End-cloud collaborative agricultural condition image acquisition system based on 5G communication and machine learning

Abstract

The invention provides an end-cloud collaborative agricultural condition image acquisition system based on 5G communication and machine learning, and relates to the technical field of end-cloud collaboration of 5G communication and machine learning. A data acquisition end is used for acquiring multispectral image data of a farmland to be observed, and compression preprocessing is carried out, so that the data transmission burden is reduced; the compressed image is sent to a data processing module, and the data processing module carries out denoising and feature extraction on the image, constructs a feature data vector and transmits the feature data vector to a cloud server; the cloud server analyzes feature data vectors by using a random forest algorithm model, judges crop types and classifies the crop types, and finally returns an analysis result to the data display module, and the data display module generates an instrument panel by using a Flask framework and visually displays agricultural condition data including key indexes such as crop types and coverage conditions. Data transmission is carried out through 5G communication to realize a terminal-cloud collaborative architecture, and efficient and real-time data transmission and processing are realized.

Description

Technical Field

[0001] The present invention relates to the field of end-cloud collaboration technology of 5G communication and machine learning, and specifically to an end-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning. Background Art

[0002] In traditional agricultural monitoring systems, the collection and processing of agricultural images mainly rely on centralized cloud computing architecture. Farmers and agricultural managers usually obtain the growth status of crops in farmland manually or with simple sensors, and then upload these data to the cloud for processing and analysis. However, with the expansion of farmland scale and the application of multispectral image acquisition technology, traditional systems face the problems of high bandwidth demand and transmission delay, which makes it impossible to effectively meet the real-time monitoring needs through the existing system architecture, especially in situations where rapid response is required for crop disease monitoring and climate change. In addition, the centralized transmission of large amounts of data also brings data privacy and security risks. Especially in large-scale agriculture, how to ensure the security of data during transmission and ensure processing efficiency has become a major bottleneck in agricultural monitoring technology.

[0003] Modern agriculture has increasingly higher requirements for the real-time and accuracy of data, but traditional centralized cloud computing architectures can hardly meet these demands. The end-cloud collaborative architecture based on 5G communications and machine learning has become a new solution for agricultural image collection and analysis. This solution performs preliminary data processing through a data processing module and delegates complex computing tasks to the cloud for analysis, thereby improving processing efficiency and feedback speed. The high bandwidth, low latency and high reliability of 5G communications ensure real-time transmission of data, enabling it to adapt to the rapid response needs of large-scale production. At the same time, machine learning technology in the cloud can improve the scientificity and accuracy of predictions and decisions, helping farmers optimize production management.

[0004] Although existing technologies have made some progress in the field of agricultural monitoring, there are still obvious shortcomings. Traditional image acquisition and analysis methods cannot meet the needs of rapidly changing agricultural environments and real-time decision-making. Data latency and processing speed are difficult to improve simultaneously. The lack of an effective end-cloud collaboration mechanism has caused a loose connection between data acquisition and analysis, resulting in inefficient data integration. In addition, existing machine learning algorithms are often trained for specific environments or data sets, lack sufficient generalization capabilities, and cannot adapt to changes in different agricultural scenarios and crop types.

[0005] The present invention aims to effectively solve the shortcomings of the above-mentioned existing technologies through an end-cloud collaborative architecture based on 5G and machine learning. The solution can reduce the delay and bandwidth requirement of data transmission and only upload key data to enhance data security and privacy protection, thereby reducing dependence on the cloud and optimizing resource utilization. At the same time, this architecture improves the flexibility and scalability of the system. The present invention provides more efficient, intelligent and secure solutions in scenarios such as vegetation index, crop type, coverage, biomass estimation and growth of agricultural conditions, providing strong support for the intelligent and precise development of modern agriculture.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0007] The purpose of the present invention is to provide an end-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning, comprising:

[0010] The data collection end is used to collect multispectral image data covering the farmland to be observed in real time and to collect multispectral image data of historical farmland, compress and pre-process the multispectral image, and send the compressed multispectral image to the data processing module;

[0011] The data processing module is used to perform denoising and feature extraction on the pre-processed multispectral image data, obtain different bands and peak intensities, combine the two to construct a feature data vector, and transmit the feature data vector to the cloud server;

[0012] The model building module is used to use the random forest algorithm, take the multispectral image data of historical farmland as the training set, and the corresponding different agricultural data as labels, and input them into the random forest algorithm for training;

[0013] The cloud server is used to input the received feature data vector into the trained random forest algorithm model, obtain various agricultural conditions, and return various agricultural conditions to the data display module;

[0014] The data display module is used to receive agricultural conditions from the cloud server and use Flask to generate a dashboard to display agricultural data.

[0015] Further, a multispectral camera is used to collect multispectral image data of farmland covering bands ranging from 100-200 nanometers, 200-400 nanometers, 400-700 nanometers, and 700-1100 nanometers, and image collection is performed in each time period;

[0016] Obtain historical multispectral images from existing databases, satellite image archives or ground sensors, and perform appropriate data cleaning and format conversion to ensure that historical data has the same format and characteristics as real-time data;

[0017] The collected multispectral images are losslessly compressed using the JPEG2000 format, the compression ratio is set to 1:10, the compressed image data is packaged into TIFF format, and the data is transmitted to the data processing module via 5G communication.

