Satellite farm data processing optimization method based on machine learning

Abstract

The invention relates to the technical field of agricultural remote sensing data processing, in particular to a satellite farm data processing optimization method based on machine learning, which comprises the following steps: S1, acquiring image data covering a farm by using multispectral and hyperspectral imaging technologies of satellites; s2, carrying out denoising and contrast enhancement processing by using an image conversion network; s3, a visual Transform model is adopted to process the image data output in the S2 so as to extract agricultural indexes; s4, performing multi-source data fusion by using an ensemble learning method, and outputting a comprehensive analysis result for use in S5; and S5, on the basis of the comprehensive analysis result in the S4, utilizing a machine learning model to predict future farm management requirements, and generating a specific farm management suggestion according to a prediction result. According to the method, the satellite image and the ground data are fused, and the visual Transform and machine learning technologies are applied, so that the farm data processing efficiency and the decision accuracy are remarkably improved, the resource management is optimized, and the crop yield and quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural remote sensing data processing, and in particular to a satellite farm data processing optimization method based on machine learning. Background Art

[0002] With the acceleration of global agricultural development, modern farm management faces more and more challenges, especially in the precise management of water resources, prevention of crop diseases, and improvement of crop yield and quality. Traditional agricultural management methods rely on empirical judgment and regular field inspections. These methods are often inefficient when dealing with large-scale farmland and cannot respond to rapidly changing environmental conditions in real time. In addition, due to the lack of real-time data analysis support, farm management decisions are often not optimized, resulting in waste of resources and yield losses.

[0003] Although satellite remote sensing technology has been used in existing technologies to monitor crop growth and environmental changes, these technologies still have some limitations. First, traditional data processing methods cannot fully utilize the large amount of data collected by satellites and ground stations, resulting in low data processing efficiency and limited result accuracy. Second, existing systems lack effective mechanisms in data fusion, feature extraction and decision support, and fail to provide a comprehensive solution to meet the complex needs of agricultural management.

[0004] Therefore, it is necessary to develop a new technical approach that can integrate advanced machine learning technology and multi-source data processing to achieve efficient analysis of farm data and real-time and accurate farm management decision support, thereby addressing the shortcomings of existing technologies and improving the scientific and technological level of agricultural production. Summary of the invention

[0005] Based on the above objectives, the present invention provides a satellite farm data processing optimization method based on machine learning.

[0006] The satellite farm data processing optimization method based on machine learning includes the following steps:

[0007] S1: Use satellite multispectral and hyperspectral imaging technology to obtain image data covering the farm, and perform preliminary screening to remove low-quality image data affected by cloud cover or transmission errors, and output high-quality data for use by the image conversion network;

[0008] S2: Receives high-quality image data from S1, performs denoising and contrast enhancement using an image transformation network, which corrects the predetermined area in the image through a spatial transformer to ensure data quality, and outputs the preprocessed image data for use by the visual Transformer model;

[0009] S3: Uses a visual Transformer model to process the image data output by S2 to extract agricultural indicators, including crop growth status, moisture content, and vegetation health index, and outputs feature data for use by the ensemble learning method;

[0010] S4: Integrate the feature data output by the visual Transformer model with the real-time climate and soil data from the ground meteorological station, use the ensemble learning method to fuse multi-source data, and output the comprehensive analysis results for use by S5;

[0011] S5: Based on the comprehensive analysis results in S4, the machine learning model is used to predict future farm management needs and generate specific farm management recommendations based on the prediction results.

[0012] Furthermore, the S1 specifically includes:

[0013] S11: Capture image data covering the farm from orbiting satellites through multispectral and hyperspectral imaging equipment. The image data includes vegetation index, moisture index, growth condition image, and soil moisture image. The image data is obtained through satellite imaging technology to provide a comprehensive view of farm land and crop conditions.

[0014] S12: The captured image data is processed initially, and low-quality image data affected by cloud cover or data transmission errors are identified and removed by using a preset automated image analysis algorithm. The screening process is to assess the quality of the image, including checking the image clarity and visual defects such as color distortion, to ensure that only image data that meets the predetermined quality standards is identified as high-quality data;

[0015] S123: Image data that is confirmed to be of high quality after preliminary screening will be formatted into a standard data format suitable for subsequent processing and transmitted to the image conversion network via a high-speed data interface.

