The integrated learning mechanism of multiple decision trees, combined with the features and samples selected randomly, effectively avoids the overfitting problem and ensures the stable performance of the model on unknown data. By optimizing the model parameters through grid search or random search method, not only the optimal parameter combination is found, the performance indicators of the model, such as accuracy, recall rate and F1 value, are improved, but also the overall efficiency of data processing is improved. These steps work together to provide strong technical support for the refined management of sweet potato planting, making agricultural production more scientific and intelligent. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a schematic diagram of the process principle of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0074] Example:

[0075] Reference Figure 1 A method for sweet potato plot area division and background removal based on UAV multispectral remote sensing, the method comprising the following steps:

[0076] S1: According to the needs of the sweet potato fertilizer test, design a field experiment plan and determine the number, area, arrangement and other parameters of the plots. Arrange necessary markers in the field, such as white or black PVC boards, QR codes, special reflective coatings, etc., to facilitate drone image recognition;

[0077] S2: Use a drone equipped with a multispectral camera to fly in the field and collect multispectral images of the sweet potato field. Set reasonable flight parameters, such as flight altitude, heading overlap rate, etc., to ensure image quality and coverage. When collecting multispectral images, synchronously record flight track information;

[0078] S3: Use Pix4D or other image processing software to process the multispectral images collected by the drone. First, perform radiation correction to eliminate the effects of lens distortion and flight attitude on the image. Then perform image stitching to generate an orthophoto. In this process, geometric correction is also required to ensure that the plane position accuracy of the image is at the centimeter level to meet the needs of subsequent analysis.

[0079] S4: Use the random forest model machine learning algorithm to automatically segment the drone imagery. The model is trained to identify the boundaries of plots in the field and automatically extract image data for each plot;

[0080] S5: Use NDVI, EVI or other vegetation indices to perform threshold segmentation and remove background pixels such as soil and weeds. Extract the average value of the remaining vegetation pixels in each plot to obtain pure sweet potato canopy spectral feature data;

[0081] S6: Based on the results of cell division and background removal, extract the multispectral reflectance data of each cell. Use statistical analysis methods to analyze the differences in spectral characteristics of different cells;

[0082] S7: Combine spectral reflectance characteristics with machine learning algorithms, such as random forest (RF) and partial least squares regression (PLSR), to achieve parameter prediction of sweet potato growth conditions (SPAD value, chlorophyll content, nitrogen content). Design a dynamic monitoring scheme for the plot based on the temporal change analysis of the plot data;

[0083] S8: Store the processed multispectral images, cell division results, background removal data, etc. in the database. Establish a data management system to facilitate users to query, analyze and download data.

[0084] In step S1, first, the number, area and arrangement of the experimental plots need to be determined according to the purpose and needs of the sweet potato fertilizer test. Usually, the number of plots should be sufficient for statistical analysis, with an area of ​​10-20 square meters. The arrangement usually adopts a randomized block design or a Latin square design. In the field, white or black PVC boards are arranged at regular intervals as reflective controls, and QR codes or special reflective paints are used to mark the boundaries of the plots so that the drone can accurately distinguish each plot during image recognition.

[0085] In step S2, before the drone flies, it is necessary to set appropriate flight parameters according to the geographical location, terrain and weather conditions of the sweet potato field. The flight altitude is usually set to 50-100 meters to ensure sufficient image resolution. The heading overlap rate is set to 60%-80%, and the lateral overlap rate is set to 70%-80% to ensure the integrity and quality of image stitching. During the flight, the multispectral camera carried by the drone will synchronously record image data and flight track information to ensure that the actual situation of the field can be accurately restored during subsequent data processing.

[0086] In step S3, radiation correction includes:

[0087] Black and white board calibration: Place a full white board (reflectivity is 100%) and a full black board (reflectivity is 0%) with standard reflectivity at the same time at the shooting site. Before and after the drone flies, take pictures of the black and white boards to obtain their images.

[0088] Image acquisition: During the flight of the drone, multispectral image data is collected, and the camera's exposure time, ISO and other parameters are recorded.

[0089] Radiometric calibration: Use black and white image to establish the conversion relationship between the image DN value (digital number) and the actual reflectivity or radiant brightness. This is usually done using the following formula: L = a\times DN+bL = a×DN+b, where L is the radiant brightness or reflectivity, DN is the digital number of the image, and a and b are calibration coefficients calculated using the black and white image.