[0018] Furthermore, after receiving the data, the data processing module performs wavelet transform on each image and decomposes it into low-frequency and high-frequency parts. Let f ij (t) is the signal to be processed, which can be expressed as follows using wavelet transform:

[0019]

[0020] In the formula, a represents the scale parameter, which controls the width of the wavelet, b represents the translation parameter, which controls the position of the wavelet, ψ represents the mother wavelet function, which is used to generate the wavelet, and W ij (a, b) represents the wavelet transform coefficients of the jth band of the signal to be processed at different scales a and positions b, f ij (t) is the signal to be processed in the jth range of the i-th acquisition, and the value range of j is (1 to 4), indicating multispectral image data covering the band ranges of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm;

[0021] The high frequency part is processed using the threshold method, and the denoising threshold is set by the following formula:

[0022]

[0023] In the formula, σ represents the standard deviation of the wavelet coefficients, reflecting the noise level, r is the number of wavelet coefficients, and λ represents the denoising threshold, which is used to determine the wavelet coefficients retained in the denoising process;

[0024] Soft threshold processing is performed on the high-frequency part to reduce the wavelet coefficients through the threshold, retaining the signal characteristics while removing noise:

[0025] C′=sgn(C)·max(|C|-λ,0)

[0026] Where C is the original wavelet coefficient, C′ is the denoised wavelet coefficient, sgn(C) represents the sign of C, which is 1 if C>0, -1 if C<0, and 0 if C=0. It is used to maintain the direction of the original wavelet coefficient.

[0027] Use the inverse wavelet transform to reconstruct the denoised wavelet coefficients into a clear signal:

[0028]

[0029] In the formula, f re (t) is the reconstructed signal, which is a clear image or data after denoising. C′(a,b) is the wavelet coefficient after denoising. a,b (t) is the mother wavelet function, which is used to reconstruct the signal.

[0030] Furthermore, the intensity value of each band is obtained by analyzing the denoised signal. For each band j (covering the band range of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm), I j Represents the intensity value of a specific band, which refers to the average reflection intensity of pixels in a certain band:

[0031]

[0032] Where N is the selected area A j The total number of pixels, A j is the pixel area corresponding to the jth band;

[0033] Define the local maximum, a pixel f re (x,y) is considered a local maximum when its intensity value is greater than the intensity values ​​of all its surrounding neighboring pixels, as expressed by the following inequality:

[0034] f re (x,y)>f re (x+q,y+p)and all(q,p)∈N a

[0035] Where N a is a 3x3 neighborhood, which means checking the surrounding 8 adjacent pixels;

[0036] Find the local maximum and calculate the first-order derivative (gradient) of the signal intensity, which represents the rate of change of the intensity. Let S(x,y)=f re (x, y) represents the intensity value of the pixel, then the first-order derivative can be expressed as, horizontally:

[0037]

[0038] Vertical direction:

[0039]

[0040] One pixel f re The condition for (x,y) to be a wave peak is that its first-order derivative changes sign at that point, that is:

[0041]

[0042] and Changes from positive to negative near (x,y) or Changes from positive to negative near (x,y);

[0043] After determining the local maximum, for all identified local peaks (x k ,y k ), the corresponding intensity value peak can be calculated j :

[0044] peak j =max{f re (x k ,y k )}

[0045] Where k represents the index of all local peaks found.

[0046] Furthermore, the feature data vector is composed of the band intensity I j and peak intensity peak j By combining with NDVI (vegetation index), a more comprehensive feature data vector F is obtained:

[0047] F=[I j peak j ,NDVI]

[0048] In the formula, I j ,peak j are the band and peak intensity values ​​of the j-th band value, respectively, where j represents the covered band ranges of 100-200 nm, 200-400 nm, 400-700 nm, and 700-1100 nm;

[0049] Among them, NDVI (Normalized Difference Vegetation Index) is:

[0050]

[0051] In the formula, R 700-1100 Indicates the reflectivity in the near-infrared band (700 nm to 1100 nm), R 400-700 Indicates the reflectivity in the red light band (400 nm to 700 nm);

[0052] Set the band intensity I j Normalized to obtain the reflectivity R j :

[0053]

[0054] In the formula, I max is the maximum reflection intensity in all bands, ensuring that R j The value is between 0 and 1.

[0055] Furthermore, the collected historical multispectral image data of farmland, including reflectance data of different bands, are used to calculate the historical feature data vector F h , and the corresponding crop label Y h Input them into random forest respectively to construct training set:

[0056] T={(F h，u , Y h,u )|u=1,2,...,M}

[0057] In the formula, T is the training set, M is the total number of samples, and F h,u Represents the characteristic data vector of the u-th sample in the historical data, Y h,u Represents the crop label corresponding to the u-th sample in the historical data;

[0058] The random forest consists of M decision trees, and the construction process is as follows:

[0059] Randomly extract M from the training set T o samples, forming the training set T of the mth tree m :

[0060]

[0061] Where M o is the number of samples per tree (less than the total number of samples M), o is the sampling index, Represents the feature vector of the vth sample in history, Represents the label of the vth sample in history;

[0062] At each node split, from the total feature F h Randomly select mtry features for optimal splitting, use the selected features to build a decision tree until the preset stopping condition (maximum depth or minimum number of sample splits) is reached, and for each tree T m , input feature vector F h Row Classification:

[0063]

[0064] In the formula, Represents the prediction result of the mth tree, representing the classification output through the tree, T m Represents the model of the mth tree, using the feature vector F for classification;

[0065] For each sample F h , the final prediction result Y of random forest h Generate by voting on all trees:

[0066]

[0067] Where Y h is the final prediction result of the random forest, generated by voting of all trees. For the prediction results of each tree, the model is the majority function, and the crop category with the most occurrences is selected.