[0016] Further, the S12 includes:

[0017] S121: First, a preset automatic image analysis algorithm is used to perform a preliminary evaluation on the received image, specifically a cloud detection algorithm, which identifies the cloud layer in the coverage area by analyzing the spectral characteristics and texture information of the image. The specific formula is expressed as: in, Represents pixel The probability of cloud cover is and They represent the spectral characteristics and texture characteristics of the point respectively, and f is the trained classification function used to determine whether the point is covered by clouds;

[0018] S122: For image areas not covered by clouds, data integrity and quality checks are performed, including evaluation of image clarity and color distortion. Clarity is evaluated by calculating the edge gradient of the image using the Sobel edge detection algorithm. The specific formula is: Among them G x and G y is the gradient of the image in the horizontal and vertical directions. Color distortion detection is performed by calculating the deviation of the image color histogram. The specific formula is: Among them, H i Represents the frequency of the i-th color in the color histogram of the image, is the average frequency of the histogram, n is the total number of color levels;

[0019] S123: Based on the evaluation results of cloud detection, clarity and color distortion, a preset quality threshold standard is used to finally determine whether the image is of high quality. Only image data with a low probability of cloud detection and clarity and color histogram deviation that meet the standards are classified as high-quality data.

[0020] Furthermore, the S2 specifically includes:

[0021] S21: receiving the recognized high-quality image data;

[0022] S22: De-noising and contrast enhancement are performed through an image conversion network. De-noising is performed by applying a convolutional neural network in deep learning. The convolutional neural network structure is used to effectively identify and suppress noise patterns in images. Contrast enhancement is performed by adjusting the brightness and contrast parameters of the image and using adaptive histogram equalization technology to improve the clarity and visual effect of the image.

[0023] S23: The spatial transformer module is embedded in the image transformation network to correct the predetermined area in the image, especially the distorted area caused by the satellite angle or illumination change. The spatial transformer automatically adjusts the target area in the image by learning the geometric transformation relationship between the input image and a set of training images to ensure the accuracy and consistency of subsequent processing;

[0024] S24: The image data processed by the image conversion network will be formatted and encapsulated and converted into a standard format used for feature extraction to ensure that all processed image data meets the input requirements of feature extraction, including data type, resolution, and color depth.

[0025] Furthermore, the space converter submodule includes:

[0026] Image collection and preprocessing: The spatial transformer submodule first collects a set of training images, which include farm images captured at different satellite angles and lighting conditions. The training images are preprocessed, including standardizing the image size, adjusting brightness and contrast, to ensure the consistency of the input;

[0027] Feature point detection and matching: Use the feature detection algorithm to identify feature points between the training image and the new input image. The specific formula of the feature detection algorithm is: in, is an image in scale space, is the Gaussian kernel, σ is the scale parameter, * represents the convolution operation, It is an input image, and then feature matching technology is used to match feature points to determine similar areas between images;

[0028] Geometric transformation model learning: Based on the detected matching feature points, the spatial transformer submodule uses a transformation model, including affine transformation or perspective transformation, to learn the geometric transformation relationship between the input image and the training image. The transformation model is represented by the following matrix:

[0029] Where x′, y′ are the coordinate points after transformation, x, y are the original coordinate points, and a, b, c, d, e, f, g, h are the transformation parameters determined by least squares optimization;

[0030] Image correction and output: Apply the learned geometric transformation model to automatically adjust the target area in the image to correct the distortion caused by satellite angle or illumination changes, and format the corrected image data.

[0031] Further, the S3 includes:

[0032] S31: receiving corrected and enhanced high-quality image data;

[0033] S32: The visual Transformer model is used to process the input image data. Specifically, the image data is divided into multiple small segments, and then the small segments are input as sequences into the Transformer network based on the self-attention mechanism. The visual Transformer model learns the complex relationship between different segments through the following self-attention calculation formula:

[0034] Where Q, K, and V represent query, key, and value matrices, respectively. k is the dimension of the key vector, V is the set of value vectors corresponding to the key vector, is the formula for calculating the attention score, where QK T represents the dot product of the query vector and the key vector, which is used to measure the similarity between them.k is the dimension of the key vector, and softmax is the function used to convert the attention score (i.e. QK T The result of the self-attention mechanism is used to assist the model to focus on the predetermined area in the input image.

[0035] S33: Based on the output of the visual Transformer model, specific agricultural indicators are extracted from the output of the deep learning model, including crop growth status, moisture content, and vegetation health index;

[0036] S34: Process the extracted agricultural indicators into standardized data formats, including data normalization, encoding and packaging.