[0090] Apply correction: Apply the above transformation to all images and perform radiometric correction on each pixel to eliminate inconsistencies in camera response and atmospheric effects.

[0091] Geometric correction includes:

[0092] Control point selection: Select several control points on the ground that are easy to identify in the image. These points should be evenly distributed throughout the study area and have precise geographic coordinates (latitude and longitude).

[0093] Control point marking: The precise location of these control points is marked in the image, usually by clicking on the image at the exact location of the control points.

[0094] Geometric model selection: Select an appropriate geometric correction model, such as a polynomial model, an affine transformation model, or a projection transformation model, to describe the relationship between image pixels and the actual position on the ground.

[0095] Correction parameter calculation: Use the ground coordinates and image coordinates of the control points to calculate the parameters of the geometric correction model.

[0096] Coordinate transformation: Apply the calculated correction parameters to resample the image and map each pixel to its correct ground position. This process usually involves the following steps:

[0097] Reprojection: Convert the pixel coordinates of the original image to coordinates in a new coordinate system.

[0098] Resampling: In the new coordinate system, the pixel values ​​are calculated by interpolation method to generate a corrected image.

[0099] In step S4, the random forest model is used for training. The specific steps are as follows:

[0100] Model training: Data preparation: Input features include multispectral remote sensing data from drones, sweet potato vegetation index, terrain parameters, etc. Data partitioning: The dataset is divided into training set and validation set with a ratio of 8:2 or 7:3.

[0101] Feature selection: Based on the feature importance assessment, the features with greater impact on the sweet potato plot area division and background removal were selected.

[0102] Model construction: Build multiple decision trees. Each tree randomly selects some features and samples for training during training. Each tree is fully grown without pruning.

[0103] Parameter optimization: Optimize the parameters of random forest, including the number of trees, tree depth, and number of feature selections, through grid search or random search methods;

[0104] In step S4, the feature importance evaluation formula is:

[0105] Where:

[0106] Importance(f) indicates the importance of feature f;

[0107] N represents the number of decision trees;

[0108] ΔIi(f) represents the reduction in impurity of feature f in the i-th tree; the reduction in impurity is calculated using the Gini impurity and information gain indicators;

[0109] Grid search finds the optimal parameter settings by traversing all possible parameter combinations. The calculation formula is:

[0110] BestParams=argmaxθ∈ΘPerformance(Model(θ,X train ,y train ),X val ,y val )

[0111] In the formula

[0112] BestParamsBestParams represents the optimal parameter combination;

[0113] Θ represents the set of all possible parameter combinations;

[0114] Model(θ,X train ,y train ) means using parameter θ in the training set (X train ,y train )The model trained on

[0115] Performance represents the model performance evaluation function, including accuracy, recall, and F1 value;

[0116] (X val ,y val ) represents the validation set.

[0117] In step S5, threshold segmentation is performed, and a suitable threshold needs to be determined first. This threshold can be determined by visual interpretation or statistical methods. A set of typical sweet potato vegetation pixels and a non-vegetation pixel can be selected, their NDVI or EVI values ​​can be calculated, and then the difference between them can be used as the threshold. Once the threshold is determined, the NDVI or EVI values ​​of all pixels can be compared. Those pixels greater than the threshold are considered to be vegetation pixels, while those less than the threshold are considered to be background pixels. In this way, the sweet potato vegetation can be separated from other background elements. .

[0118] In step S6, the specific process includes:

[0119] Data extraction: Extract multispectral reflectance data for each cell from the processed drone image. Assume there are n cells, and each cell has spectral data of m bands.

[0120] Data preprocessing: To ensure data quality, perform the following preprocessing steps:

[0121] Outlier removal: Identify and remove abnormal data points due to noise or other factors.

[0122] Data normalization: Spectral data were normalized to the same range using the min-max normalization method.

[0123] Organize data structure: Organize the extracted spectral data into a matrix, where rows represent plots and columns represent spectral reflectance values ​​of different bands. Make sure that the data of each plot corresponds to the corresponding fertilization treatment label.

[0124] Perform one-way ANOVA:

[0125] Hypothesis Testing:

[0126] Null hypothesis (H0): The mean spectral reflectance of plots in all fertilization treatments is equal.