[0068] Furthermore, the received feature data vector is input into the trained random forest algorithm model to predict the agricultural situation:

[0069] Y=model(F)

[0070] Where Y is the prediction result of the feature data F by the random forest model; Y refers to different agricultural conditions. If the goal of the model is classification, then Y represents the predicted crop type; if the goal of the model is regression (estimation of biomass or coverage), then Y represents the corresponding continuous value;

[0071] Y can be crop type, vegetation coverage, biomass, growth status, and for crop vegetation coverage:

[0072] Coverage = f(NDVI,EVI,SAVI)

[0073] Among them, EVI is an improvement on NDVI:

[0074]

[0075] In the formula, R 440-510 Indicates the reflectivity of the blue light band, mainly used to reduce the impact of soil background;

[0076] SAVI aims to provide a more accurate assessment of vegetation cover by adjusting for the effects of soil reflectance:

[0077]

[0078] Where, is the soil adjustment factor, which is usually taken as 0.5;

[0079] Biomass estimation: Biomass can be calculated based on NDVI, temperature, and precipitation:

[0080] Biomass = α·NDVI β +y·T+δ·P

[0081] Where Biomass represents the estimated biomass (kg / hectare), T is temperature (℃), P is precipitation (mm), α, β, γ, δ are parameters fitted based on historical data, reflecting the contribution of NDVI, temperature and precipitation to biomass;

[0082] Growth data can be obtained by monitoring the changes in NDVI:

[0083]

[0084] In the formula, NDVI t Indicates the NDVI value at the current moment, NDVI t-1 It represents the NDVI value at the previous moment, and n is the length of the period for calculating the growth change (1 day or 7 days);

[0085] The analysis results are collated and returned to the data display module to create a result data structure containing the following information: crop type Y, coverage: vegetation coverage derived and evaluated by NDVI values, biomass: estimated biomass, growth data: time series changes of NDVI.

[0086] Furthermore, in the Flask application, the data returned by the cloud server is parsed into a format suitable for front-end display, and a dashboard layout is designed, including a title: displaying the name of the dashboard, "Agricultural Monitoring Dashboard", crop types: displaying classification results, using charts or tables, NDVI curves: displaying the trend of NDVI over time, coverage: using heat maps or pie charts to display the coverage of different areas, biomass estimation: displaying through bar charts or other charts, and growth data: displaying growth trends, represented by line graphs.

[0087] Compared with the prior art, the present invention has the following beneficial effects:

[0088] This invention significantly improves the high latency, bandwidth pressure and data security problems existing in traditional agricultural monitoring systems by introducing an end-cloud collaborative architecture combined with 5G communication technology. The high bandwidth and low latency characteristics of 5G technology ensure the real-time transmission of farmland multispectral image data, enabling the system to provide timely monitoring and feedback in situations that require rapid response, such as crop diseases and climate change. This greatly reduces the slow response problem of traditional systems caused by high bandwidth requirements and transmission delays.

[0089] The present invention significantly improves multiple technical problems in traditional agricultural monitoring systems through the collaborative work of the data acquisition end, data processing module, model building module, cloud server and data display module. The data acquisition end adopts advanced multispectral sensors, which can collect high-quality image data in real time and reduce the amount of data through compression preprocessing, thereby reducing the demand for bandwidth and transmission delay and improving the efficiency of data transmission; the data processing module performs denoising and feature extraction to ensure data quality, providing a reliable basis for subsequent analysis.

[0090] In the cloud, the random forest algorithm efficiently analyzes the feature data vectors, which can quickly and accurately determine the crop type and health status, solving the problem of insufficient adaptability of existing machine learning algorithms. At the same time, the data display module generates a visual dashboard through Flask, which intuitively presents the analysis results and helps agricultural managers make decisions quickly. In summary, the present invention not only improves the real-time and accuracy of agricultural monitoring, but also provides an efficient and safe monitoring solution for modern agriculture, promoting the intelligent development of agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a schematic diagram of the overall system module of the present invention. DETAILED DESCRIPTION

[0092] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0093] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0094] Example:

[0095] Please refer to the attached drawings of the specification, the present invention provides a technical solution:

[0096] A terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning, comprising:

[0097] The data collection end is used to collect multispectral image data covering the farmland to be observed in real time and to collect multispectral image data of historical farmland, compress and pre-process the multispectral image, and send the compressed multispectral image to the data processing module;

[0098] Select the high-resolution MicaSense RedEdge (multispectral camera), which covers the band range of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm and has the multispectral imaging capability required for farmland monitoring. Set the camera's shooting parameters, including shutter speed, aperture and ISO, to adapt to the optimal lighting conditions, and collect multispectral images of the farmland under the optimal lighting conditions; connect the MicaSense RedEdge to a remote control device (such as a remote control or smartphone application), set the time interval of the remote control device to regularly collect images at different growth stages, and choose to collect images every 1 hour, 2 hours or at a fixed time every day. During the set time period, use the remote control device to activate the camera to shoot, which can be a single shot or set to continuous shooting mode to ensure that multiple images are captured in a dynamically changing environment to obtain dynamic information at different growth stages;

[0099] The camera has a band reflection intensity recording function. Enable this function in the shooting settings to ensure that the reflection intensity data of each band is automatically recorded after shooting. Whenever an image is collected, the system automatically generates the reflection intensity value of each band and associates it with the corresponding image file to ensure that the image and data files are saved synchronously.

[0100] Collect historical multispectral image data for model training, validation and trend analysis, obtain historical multispectral images from existing databases, satellite image archives or ground sensors, and perform appropriate data cleaning and format conversion to ensure that historical data has the same format and characteristics as real-time data;

[0101] Import the collected multispectral images into the image processing software, compress them losslessly using the JPEG2000 format, set the compression ratio to 1:10, and start the image compression process. The software will generate a compressed image file to ensure that the original data has not been changed;

[0102] In the image processing software, import the compressed JPEG2000 image file, select the "Export" or "Save As" function, select the TIFF format, ensure that all compressed images are packaged into a TIFF file, and save the file to ensure the integrity and accessibility of all data. Use data transmission tools (such as FTP, HTTP, etc.) to transfer the packaged TIFF format image data to the data processing module via 5G communication to ensure the integrity and security of the data during the process.