[0037] Further, the S33 includes:

[0038] S331: first, receiving output data from the visual Transformer model, including a deep feature description of the image segment, and highlighting agricultural information in the image;

[0039] S332: Parse the feature vector provided by the visual Transformer model, which is high-dimensional and contains rich information captured from each image segment, including color, texture, edge, and shape features;

[0040] S333: Based on the parsed feature vectors, apply machine learning classification and regression algorithms to quantify and classify agricultural indicators;

[0041] S334: Finally, the quantified and classified agricultural indicators are processed into a standardized format.

[0042] Further, the S4 includes:

[0043] S41: First, it receives two types of data inputs: one is the feature data output by the visual Transformer model, which is used to provide detailed visual information on crop growth status, moisture content, and vegetation health index; the other is real-time climate and soil data from ground-based meteorological stations, including agricultural parameters such as temperature, humidity, rainfall, and soil composition;

[0044] S42: Use ensemble learning methods to fuse image feature data from the visual Transformer model with ground meteorological and soil data. Specifically, use the ensemble method of random forest or gradient boosting machine to process multi-source data to identify and integrate information from various data sources and improve the prediction performance of decision support.

[0045] S43: Format the comprehensive analysis results after data fusion processing, including converting the analysis results into a format that is easy to interpret and apply, to ensure accurate transmission of information and efficient application.

[0046] Further, the S5 includes:

[0047] S51: Receive the comprehensive analysis results from S4, which integrate the image feature data provided by the visual Transformer model with the real-time climate and soil data from the ground meteorological station;

[0048] S52: Uses time series forecasting models to predict irrigation needs or crop disease risks in the next few days. This model can automatically learn and identify potential management needs based on historical and current data trends;

[0049] S53: Generate specific farm management recommendations based on the output of the time series forecasting model, including irrigation plans, fertilization schedules, and disease control measures to support farm managers in making scientific decisions.

[0050] Furthermore, the S52 specifically includes:

[0051] S521: First, integrate the received comprehensive analysis results, including the image feature data provided by the visual Transformer model and the real-time climate and soil data from the ground meteorological station;

[0052] S522: Select the autoregressive integrated moving average model to analyze and predict time series data that show obvious trends or seasonality. The specific formula is:

[0053] Among them, Y t is the target variable at time point t, c is a constant term, φ i is the coefficient of the autoregressive term, θ i is the coefficient of the moving average term, t is the error term, p and q are the orders of autoregression and moving average of the model respectively;

[0054] S523: Use historical datasets to train selected time series models, including tuning model parameters to minimize forecast error, and perform model validation to ensure the model can accurately predict future irrigation needs and crop disease risks;

[0055] S524: Based on the trained time series model, perform forecasting operations to estimate the irrigation demand or crop disease risk in the next few days. The forecast is based on the latest climate conditions, soil conditions, and crop growth status variables to ensure the real-time and accuracy of the forecast results;

[0056] S525: The prediction results are formatted and outputted so that they can be directly used for generating management suggestions in S53 to form specific farm management suggestions.

[0057] Beneficial effects of the present invention:

[0058] The present invention, by integrating satellite image data and ground meteorological data, can provide a comprehensive data view, enabling farm managers to obtain more detailed and comprehensive farm status information. This multi-source data fusion method not only increases the amount of information for decision-making, but also improves the accuracy and real-time nature of decision-making, especially in water resource management and disease prevention and control.

[0059] The visual Transformer model and machine learning technology used in the present invention significantly improve the efficiency and effect of data processing. Through the self-attention mechanism, the visual Transformer can effectively identify and emphasize key areas in the image, thereby more accurately extracting features related to crop growth. The application of this deep learning method makes it faster and more accurate to extract useful information from large amounts of satellite data, greatly improving the ability of agricultural data analysis.

[0060] The present invention, through the time series prediction model, allows farm managers to predict future irrigation needs, crop growth trends and potential disease risks, so as to formulate corresponding management measures in advance. This not only optimizes the use of resources and reduces waste, but also helps to improve crop yield and quality, bringing direct economic benefits to the farm. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 A schematic diagram of a satellite farm data processing optimization method according to an embodiment of the present invention;

[0063] Figure 2 Schematic diagram of the process of using a visual Transformer model in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] 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.

[0065] 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 "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.