[0127] Alternative hypothesis (H1): The mean spectral reflectance of plots in at least one fertilization treatment is different from that of other treatments.

[0128] Calculate statistics:

[0129] Calculate the within-group mean, between-group mean, and overall mean.

[0130] Calculate within-group and between-group variances.

[0131] Calculate the F value, which is the ratio of the between-group variance to the within-group variance, F = MSB / MSE, where MSB is the between-group mean square and MSE is the within-group mean square.

[0132] Determine the significance level: Choose a significance level (α), usually 0.05.

[0133] Find the F distribution table: Find the critical value in the F distribution table based on the degrees of freedom (between groups and within groups) and the significance level.

[0134] Make a decision: If the calculated F value is greater than the critical value, reject the null hypothesis and think that different fertilization treatments have a significant effect on spectral reflectance. If the F value is not greater than the critical value, fail to reject the null hypothesis and think that different fertilization treatments have no significant effect on spectral reflectance.

[0135] Interpretation of results: If the null hypothesis is rejected, further multiple comparison analysis is performed to determine which specific treatments have significant differences. Record and report the ANOVA's F value, P value, group mean, standard deviation and other statistical results.

[0136] Where n represents the number of cells; m represents the number of spectral bands; MSB represents the mean square between groups; MSE represents the mean square within groups; F value represents the ratio of the variance between groups to the variance within groups; P value represents the probability value, which indicates the probability of observing the current result or a more extreme result.

[0137] In step S7, regular drone flight: drone flight is performed regularly over sweet potato fields according to a predetermined plan to collect the latest multispectral image data. Data processing: preprocessing of the collected data, including radiation correction, geometric correction, vegetation index calculation, etc. Model update: based on new data, the prediction model is continuously updated and adjusted to improve the accuracy and reliability of the model.

[0138] In step S8, the data management system includes the following aspects: User authority management: define different user roles and corresponding authority levels to ensure that only authorized personnel can access sensitive data.

[0139] Data backup and recovery: Regularly back up important data in the database to prevent data loss in the event of an accident.

[0140] Logging: Record all data operation logs to facilitate tracking and analysis of potential security issues.

[0141] Privacy protection: Comply with relevant privacy laws and policies and protect the confidentiality of personal information.

[0142] From the above, it can be seen that in the present invention, the use of random forest model for training brings significant advantages. First, the comprehensiveness of data preparation and the accuracy of feature selection jointly improve the prediction accuracy of the model. By integrating multiple features such as multispectral remote sensing data, vegetation index and terrain parameters, the model can comprehensively consider the various factors affecting sweet potato growth, thereby more accurately predicting the results of cell division and background removal. Feature importance assessment further screens out the features that have the greatest impact on the model, reduces the interference caused by irrelevant factors, makes the model more focused on key information, and improves the accuracy and reliability of prediction.

[0143] In the present invention, the construction of the random forest model and the parameter optimization process enhance the generalization ability and processing efficiency of the model. The integrated learning mechanism of multiple decision trees, combined with the features and samples selected randomly, effectively avoids the overfitting problem and ensures the stable performance of the model on unknown data. By optimizing the model parameters through grid search or random search method, not only the optimal parameter combination is found, the performance indicators of the model, such as accuracy, recall rate and F1 value, are improved, but also the overall efficiency of data processing is improved. These steps work together to provide strong technical support for the refined management of sweet potato planting, making agricultural production more scientific and intelligent.