[0103] The data processing module is used to perform denoising and feature extraction on the pre-processed multispectral image data, obtain different bands and peak intensities, combine the two to construct a feature data vector, and transmit the feature data vector to the cloud server;

[0104] After receiving the data, the data processing module performs wavelet transform on each image and decomposes it into low-frequency and high-frequency parts. Let f ij (t) is the signal to be processed, which can be expressed as follows using wavelet transform:

[0105]

[0106] In the formula, a represents the scale parameter, which controls the width of the wavelet, b represents the translation parameter, which controls the position of the wavelet, ψ represents the mother wavelet function, which is used to generate the wavelet, and W ij (a, b) represents the wavelet transform coefficients of the jth band of the signal to be processed at different scales a and positions b, f ij (t) is the signal to be processed in the jth range of the i-th acquisition, and the value range of j is (1 to 4), indicating multispectral image data covering the band ranges of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm;

[0107] Among them, the scale parameter a controls the width of the wavelet: the larger the scale, the wider the wavelet, and the smoother the captured signal features; the smaller the scale, the narrower the wavelet, and more details can be captured; the translation parameter b controls the position of the wavelet, and changing b can slide the wavelet on the time axis; W ij (a, b) represents the wavelet transform coefficients of the signal to be processed at different scales a and positions b of the jth band collected at the i-th time. By selecting different a and b, the characteristic information of the signal at various scales and positions can be obtained. Wavelet transform is a signal processing technology used to analyze the characteristics of signals at different scales and time positions. Through this transform, complex signals can be decomposed into low-frequency and high-frequency components, thereby extracting the main features and details of the signal.

[0108] Use wavelet transform to decompose the signal into multiple levels, get the low-frequency part (approximation) and high-frequency part (details), and apply denoising technology in the high-frequency level:

[0109] The high frequency part is processed using the threshold method, and the denoising threshold is set by the following formula:

[0110]

[0111] In the formula, σ represents the standard deviation of the wavelet coefficients, reflecting the noise level, r is the number of wavelet coefficients, and λ represents the denoising threshold, which is used to determine the wavelet coefficients retained in the denoising process;

[0112] This formula is used to set the denoising threshold λ, where σ represents the standard deviation of the wavelet coefficients, reflecting the level of noise; r is the number of wavelet coefficients, and this threshold is used to determine which wavelet coefficients should be considered as noise and removed in subsequent processing; the standard deviation σ of the wavelet coefficients, the higher the noise level, the larger λ, which means that more noise components need to be removed; the number of wavelet coefficients r, the larger the number, the larger λ, which means that when processing on a larger data set, more noise factors need to be considered; the denoising threshold λ represents the denoising standard selected in wavelet denoising, reflecting the size of the noise level;

[0113] Soft threshold processing is performed on the high-frequency part to reduce the wavelet coefficients through the threshold, retaining the signal characteristics while removing noise:

[0114] C'=sgn(C)·max(|C|-λ,0)

[0115] Where C is the original wavelet coefficient, C′ is the denoised wavelet coefficient, sgn(C) represents the sign of C, which is 1 if C>0, -1 if C<0, and 0 if C=0. It is used to maintain the direction of the original wavelet coefficient.

[0116] This formula is used to denoise the high-frequency wavelet coefficient C. It can effectively reduce the noise in the wavelet coefficient while retaining the signal characteristics through the soft threshold method. The larger the absolute value of the original wavelet coefficient C, the more signal characteristics will be retained after threshold processing. The larger the denoising threshold λ, the more high-frequency components will be removed. The denoised wavelet coefficient C′, through processing, reduces the noise and retains important signal characteristics.

[0117] Use the inverse wavelet transform to reconstruct the denoised wavelet coefficients into a clear signal:

[0118]

[0119] In the formula, f re(t) is the reconstructed signal, which is a clear image or data after denoising. C′(a,b) is the wavelet coefficient after denoising. a,b (t) is the mother wavelet function, used to reconstruct the signal;

[0120] The inverse wavelet transform reconstructs the denoised wavelet coefficients C′ into a clear signal f re (t), this process allows the useful features of the original signal to be restored from the denoised information; the denoised wavelet coefficients C′(a,b), its value affects the quality and characteristics of the final reconstructed signal; the reconstructed signal f re (t), its clarity and accuracy depend on the processing effect of wavelet coefficients;

[0121] The intensity value of each band is obtained by analyzing the denoised signal. For each band j (covering the band range of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm), I j Represents the intensity value of a specific band, which refers to the average reflection intensity of pixels in a certain band:

[0122]

[0123] Where N is the selected area A j The total number of pixels in the band, A j is the pixel area corresponding to the jth band, usually the area in the image corresponding to the wavelength range of the band, I j represents the average reflection intensity value of the jth band, which represents the average reflection intensity of all pixels in a specific band, f re (x, y) represents the pixel reflection intensity value at position (x, y), which represents the signal intensity after denoising. j The pixel reflection intensity f re When (x,y) increases, the average reflection intensity I j will increase;

[0124] Define the local maximum, a pixel f re (x,y) is considered a local maximum when its intensity value is greater than the intensity values ​​of all its surrounding neighboring pixels, as expressed by the following inequality:

[0125] f re (x,y)>f re (x+q,y+p)and all(q,p)∈N a

[0126] Where N a is a 3x3 neighborhood, indicating that the 8 neighboring pixels around are checked. Specifically, Na Contains 8 pixels centered at (x, y) (left, right, top, bottom, and four diagonal directions);

[0127] The local maximum refers to a pixel f re The intensity value of (x, y) is greater than the intensity values ​​of its eight adjacent pixels, that is, the intensity of this point is the highest among its surroundings. re If (x, y) is a local maximum, its intensity is higher than that of its neighboring pixels, indicating that the pixel is the most significant feature point in the area;