[0066] like Figure 1-Figure 2 As shown, the satellite farm data processing optimization method based on machine learning includes the following steps:

[0067] S1: Use satellite multispectral and hyperspectral imaging technology to obtain image data covering the farm, and perform preliminary screening to remove low-quality image data affected by cloud cover or transmission errors, and output high-quality data for use by the image conversion network;

[0068] S2: Receives high-quality image data from S1, performs denoising and contrast enhancement using an image transformation network, which corrects the predetermined area in the image through a spatial transformer to ensure data quality, and outputs the preprocessed image data for use by the visual Transformer model;

[0069] S3: Uses a visual Transformer model to process the image data output by S2 to extract agricultural indicators, including crop growth status, moisture content, and vegetation health index, and outputs feature data for use by the ensemble learning method. The visual Transformer model analyzes the global relationship between fragments and identifies agricultural indicators by serializing image fragments and applying a self-attention mechanism;

[0070] S4: Integrate the feature data output by the visual Transformer model with the real-time climate and soil data from the ground meteorological station, use the ensemble learning method to fuse multi-source data, and output the comprehensive analysis results for use by S5;

[0071] S5: Based on the comprehensive analysis results in S4, the machine learning model is used to predict future farm management needs and generate specific farm management recommendations based on the prediction results.

[0072] S1 specifically includes:

[0073] S11: Capture image data covering the farm from orbiting satellites through multispectral and hyperspectral imaging equipment. The image data includes vegetation index (NDVI), moisture index, growth condition imagery, and soil moisture imagery. The imagery is obtained through satellite imaging technology to provide a comprehensive view of farm land and crop conditions.

[0074] S12: The captured image data is processed initially, and low-quality image data affected by cloud cover or data transmission errors are identified and removed by using a preset automated image analysis algorithm. The screening process is to assess the quality of the image, including checking the image clarity and visual defects such as color distortion, to ensure that only image data that meets the predetermined quality standards is identified as high-quality data;

[0075] S123: Image data that is confirmed to be of high quality after preliminary screening will be formatted into a standard data format suitable for subsequent processing and transmitted to the image conversion network through a high-speed data interface, ensuring that the data received by the image conversion network is consistent and analyzable, providing a solid data foundation for subsequent deep learning and feature extraction operations; through the detailed description of the above satellite data acquisition, it is ensured that each step from data capture to preliminary screening is highly accurate and efficient, providing accurate and real-time remote sensing data support for farm management, thereby optimizing the management and operation of the entire farm.

[0076] S12 includes:

[0077] S121: First, a preset automatic image analysis algorithm is used to perform a preliminary evaluation on the received image, specifically a cloud detection algorithm, which identifies the cloud layer in the coverage area by analyzing the spectral characteristics and texture information of the image. The specific formula is expressed as: in, Represents pixel The probability of cloud cover is and They represent the spectral characteristics and texture characteristics of the point respectively, and f is the trained classification function used to determine whether the point is covered by clouds;

[0078] S122: For image areas not covered by clouds, data integrity and quality checks are performed, including evaluation of image clarity and color distortion. Clarity is evaluated by calculating the edge gradient of the image using the Sobel edge detection algorithm. The specific formula is: Among them G x and G y is the gradient of the image in the horizontal and vertical directions. Color distortion detection is performed by calculating the deviation of the image color histogram. The specific formula is: Among them, H iRepresents the frequency of the i-th color in the color histogram of the image, is the average frequency of the histogram, n is the total number of color levels;

[0079] S123: Based on the evaluation results of cloud detection, clarity and color distortion, the preset quality threshold standard is used to finally determine whether the image is of high quality. Only image data with low cloud detection probability and clarity and color histogram deviation that meet the standards are classified as high-quality data. Through the application of the above detailed steps and calculation formulas, low-quality image data caused by cloud occlusion or data transmission errors can be effectively identified and eliminated, ensuring that the data output to subsequent steps is highly reliable and consistent, providing solid data support for the precise management of farms.

[0080] S2 specifically includes:

[0081] S21: Receive the identified high-quality image data. After preliminary screening, this data eliminates most of the problems of cloud cover and transmission errors, ensuring the basic quality of the data;

[0082] S22: De-noising and contrast enhancement are performed through an image conversion network. De-noising is performed by applying a convolutional neural network (CNN) in deep learning. The convolutional neural network structure is used to effectively identify and suppress noise patterns in images. Contrast enhancement is performed by adjusting the brightness and contrast parameters of the image and using adaptive histogram equalization technology to improve the clarity and visual effect of the image.

[0083] S23: The spatial transformer module is embedded in the image transformation network to correct the predetermined area in the image, especially the distorted area caused by the satellite angle or illumination change. The spatial transformer automatically adjusts the target area in the image by learning the geometric transformation relationship between the input image and a set of training images to ensure the accuracy and consistency of subsequent processing;

[0084] S24: The image data processed by the image conversion network will be formatted and packaged and converted into a standard format used for feature extraction to ensure that all processed image data meet the input requirements of feature extraction, including data type, resolution, and color depth. Through the above detailed description, intelligent data preprocessing can effectively improve data quality and provide uniform and high-quality input data for subsequent deep learning and feature analysis. This process ensures the efficiency and accuracy of the entire data processing process, which is crucial for the satellite farm data processing optimization method based on machine learning.