[0144] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for sweet potato plot area division and background removal based on unmanned aerial vehicle multispectral remote sensing, characterized in that: The method comprises the following steps: S1: According to the needs of the sweet potato fertilizer test, design the field experiment plan and determine the number, area, and arrangement parameters of the plots; arrange necessary markers in the field, such as white or black PVC boards, QR codes, and special reflective coatings, to facilitate drone image recognition; S2: Use a drone equipped with a multispectral camera to fly in the field to collect multispectral images of sweet potato fields; set reasonable flight parameters, including flight altitude and heading overlap rate, to ensure image quality and coverage; when collecting multispectral images, synchronously record flight track information; S3: Use Pix4D or other image processing software to process the multispectral images collected by the drone. First, perform radiation correction to eliminate the influence of lens distortion and flight attitude on the image. Then perform image stitching to generate orthophotos. In this process, geometric correction is also required to ensure that the plane position accuracy of the image is at the centimeter level to meet the needs of subsequent analysis. S4: Use the random forest model machine learning algorithm to automatically segment the drone imagery; train the model to identify the boundaries of plots in the field and automatically extract the image data of each plot; S5: Using NDVI, EVI or other vegetation indices, perform threshold segmentation to remove soil and weed background pixels; extract the average value of the remaining vegetation pixels in each plot to obtain pure sweet potato canopy spectral feature data; S6: Based on the results of cell division and background removal, extract the multispectral reflectance data of each cell; use statistical analysis methods to analyze the differences in spectral characteristics of different cells S7: Combine spectral reflectance characteristics with machine learning algorithms to achieve parameter prediction of sweet potato growth conditions. Design a dynamic monitoring solution for the plot based on the time series change analysis of plot data. S8: Store the processed multispectral images, cell division results, and background removal data in a database; establish a data management system to facilitate users to query, analyze, and download data.

2. A method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S1, first, the number, area and arrangement of the experimental plots need to be determined according to the purpose and requirements of the sweet potato fertilizer test; generally, the number of plots should be sufficient for statistical analysis, with an area of ​​10-20 square meters; The arrangement method usually adopts a random block design or a Latin square design; in the field, white or black PVC boards are arranged at certain intervals as reflective controls, and QR codes or special reflective paints are used to mark the boundaries of the plots so that the drone can accurately distinguish each plot during image recognition.

3. A method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S2, before the UAV flies, it is necessary to set appropriate flight parameters according to the geographical location, terrain and weather conditions of the sweet potato field; the flight altitude is usually set to 50-100 meters to ensure sufficient image resolution; the heading overlap rate is set to 60%-80%, and the lateral overlap rate is set to 70%-80% to ensure the integrity and quality of image stitching; during the flight, the multispectral camera carried by the UAV will synchronously record image data and flight track information to ensure that the actual situation of the field can be accurately restored during subsequent data processing.

4. A method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S3, the radiation correction includes: Black and white board calibration: Place a full white board and a full black board with standard reflectivity at the same time at the shooting site; shoot the black and white boards before and after the drone flies to obtain their images; Image acquisition: During the flight of the drone, multispectral image data is collected, and the camera exposure time and ISO parameters are recorded; Radiometric calibration: Use black and white image to establish the conversion relationship between the image DN value and the actual reflectivity or radiant brightness; this is usually done using the following formula: L = a\timesDN+bL = a×DN+b, where L is the radiant brightness or reflectivity, DN is the digital number of the image, and a and b are calibration coefficients calculated using the black and white image; Apply correction: Apply the above transformation relationship to all images and perform radiometric correction on each pixel to eliminate inconsistencies in camera response and atmospheric effects; Geometric correction includes: Control point selection: Select several control points on the ground that are easy to identify in the image. These points should be evenly distributed throughout the study area and have precise geographic coordinates. Control point marking: accurately marking the location of these control points in the image, usually by clicking on the exact location of the control points on the image; Geometric model selection: Select an appropriate geometric correction model, such as a polynomial model, an affine transformation model, or a projective transformation model, to describe the relationship between image pixels and the actual position on the ground; Correction parameter calculation: Calculate the parameters of the geometric correction model using the ground coordinates and image coordinates of the control points; Coordinate transformation: Apply the calculated correction parameters to resample the image to map each pixel to its correct ground position; this process usually involves the following steps: Reprojection: Convert the pixel coordinates of the original image to coordinates in the new coordinate system; Resampling: In the new coordinate system, the pixel values ​​are calculated by interpolation method to generate a corrected image.

5. The method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S4, a random forest model is used for training, and the specific steps are as follows: Model training: Data preparation: Input features include multispectral remote sensing data from drones, sweet potato vegetation index, and terrain parameters; Data partitioning: Divide the dataset into training and validation sets with a ratio of 8:2 or 7:3; Feature selection: Based on the feature importance assessment, the features with greater impact on sweet potato plot area division and background removal were selected; Model construction: Build multiple decision trees. Each tree randomly selects some features and samples for training. Each tree is fully grown without pruning. Parameter optimization: Optimize the parameters of random forest, including the number of trees, tree depth, and number of feature selections, through grid search or random search methods.