[0128] Find the local maximum and calculate the first-order derivative (gradient) of the signal intensity, which represents the rate of change of the intensity. Let S(x,y) = f re (x, y) represents the intensity value of the pixel, which is the intensity value of the pixel (x, y), usually referring to the signal value after denoising. The first-order derivative represents the rate of change of intensity, that is, the change of pixel intensity in the horizontal and vertical directions. In the horizontal direction:

[0129]

[0130] Vertical direction:

[0131]

[0132] If the derivative of S(x,y) in a certain direction becomes larger (for example If the derivative is positive, it means that the intensity is increasing; if the derivative becomes smaller (negative), it means that the intensity is decreasing, which means that the change in the sign of the first-order derivative reflects the increase or decrease in intensity;

[0133] One pixel f re The condition for (x,y) to be a wave peak is that its first-order derivative changes sign at that point, that is:

[0134]

[0135] and Changes from positive to negative near (x,y) or Changes from positive to negative near (x,y);

[0136] The condition for the peak is that the intensity value of the pixel shows a certain change in both the horizontal and vertical directions (that is, the first-order derivative changes from positive to negative at this point). and When the conditions are met at the same time, it means that the pixel is a local maximum, so it can be considered as a feature point (peak) in the image;

[0137] After determining the local maximum, for all identified local peaks (x k ,y k), the corresponding intensity value peak can be calculated j :

[0138] peak j =max{f re (x k ,y k )}

[0139] Where k represents the index of all local peaks found, peak j represents the local peak intensity of the jth band, which represents the maximum intensity value among all local peak pixels found; peak j is the maximum value of the intensity values ​​of all local peaks found, so it is consistent with the intensity of the local peak f re (x k ,y k ) is directly related to the size of the peak. When the local peak intensity increases, the peak j It may also increase;

[0140] The characteristic data vector is composed of the band intensity I j and peak intensity peak j By combining with NDVI (vegetation index), a more comprehensive feature data vector F is obtained:

[0141] F=[I j , peak j ,NDVI]

[0142] In the formula, I j ,peak j are the band and peak intensity values ​​of the jth band value, respectively, j represents the coverage band range of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm; the feature vector F is used in the subsequent classification algorithm and can provide a comprehensive analysis of the crop status; the changes in the intensity values ​​of each band and peak reflect the response of different bands to the health status of the crop; the feature vector F, its dimension and content directly affect the input effect of the classification algorithm, reflecting the spectral characteristics of crops in different bands, NDVI is the normalized vegetation index; I j ,peak j ,NDVI is the independent variable, which reflects the spectral characteristics of different bands and the health status of vegetation;

[0143] Among them, NDVI (Normalized Difference Vegetation Index) is:

[0144]

[0145] In the formula, R 700-1100Indicates the reflectance in the near-infrared band (700 nm to 1100 nm), which is often used to assess the health of plants because healthy plants usually have higher reflectance in this band. 400-700 Indicates the reflectance in the red light band (400 nm to 700 nm). Healthy plants usually have a low reflectance in this band because they absorb a lot of red light for photosynthesis.

[0146] Typically, NDVI values ​​range from -1 to +1, with values ​​close to 1 indicating good vegetation cover, values ​​close to 0 indicating no vegetation (such as urban areas or bare soil), and negative values ​​typically indicating water bodies or other non-vegetative cover; furthermore, as R 700-1100 Increase or R 400-700 decreases, NDVI values ​​usually increase, indicating improved vegetation health;

[0147] Set the band intensity I j Normalized to obtain the reflectivity R j :

[0148]

[0149] In the formula, I max is the maximum reflection intensity in all bands, which plays a normalization role and ensures that R j The value of R is between 0 and 1. j Represents the reflectance of the jth band, which is the normalized intensity value, ranging from 0 to 1;

[0150] This formula calculates the reflectivity R of each band j , used to analyze the light reflection characteristics of different bands. Reflectivity is an important indicator for evaluating crop health and soil quality; I j The larger the value, the higher the reflectivity;

[0151] The data processing module is used to perform denoising and feature extraction on the pre-processed multispectral image data, obtain different bands and peak intensities, combine the two to construct a feature data vector, and transmit the feature data vector to the cloud server;

[0152] The collected historical multispectral image data of farmland, including reflectance data of different bands, are used to calculate the historical feature data vector F h , and the corresponding crop label Y h Input them into random forest respectively to construct training set:

[0153] T={(F h，u , Y h,u )|u=1,2,...,M}

[0154] Where T is the training set, which contains multiple samples. Each sample consists of a feature vector and a label. M is the total number of samples, which indicates the number of samples in the training set. h,u It represents the characteristic data vector of the u-th sample in the historical data, reflecting the band reflection intensity, peak intensity, NDVI and other information of the sample. h,u Indicates the crop label corresponding to the u-th sample in the historical data, indicating the actual crop type of the sample;

[0155] The random forest consists of M decision trees, and the construction process is as follows:

[0156] Randomly extract M from the training set T o samples, forming the training set T of the mth tree m :

[0157]

[0158] Where M o is the number of samples per tree (less than the total number of samples M), o is the sampling index, Represents the feature vector of the vth sample in history, represents the label of the vth sample in history, T m Represents the training set of the mth tree, including M o samples; randomly select M from the training set T o The sub-training set T is formed by samples m This random sampling method helps to improve the generalization ability of the model. o It can improve the randomness of the model and prevent overfitting;

[0159] At each node split, from the total feature F h Randomly select m try The best segmentation is performed based on m features. try The number of features selected at each node split is usually less than the total number of features to increase the diversity of the tree; the decision tree is constructed using the selected features to ensure the diversity of each tree and reduce the risk of overfitting until the preset stopping condition (maximum depth or minimum number of sample splits) is reached. m , input feature vector F h Row Classification:

[0160]

[0161] In the formula, Represents the prediction result of the mth tree, indicating that the feature vector F is predicted by this tree h The output of classification, T m Represents the model of the mth tree, using the feature vector F for classification; the feature vector Fh Different values ​​of will directly affect the prediction results of each tree

[0162] For each sample F h , the final prediction result Y of random forest h Generate by voting on all trees:

[0163]

[0164] Where Y h The final prediction result of the random forest is generated by voting of all trees. It reflects the comprehensive judgment of multiple trees and is usually the classification result with the most occurrences. The prediction results of each tree. The combination of different trees can improve the robustness of classification. The model is the majority function, and the crop category with the most occurrences is selected.