[0085] The space transformer submodule includes:

[0086] Image collection and preprocessing: The spatial transformer submodule first collects a set of training images, which include farm images captured at different satellite angles and lighting conditions. The training images are preprocessed, including standardizing the image size, adjusting brightness and contrast, to ensure the consistency of the input;

[0087] Feature point detection and matching: Use a feature detection algorithm, such as Scale Invariant Feature Transform (SIFT) or Speeded Up Robust Features (SURF), to identify feature points between the training image and the new input image. The specific formula of the feature detection algorithm is: in, is an image in scale space, is the Gaussian kernel, σ is the scale parameter, * represents the convolution operation, It is an input image, and then feature matching techniques such as nearest neighbor search are used to match feature points to determine similar areas between images;

[0088] Geometric transformation model learning: Based on the detected matching feature points, the spatial transformer submodule uses a transformation model, including affine transformation or perspective transformation, to learn the geometric transformation relationship between the input image and the training image. The transformation model is represented by the following matrix:

[0089] Where x′, y′ are the coordinate points after transformation, x, y are the original coordinate points, and a, b, c, d, e, f, g, h are the transformation parameters determined by least squares optimization;

[0090] Image correction and output: Apply the learned geometric transformation model to automatically adjust the target area in the image to correct the distortion caused by satellite angle or illumination changes, and format the corrected image data; through the above steps, the spatial transformer submodule can effectively correct the distorted area in the input image to ensure that the image data quality meets the requirements of subsequent processing. This process not only improves the data availability, but also enhances the performance and reliability of the entire system.

[0091] S3 includes:

[0092] S31: receiving corrected and enhanced high-quality image data, which has been subjected to noise, contrast adjustment and spatial correction to ensure that the image data input to feature extraction has a high degree of clarity and accuracy;

[0093] S32: The visual Transformer model is used to process the input image data. Specifically, the image data is divided into multiple small fragments (patches), and then the small fragments are input as sequences into the Transformer network based on the self-attention mechanism. The visual Transformer model learns the complex relationship between different fragments through the following self-attention calculation formula:

[0094] Where Q, K, and V represent query, key, and value matrices respectively. k is the dimension of the key vector, V is the set of value vectors corresponding to the key vector, is the formula for calculating the attention score, where QK T represents the dot product of the query vector and the key vector, which is used to measure the similarity between them. k is the dimension of the key vector. The square root and division operations are used to normalize these scores to prevent numerical calculation problems when training deep learning models. Softmax is the function used to convert the attention scores (i.e., QK T The results of () are converted into probability forms, which indicate the importance of each key (i.e., image fragment) given the query. The self-attention mechanism is used to assist the model to focus on the predetermined area in the input image;

[0095] S33: Based on the output of the visual Transformer model, specific agricultural indicators are extracted from the output of the deep learning model, including crop growth status, moisture content, and vegetation health index. This process involves parsing the features of the Transformer model output and applying specific algorithms to quantify and classify these indicators.

[0096] S34: Process the extracted agricultural indicators into a standardized data format, including data normalization, encoding and encapsulation, to ensure the compatibility and operability of the data; Through the above description, key agricultural indicators can be effectively extracted from the preprocessed images. These indicators are crucial for understanding and managing the status of the farm. The application of the visual Transformer model provides an efficient way to process complex image data and extract valuable information from it, thereby supporting precise agricultural decision-making.

[0097] S33 includes:

[0098] S331: First, receive the output data from the visual Transformer model, including the deep feature description of the image fragment, which has been enhanced by the self-attention mechanism and highlights the agricultural information in the image;

[0099] S332: Parse the feature vector provided by the visual Transformer model, which is high-dimensional and contains rich information captured from each image segment, including color, texture, edge and shape features, which are directly related to the growth status of crops, moisture content and vegetation health index;

[0100] S333: Based on the parsed feature vectors, apply machine learning classification and regression algorithms to quantify and classify agricultural indicators. For example, support vector machines (SVM) can be used to classify crop health status, or linear regression can be used to estimate moisture content; Support vector machines (SVM):

[0101] Among them, y i is the label of the training sample, x i is the feature of the training sample, x is the new input feature, K is the kernel function, α i and b are parameters determined by training, linear regression: y = β0 + β1x1 + β2x2 + … + β n x n , where y is the target to be predicted (such as moisture content), x1, x2, …, x n are input features, β0,β1,…,β n is the regression coefficient;

[0102] S334: Finally, the quantified and classified agricultural indicators are processed into a standardized format, which includes converting the indicator values ​​into a format suitable for use by decision support systems to ensure the consistency and interpretability of the data; through the above detailed steps, agricultural indicator extraction can effectively extract and quantify key agricultural indicators from the output of the deep learning model, which is crucial for farm management and decision making, ensuring the accurate conversion from image data to actual usable agricultural information, and enhancing the application value and operational efficiency.