6. A method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S4, the feature importance evaluation formula is: Where: Importance(f) indicates the importance of feature f; N represents the number of decision trees; ΔIi(f) represents the reduction in impurity of feature f in the i-th tree; the reduction in impurity is calculated using the Gini impurity and information gain indicators; Grid search finds the optimal parameter settings by traversing all possible parameter combinations. The calculation formula is: BestParams=argmaxθ∈ΘPerformance(Model(θ,X train ,y train ), X val ,and val ) In the formula BestParamsBestParams represents the optimal parameter combination; Θ represents the set of all possible parameter combinations; Model(θ,X train ,y train ) means using parameter θ in the training set (X train ,y train )The model trained on Performance represents the model performance evaluation function, including accuracy, recall, and F1 value; (X val ,y val ) represents the validation set.

7. The method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S5, threshold segmentation is performed, and a suitable threshold needs to be determined first; this threshold is determined by visual interpretation or statistical methods; a group of typical sweet potato vegetation pixels and a non-vegetation pixel are selected, their NDVI or EVI values ​​are calculated, and then the difference between them is used as the threshold; once the threshold is determined, the NDVI or EVI values ​​of all pixels are compared; those pixels greater than the threshold are considered to be vegetation pixels, and those less than the threshold are considered to be background pixels; in this way, the sweet potato vegetation is separated from other background elements.

8. The method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In step S6, the specific process includes: Data extraction: Extract multispectral reflectance data for each cell from the processed drone image; assuming there are n cells, each cell has spectral data of m bands; Data preprocessing: To ensure data quality, perform the following preprocessing steps: Outlier removal: Identify and remove abnormal data points due to noise or other factors; Data normalization: Spectral data were normalized to the same range using the min-max normalization method. Organize data structure: Organize the extracted spectral data into a matrix, where rows represent plots and columns represent spectral reflectance values ​​of different bands; ensure that the data of each plot corresponds to the corresponding fertilization treatment label; Perform one-way ANOVA: Hypothesis Testing: Null hypothesis: The mean phase of the spectral reflectance of plots in all fertilization treatments; Alternative hypothesis: The mean spectral reflectance of plots in at least one fertilization treatment is different from that of the other treatments; Calculate statistics: Calculate the within-group mean, between-group mean, and total mean; Calculate within-group variance and between-group variance; Calculate the F value, which is the ratio of the between-group variance to the within-group variance, F = MSB / MSE, where MSB is the between-group mean square and MSE is the within-group mean square; Determine the significance level: Select a significance level, usually 0.05; Find the F distribution table: Find the critical value in the F distribution table based on the degrees of freedom and significance level; Make a decision: If the calculated F value is greater than the critical value, reject the null hypothesis and think that different fertilization treatments have a significant effect on spectral reflectance; if the F value is not greater than the critical value, fail to reject the null hypothesis and think that different fertilization treatments have no significant effect on spectral reflectance; Interpretation of results: If the null hypothesis is rejected, further multiple comparison analysis is performed to determine which specific treatments have significant differences; record and report the ANOVA F value, P value, mean and standard deviation statistical results of each group; Where n represents the number of cells; m represents the number of spectral bands; MSB represents the mean square between groups; MSE represents the mean square within groups; F value represents the ratio of the variance between groups to the variance within groups; P value represents the probability value, which indicates the probability of observing the current result or a more extreme result.

9. The method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In the step S7, regular drone flight: according to a predetermined plan, drone flights are performed regularly on sweet potato fields to collect the latest multispectral image data; data processing: pre-processing the collected data, including radiation correction, geometric correction, and vegetation index calculation; model updating: based on new data, the prediction model is continuously updated and adjusted to improve the accuracy and reliability of the model.

10. The method for sweet potato cell area division and background removal based on unmanned aerial vehicle multispectral remote sensing as claimed in claim 1, characterized in that: In the step S8, the data management system; It includes the following aspects: User authority management: define different user roles and corresponding authority levels to ensure that only authorized personnel can access sensitive data; at the same time, regularly back up important data in the database to prevent data loss in the event of an accident; record all data operation logs to facilitate tracking and analysis of potential security issues.

Images