[0165] By calculating the feature importance score, we can have a deeper understanding of the decision-making mechanism within the random forest model and identify the features that have the greatest impact on the classification results. Random forests can evaluate which features have the greatest impact on the classification results by calculating the importance of the features. A common method is based on the reduced Gini coefficient or information gain. Let feature F k The importance rating is I k

[0166]

[0167] In the formula, I k For feature F k The importance score indicates the contribution of the feature to the model prediction performance. M is the total number of decision trees. ΔGini m (F k ) is the feature F used in the mth tree k The Gini coefficient reduced by splitting. The Gini coefficient is an indicator to measure the purity of the node. The larger the reduction value, the greater the contribution of the feature to the classification effect.

[0168] This formula calculates the characteristic F k Importance score, which reflects the influence of the feature on the classification results in the random forest model. Feature importance assessment helps to understand the decision-making process of the model and can be used for feature selection to help simplify the model and reduce overfitting; Gini coefficient ΔGini m (F k ), indicating that in the mth tree, feature F is used kThe Gini coefficient reduced by splitting; the Gini coefficient is an indicator to measure the purity of a node. The smaller the Gini coefficient, the higher the purity of the node. The importance of a feature can be evaluated by reducing the Gini coefficient, and the contribution of the feature to the classification ability of the model can be quantified;

[0169] The total number of decision trees, M, indicates the number of trees used in the model. The more trees there are, the more accurate the overall average will be. The feature F is used in each tree. k The Gini coefficient ΔGini reduced by splitting m (F k ), the specific calculation method is the difference between the Gini coefficients before and after the split node. If the feature F k If the feature is used for splitting multiple times in the decision tree and the reduced Gini coefficient is large, it means that the feature has a high importance; k For feature F k The importance score of feature F is higher. k The greater the contribution to the model’s predictive performance, the more effective the feature is in the decision-making process;

[0170] The Gini coefficient is an indicator to measure the purity of a data set. Its value range is [0,1], where 0 indicates perfect purity (all samples belong to the same class) and 1 indicates complete mixing (samples are evenly distributed in all classes). When using features for splitting, the Gini coefficient will decrease, indicating that the purity of each child node after the split is higher. If feature F k The frequent reduction of Gini coefficient in multiple trees indicates that this feature has a significant contribution in the classification task;

[0171] Features with high importance rating I k Indicates that this feature plays a key role in the classification process. The decision tree model relies on this feature to make classification decisions. If the I of some features k If the value is low, it means that these features contribute little to the model and may have little impact on the predictive performance of the model, so they can be considered for removal.

[0172] The cloud server is used to input the received feature data vector into the trained random forest algorithm model, obtain various agricultural conditions, and return various agricultural conditions to the data display module;

[0173] The received feature data vector is input into the trained random forest algorithm model to predict the agricultural situation:

[0174] Y=model(F)

[0175] Where Y is the prediction result of the feature data F by the random forest model; Y refers to different agricultural conditions. If the goal of the model is classification, then Y represents the predicted crop type; if the goal of the model is regression (estimation of biomass or coverage), then Y represents the corresponding continuous value;

[0176] Y can be crop type, vegetation coverage, biomass, growth status, and for crop vegetation coverage:

[0177] Coverage = f(NDVU,EVI,SAVI)

[0178] Among them, EVI is an improvement on NDVI:

[0179]

[0180] In the formula, R 440-510 Represents the reflectance of the blue light band, which is mainly used to reduce the impact of soil background. Generally, as the health of vegetation improves, R 700-1100 Increase, R 400-700 decrease, which eventually leads to an increase in EVI;

[0181] SAVI is designed to provide a more accurate assessment of vegetation cover by adjusting for the effects of soil reflectance:

[0182]

[0183] Where is the soil adjustment factor, usually set to 0.5, which is used to adjust the effect of soil reflectance. Increased vegetation reflectance will increase the SAVI value, reflecting a higher vegetation coverage.

[0184] Biomass estimation: Biomass can be calculated based on NDVI, temperature, and precipitation:

[0185] Biomass = α·NDVI β +y·T+δ·P

[0186] Where Biomass represents the estimated biomass (kg / hectare), T is temperature (℃), P is precipitation (mm), α, β, γ, δ are parameters fitted based on historical data, reflecting the contribution of NDVI, temperature and precipitation to biomass. Generally, an increased MDVI value will lead to an increase in biomass because it reflects a healthier vegetation condition. Similarly, an increase in temperature and precipitation will also have a positive effect on biomass.

[0187] The higher the vegetation index NDVI, the greater the biomass. The estimated biomass Biomass is usually expressed in kilograms per hectare. As the NDVI increases, the biomass also increases.

[0188] Growth data can be obtained by monitoring the changes in NDVI:

[0189]

[0190] In the formula, NDVI t Indicates the NDVI value at the current moment, NDVI t-n Indicates the NDVI value at the previous n moments, NDVI t-1 Indicates the NDVI value of the previous moment, n is the length of the period for calculating the growth change (1 day or 7 days). Generally, if the current NDVI value is higher than the previous moment, the growth increases, otherwise it decreases;

[0191] Growth rate Growth Rte, reflects the growth change of vegetation through calculation. The larger the value, the faster the growth rate. This formula calculates the change rate of vegetation growth, reflecting the change of vegetation status over time. A positive value indicates that the vegetation growth status has improved, while a negative value indicates that the growth status has deteriorated.