[0103] S4 includes:

[0104] S41: First, it receives two types of data inputs: one is the feature data output by the visual Transformer model, which is used to provide detailed visual information on crop growth status, moisture content, and vegetation health index; the other is real-time climate and soil data from ground-based meteorological stations, including agricultural parameters such as temperature, humidity, rainfall, and soil composition;

[0105] S42: An integrated learning method is used to fuse the image feature data from the visual Transformer model with the ground meteorological and soil data. The integrated learning method improves the accuracy of the overall prediction by combining the prediction results of multiple learning algorithms. The specific technology uses the integration method of random forest or gradient boosting machine to process multi-source data to identify and integrate information from various data sources and improve the prediction performance of decision support. For example, the integration method of random forest is expressed as:

[0106] Among them, X represents the input feature data set, N is the number of decision trees, DecisionTree i is the prediction result of the i-th decision tree, and Y is the final output prediction;

[0107] S43: Format the comprehensive analysis results after data fusion processing, including converting the analysis results into a format that is easy to interpret and apply, to ensure accurate transmission and application efficiency of information; through the above steps, fusion analysis can efficiently integrate data from different sources and provide a comprehensive data view, thereby enhancing the information support basis for farm management decisions.

[0108] S5 includes:

[0109] S51: Receive the comprehensive analysis results from S4, which integrate the image feature data provided by the visual Transformer model with the real-time climate and soil data from the ground meteorological station. The comprehensive analysis results provide a comprehensive data foundation for the machine learning model to support accurate prediction analysis;

[0110] S52: Uses time series forecasting models to predict irrigation needs or crop disease risks in the next few days. This model can automatically learn and identify potential management needs based on historical and current data trends;

[0111] S53: Generate specific farm management recommendations based on the output of the time series forecasting model, including irrigation plans, fertilization schedules, and disease control measures, and convert the forecast results into practical operational steps to support farm managers in making scientific decisions.

[0112] S52 specifically includes:

[0113] S521: First, integrate the received comprehensive analysis results, including the image feature data provided by the visual Transformer model and the real-time climate and soil data from the ground meteorological station. This data is used to build a complete data set required for time series analysis, ensuring that all relevant factors affecting irrigation demand and crop disease risks are included;

[0114] S522: Select the Autoregressive Integrated Moving Average (ARIMA) model to analyze and predict time series data that show obvious trends or seasonality. The specific formula is:

[0115] Among them, Y t is the target variable at time point t (such as irrigation demand), c is a constant term, φ i is the coefficient of the autoregressive term, θ i is the coefficient of the moving average term, t is the error term, p and q are the orders of autoregression and moving average of the model respectively;

[0116] S523: Use historical datasets to train selected time series models, including tuning model parameters to minimize forecast error, and perform model validation to ensure the model can accurately predict future irrigation needs and crop disease risks;

[0117] S524: Based on the trained time series model, perform forecasting operations to estimate the irrigation demand or crop disease risk in the next few days. The forecast is based on the latest climate conditions, soil conditions, and crop growth status variables to ensure the real-time and accuracy of the forecast results;

[0118] S525: Format the prediction results and output them so that they can be directly used to generate management suggestions in S53 to form specific farm management suggestions. Through the above steps, machine learning prediction can effectively use time series prediction models to predict future farm management needs, providing a scientific basis to support farm managers in making effective resource allocation and disease management decisions.

[0119] The present invention is intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A satellite farm data processing optimization method based on machine learning, characterized in that: The following steps are involved: S1: Use satellite multispectral and hyperspectral imaging technology to obtain image data covering the farm, and perform preliminary screening to remove low-quality image data affected by cloud cover or transmission errors, and output high-quality data for use by the image conversion network; S2: Receives high-quality image data from S1, performs denoising and contrast enhancement using an image transformation network, which corrects the predetermined area in the image through a spatial transformer to ensure data quality, and outputs the preprocessed image data for use by the visual Transformer model; S3: Uses a visual Transformer model to process the image data output by S2 to extract agricultural indicators, including crop growth status, moisture content, and vegetation health index, and outputs feature data for use by the ensemble learning method; S4: Integrate the feature data output by the visual Transformer model with the real-time climate and soil data from the ground meteorological station, use the ensemble learning method to fuse multi-source data, and output the comprehensive analysis results for use by S5; S5: Based on the comprehensive analysis results in S4, the machine learning model is used to predict future farm management needs and generate specific farm management recommendations based on the prediction results.

2. The satellite farm data processing optimization method based on machine learning according to claim 1 is characterized in that: The S1 specifically includes: S11: Capture image data covering the farm from orbiting satellites through multispectral and hyperspectral imaging equipment. The image data includes vegetation index, moisture index, growth condition image, and soil moisture image. The image data is obtained through satellite imaging technology to provide a comprehensive view of farm land and crop conditions. S12: The captured image data is processed initially, and low-quality image data affected by cloud cover or data transmission errors are identified and removed by using a preset automated image analysis algorithm. The screening process is to assess the quality of the image, including checking the image clarity and visual defects such as color distortion, to ensure that only image data that meets the predetermined quality standards is identified as high-quality data; S123: Image data that is confirmed to be of high quality after preliminary screening will be formatted into a standard data format suitable for subsequent processing and transmitted to the image conversion network via a high-speed data interface.

3. The satellite farm data processing optimization method based on machine learning according to claim 2 is characterized in that: The S12 includes: S121: First, a preliminarily evaluation of the received image is performed using a preset automated image analysis algorithm, specifically a cloud detection algorithm, which identifies the cloud layer in the coverage area by analyzing the spectral characteristics and texture information of the image. The specific formula is: C(x,y)=f(S(x,y) , T(x,y)), where C(x,y) represents the cloud coverage probability of the pixel point (x,y), S(x,y) and T(x,y) represent the spectral characteristics and texture characteristics of the point respectively, and f is the trained classification function used to determine whether the point is covered by clouds; S122: For image areas not covered by clouds, data integrity and quality checks are performed, including evaluation of image clarity and color distortion. Clarity is evaluated by calculating the edge gradient of the image using the Sobel edge detection algorithm. The specific formula is: Among them G x and G y is the gradient of the image in the horizontal and vertical directions. Color distortion detection is performed by calculating the deviation of the image color histogram. The specific formula is: Among them, H i Represents the frequency of the i-th color in the color histogram of the image, is the average frequency of the histogram, n is the total number of color levels; S123: Based on the evaluation results of cloud detection, clarity and color distortion, a preset quality threshold standard is used to finally determine whether the image is of high quality. Only image data with a low probability of cloud detection and clarity and color histogram deviation that meet the standards are classified as high-quality data.

4. The satellite farm data processing optimization method based on machine learning according to claim 1 is characterized in that: The S2 specifically includes: S21: receiving the recognized high-quality image data; S22: De-noising and contrast enhancement are performed through an image conversion network. De-noising is performed by applying a convolutional neural network in deep learning. The convolutional neural network structure is used to effectively identify and suppress noise patterns in images. Contrast enhancement is performed by adjusting the brightness and contrast parameters of the image and using adaptive histogram equalization technology to improve the clarity and visual effect of the image. S23: The spatial transformer module is embedded in the image transformation network to correct the predetermined area in the image, especially the distorted area caused by the satellite angle or illumination change. The spatial transformer automatically adjusts the target area in the image by learning the geometric transformation relationship between the input image and a set of training images to ensure the accuracy and consistency of subsequent processing; S24: The image data processed by the image conversion network will be formatted and encapsulated and converted into a standard format used for feature extraction to ensure that all processed image data meets the input requirements of feature extraction, including data type, resolution, and color depth.

5. The satellite farm data processing optimization method based on machine learning according to claim 1 is characterized in that: The space converter submodule comprises: Image collection and preprocessing: The spatial transformer submodule first collects a set of training images, which include farm images captured at different satellite angles and lighting conditions. The training images are preprocessed, including standardizing the image size, adjusting brightness and contrast, to ensure the consistency of the input; Feature point detection and matching: Use the feature detection algorithm to identify feature points between the training image and the new input image. The specific formula of the feature detection algorithm is: in, is an image in scale space, is the Gaussian kernel, σ is the scale parameter, * represents the convolution operation, It is an input image, and then feature matching technology is used to match feature points to determine similar areas between images; Geometric transformation model learning: Based on the detected matching feature points, the spatial transformer submodule uses a transformation model, including affine transformation or perspective transformation, to learn the geometric transformation relationship between the input image and the training image. The transformation model is represented by the following matrix: Where x′, y′ are the coordinate points after transformation, x, y are the original coordinate points, and a, b, c, d, e, f, g, h are the transformation parameters determined by least squares optimization; Image correction and output: Apply the learned geometric transformation model to automatically adjust the target area in the image to correct the distortion caused by satellite angle or illumination changes, and format the corrected image data.