[0192] The analysis results are collated and returned to the data display module to provide comprehensive information on crop conditions. This data structure will be used for subsequent decision support and visualization. Create a result data structure containing the following information: crop type Y, NDVI, coverage: vegetation coverage derived and evaluated by NDVI values, biomass: estimated biomass, growth data: time series changes in NDVI.

[0193] The data display module is used to receive agricultural conditions from the cloud server and use Flask to generate a dashboard to display agricultural data;

[0194] In the Flask application, the data returned by the cloud server is parsed into a format suitable for front-end display. Appropriate folders are created in the project directory. / static: is used to store static resources, such as CSS style sheets and JavaScript scripts, and / templates: is used to store HTML template files to facilitate dynamic generation of web page content. Design a dashboard layout, including a title: display the name of the dashboard, "Agricultural Monitoring Dashboard";

[0195] Create an area to display the classification results of crop types: Use a chart (such as a pie chart or bar chart) to visually display the classification results of each type of crop. You can also choose to use a table to list the crop names and corresponding quantities;

[0196] Design an area to show the changing trend of NDVI values: NDVI curve: Show the trend of NDVI over time, or use a line graph with time on the X-axis and NDVI value on the Y-axis. Make sure the graph can clearly show the changing trend of NDVI over time so that users can see the changes in vegetation health;

[0197] Set up an area to show the coverage of different areas: Coverage: Use a heat map (thermal map) to show the vegetation coverage of different areas. The color depth indicates the coverage degree. You can also use a pie chart to show the overall coverage, and each part represents the coverage ratio of different areas;

[0198] Create a region to display the biomass estimation results: Biomass estimation: Use a bar chart to display the biomass estimation values ​​of different crops or regions. The X-axis represents the crop type or region, and the Y-axis represents the biomass value. Make sure the bar chart is clear and easy to read so that users can compare the biomass of different crops;

[0199] Set up an area to display crop growth data: Growth data: Use a line graph to display growth trends. The X-axis represents time and the Y-axis represents growth indicators (such as biomass or NDVI). This graph can help users intuitively understand the growth of crops.

[0200] In the / static / css and / static / js directories, create style sheets and script files respectively: CSS is used to design the appearance of the dashboard, including fonts, colors, chart styles, layouts, etc.; JavaScript is used to handle the dynamic generation and interactive effects of charts. You can use chart libraries (such as Chart.js or D3.js) to render charts; finally, combine all the parts together to form a consistent dashboard interface, and ensure that the layout of each chart and data display part is reasonable so that users can easily browse and understand the data.

[0201] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0202] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0203] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0204] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning, characterized in that: include: The data collection end is used to collect multispectral image data covering the farmland to be observed in real time and to collect multispectral image data of historical farmland, compress and pre-process the multispectral image, and send the compressed multispectral image to the data processing module; The data processing module is used to perform denoising and feature extraction on the pre-processed multispectral image data, obtain different bands and peak intensities, combine the two to construct a feature data vector, and transmit the feature data vector to the cloud server; The model building module is used to use the random forest algorithm, take the multispectral image data of historical farmland as the training set, and the corresponding different agricultural data as labels, and input them into the random forest algorithm for training; The cloud server is used to input the received feature data vector into the trained random forest algorithm model, obtain various agricultural conditions, and return various agricultural conditions to the data display module; The data display module is used to receive agricultural conditions from the cloud server and use Flask to generate a dashboard to display agricultural data.

2. According to claim 1, a terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning is characterized in that: Use a multispectral camera to collect multispectral image data of farmland covering bands ranging from 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm, and collect images for each time period; Obtain historical multispectral images from existing databases, satellite image archives or ground sensors, and perform appropriate data cleaning and format conversion to ensure that historical data has the same format and characteristics as real-time data; The collected multispectral images are losslessly compressed using the JPEG2000 format, the compression ratio is set to 1:10, the compressed image data is packaged into TIFF format, and the data is transmitted to the data processing module via 5G communication.

3. According to claim 2, a terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning is characterized in that: After receiving the data, the data processing module performs wavelet transform on each image and decomposes it into low-frequency and high-frequency parts. Let f ij (t) is the signal to be processed, which can be expressed as follows using wavelet transform: In the formula, a represents the scale parameter, which controls the width of the wavelet, b represents the translation parameter, which controls the position of the wavelet, ψ represents the mother wavelet function, which is used to generate the wavelet, and W ij (a, b) represents the wavelet transform coefficients of the jth band of the signal to be processed at different scales a and positions b, f ij (t) is the signal to be processed in the jth range of the i-th acquisition, and the value range of j is (1 to 4), indicating multispectral image data covering the band ranges of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm; The high frequency part is processed using the threshold method, and the denoising threshold is set by the following formula: In the formula, σ represents the standard deviation of the wavelet coefficients, reflecting the noise level, r is the number of wavelet coefficients, and λ represents the denoising threshold, which is used to determine the wavelet coefficients retained in the denoising process; Soft threshold processing is performed on the high-frequency part to reduce the wavelet coefficients through the threshold, retaining the signal characteristics while removing noise: C′=sgn(C)·max(|C|-λ,0) Where C is the original wavelet coefficient, C′ is the denoised wavelet coefficient, sgn(C) represents the sign of C, which is 1 if C>0, -1 if C<0, and 0 if C=0. It is used to maintain the direction of the original wavelet coefficient. Use the inverse wavelet transform to reconstruct the denoised wavelet coefficients into a clear signal: In the formula, f re (t) is the reconstructed signal, which is a clear image or data after denoising. C′(a,b) is the wavelet coefficient after denoising. a,b (t) is the mother wavelet function, which is used to reconstruct the signal.