6. The satellite farm data processing optimization method based on machine learning according to claim 1 is characterized in that: The S3 includes: S31: receiving corrected and enhanced high-quality image data; S32: The visual Transformer model is used to process the input image data. Specifically, the image data is divided into multiple small segments, and then the small segments are input as sequences into the Transformer network based on the self-attention mechanism. The visual Transformer model learns the complex relationship between different segments through the following self-attention calculation formula: Where Q, K, and V represent query, key, and value matrices, respectively. k is the dimension of the key vector, V is the set of value vectors corresponding to the key vector, is the formula for calculating the attention score, where QK T represents the dot product of the query vector and the key vector, which is used to measure the similarity between them. k is the dimension of the key vector, and softmax is the function used to convert the attention score (i.e. QK T The result of the self-attention mechanism is used to assist the model to focus on the predetermined area in the input image. S33: Based on the output of the visual Transformer model, specific agricultural indicators are extracted from the output of the deep learning model, including crop growth status, moisture content, and vegetation health index; S34: Process the extracted agricultural indicators into standardized data formats, including data normalization, encoding and packaging.

7. The satellite farm data processing optimization method based on machine learning according to claim 6 is characterized in that: The S33 includes: S331: first, receiving output data from the visual Transformer model, including a deep feature description of the image segment, and highlighting agricultural information in the image; S332: Parse the feature vector provided by the visual Transformer model, which is high-dimensional and contains rich information captured from each image segment, including color, texture, edge, and shape features; S333: Based on the parsed feature vectors, apply machine learning classification and regression algorithms to quantify and classify agricultural indicators; S334: Finally, the quantified and classified agricultural indicators are processed into a standardized format.

8. The satellite farm data processing optimization method based on machine learning according to claim 1 is characterized in that: The S4 includes: S41: First, it receives two types of data inputs: one is the feature data output by the visual Transformer model, which is used to provide detailed visual information on crop growth status, moisture content, and vegetation health index; the other is real-time climate and soil data from ground-based meteorological stations, including agricultural parameters such as temperature, humidity, rainfall, and soil composition; S42: Use ensemble learning methods to fuse image feature data from the visual Transformer model with ground meteorological and soil data. Specifically, use the ensemble method of random forest or gradient boosting machine to process multi-source data to identify and integrate information from various data sources and improve the prediction performance of decision support. S43: Format the comprehensive analysis results after data fusion processing, including converting the analysis results into a format that is easy to interpret and apply, to ensure accurate transmission of information and efficient application.

9. The satellite farm data processing optimization method based on machine learning according to claim 1 is characterized in that: The S5 includes: S51: Receive the comprehensive analysis results from S4, which integrate the image feature data provided by the visual Transformer model with the real-time climate and soil data from the ground meteorological station; S52: Uses time series forecasting models to predict irrigation needs or crop disease risks in the next few days. This model can automatically learn and identify potential management needs based on historical and current data trends; S53: Generate specific farm management recommendations based on the output of the time series forecasting model, including irrigation plans, fertilization schedules, and disease control measures to support farm managers in making scientific decisions.

10. The satellite farm data processing optimization method based on machine learning according to claim 9 is characterized in that: The S52 specifically includes: S521: First, integrate the received comprehensive analysis results, including the image feature data provided by the visual Transformer model and the real-time climate and soil data from the ground weather station; S522: Select the autoregressive integrated moving average model to analyze and predict time series data that show obvious trends or seasonality. The specific formula is: Among them, Y t is the target variable at time point t, c is a constant term, φ i is the coefficient of the autoregressive term, θ i is the coefficient of the moving average term, t is the error term, p and q are the orders of autoregression and moving average of the model respectively; S523: Use historical datasets to train selected time series models, including tuning model parameters to minimize forecast error, and perform model validation to ensure the model can accurately predict future irrigation needs and crop disease risks; S524: Based on the trained time series model, perform forecasting operations to estimate the irrigation demand or crop disease risk in the next few days. The forecast is based on the latest climate conditions, soil conditions, and crop growth status variables to ensure the real-time and accuracy of the forecast results; S525: The prediction results are formatted and outputted so that they can be directly used for generating management suggestions in S53 to form specific farm management suggestions.

Images