4. According to claim 3, a terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning is characterized in that: The intensity value of each band is obtained by analyzing the denoised signal. For each band j (covering the band range of 100-200 nm, 200-400 nm, 400-700 nm and 700-1100 nm), I j Represents the intensity value of a specific band, which refers to the average reflection intensity of pixels in a certain band: Where N is the selected area A j The total number of pixels, A j is the pixel area corresponding to the jth band; Define the local maximum, a pixel f re (x,y) is considered a local maximum when its intensity value is greater than the intensity values ​​of all its surrounding neighboring pixels, as expressed by the following inequality: fx e (x,y)>f re (x+q,y+p)and all(q,p)∈N a Where N a is a 3x3 neighborhood, which means checking the surrounding 8 adjacent pixels; Find the local maximum and calculate the first-order derivative (gradient) of the signal intensity, which represents the rate of change of the intensity. Let S(x,y) = f re (x, y) represents the intensity value of the pixel, then the first-order derivative can be expressed as, horizontally: Vertical direction: One pixel f re The condition for (x,y) to be a wave peak is that its first-order derivative changes sign at that point, that is: and Changes from positive to negative near (x,y) or Changes from positive to negative near (x,y); After determining the local maximum, for all identified local peaks (x k ,y k ), the corresponding intensity value peak can be calculated j : peak j =max{f re (x k ,y k )} Where k represents the index of all local peaks found.

5. According to claim 4, a terminal-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning is characterized in that: Feature data The vector is composed of the band intensity I j and peak intensity peak j By combining with NDVI (vegetation index), a more comprehensive feature data vector F is obtained: F=[I j ,peak j ,NDVI] In the formula, I j ,peak j are the band and peak intensity values ​​of the j-th band value, respectively, where j represents the covered band ranges of 100-200 nm, 200-400 nm, 400-700 nm, and 700-1100 nm; Among them, NDVI (Normalized Difference Vegetation Index) is: In the formula, R 700-1100 Indicates the reflectivity in the near-infrared band (700 nm to 1100 nm), R 400-700 Indicates the reflectivity in the red light band (400 nm to 700 nm); Set the band intensity I j Normalized to obtain the reflectivity R j : In the formula, I max is the maximum reflection intensity in all bands, ensuring that R j The value is between 0 and 1.

6. The end-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning according to claim 5 is characterized by: The collected historical multispectral image data of farmland, including reflectance data of different bands, are used to calculate the historical feature data vector F h , and the corresponding crop label Y h Input them into random forest respectively to construct training set: T={(F h,u ,Y h,u )∣u=1,2,…,M} In the formula, T is the training set, M is the total number of samples, and F h,u Represents the characteristic data vector of the u-th sample in the historical data, Y h,u Represents the crop label corresponding to the u-th sample in the historical data; The random forest consists of M decision trees, and the construction process is as follows: Randomly extract M from the training set T o samples, forming the training set T of the mth tree m : Where M o is the number of samples per tree (less than the total number of samples M), o is the sampling index, Represents the feature vector of the vth sample in history, Represents the label of the vth sample in history; At each node split, from the total feature F h Randomly select m try The features are optimally split and decision trees are constructed using the selected features until the preset stopping condition (maximum depth or minimum number of sample splits) is reached. m , input feature vector F h Row Classification: In the formula, Represents the prediction result of the mth tree, which represents the classification output of the tree, T m Represents the model of the mth tree, using the feature vector F for classification; For each sample F h , the final prediction result Y of random forest h Generate by voting on all trees: Where Y h is the final prediction result of the random forest, generated by voting of all trees. For the prediction results of each tree, the model is the majority function, and the crop category with the most occurrences is selected.

7. The end-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning according to claim 6 is characterized by: The received feature data vector is input into the trained random forest algorithm model to predict the agricultural situation: Y=model(F) Where Y is the prediction result of the feature data F by the random forest model; Y refers to different agricultural conditions. If the goal of the model is classification, then Y represents the predicted crop type; if the goal of the model is regression (estimation of biomass or coverage), then Y represents the corresponding continuous value; Y can be crop type, vegetation coverage, biomass, growth status, and for crop vegetation coverage: Coverage = f(NDVI,EVI,SAVI) Among them, EVI is an improvement on NDVI: In the formula, R 440-510 Indicates the reflectivity of the blue light band, mainly used to reduce the impact of soil background; SAVI is designed to provide a more accurate assessment of vegetation cover by adjusting for the effects of soil reflectance: Where, is the soil adjustment factor, which is usually taken as 0.5; Biomass estimation: Biomass can be calculated based on NDVI, temperature, and precipitation: Biomass=α·NDVI β +γ·T+δ·P Where Biomass represents the estimated biomass (kg / hectare), T is temperature (℃), P is precipitation (mm), α, β, γ, δ are parameters fitted based on historical data, reflecting the contribution of NDVI, temperature and precipitation to biomass; Growth data can be obtained by monitoring the changes in NDVI: In the formula, NDVI t Indicates the NDVI value at the current moment, NDVI t-1 It represents the NDVI value at the previous moment, and n is the length of the period for calculating the growth change (1 day or 7 days); The analysis results are collated and returned to the data display module to create a result data structure containing the following information: crop type Y, coverage: vegetation coverage derived and evaluated by NDVI values, biomass: estimated biomass, growth data: time series changes of NDVI.

8. The end-cloud collaborative agricultural image acquisition system based on 5G communication and machine learning according to claim 7 is characterized by: In the Flask application, the data returned by the cloud server is parsed into a format suitable for front-end display, and a dashboard layout is designed, including the title: display the name of the dashboard, "Agricultural Monitoring Dashboard", crop types: display classification results, using charts or tables, NDVI curve: display the trend of NDVI over time, coverage: use heat maps or pie charts to display the coverage of different areas, biomass estimation: display through bar charts or other charts, growth data: display growth trends, represented by line graphs.

Images