A cloud computing-based agricultural carbon emission management system

Abstract

The application discloses an agricultural carbon emission management system based on cloud computing, comprising: a data acquisition module for collecting carbon emission data generated by agricultural activities; a data processing module connected with the data acquisition module for pre-processing the carbon emission data and then performing feature extraction; a model construction module connected with the data processing module for constructing a multi-modal network model, training the multi-modal network model, and obtaining a carbon emission prediction model; a trend prediction module connected with the model construction module for predicting agricultural carbon emission based on the carbon emission prediction model and obtaining a carbon emission prediction result; and a control management module connected with the trend prediction module for regulating and managing according to the carbon emission prediction result. The application can significantly improve the collection and analysis efficiency of agricultural carbon emission data, help to improve the level of sustainable agricultural development, and has important significance for environmental protection and climate change mitigation.

Description

Technical Field

[0001] This invention belongs to the field of carbon emission monitoring and control technology, and in particular relates to a cloud computing-based agricultural carbon emission management system. Background Technology

[0002] With the increasing severity of global climate change, agriculture, as a significant source of global greenhouse gas emissions, has received widespread attention for carbon emission management. Traditional agricultural carbon emission management methods suffer from several problems. For instance, current carbon emission monitoring and management typically rely on traditional monitoring methods and manual data collection and processing. This approach is not only inefficient but also prone to incomplete, inaccurate, or outdated data collection and processing. Furthermore, existing systems often lack real-time performance and integration, failing to effectively integrate data from different sources for comprehensive analysis. Agricultural carbon emission management involves multiple stages, such as land cultivation, fertilizer use, crop growth, and livestock management, and current technologies struggle to integrate data from all these stages for holistic analysis. Moreover, the lack of a unified platform for data storage, analysis, and management limits the effective utilization and management of carbon emission data. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a cloud computing-based agricultural carbon emission management system, comprising:

[0004] Preferably, the data acquisition module is used to collect carbon emission data generated by agricultural activities;

[0005] A data processing module, connected to the data acquisition module, is used to preprocess the carbon emission data and then extract its features.

[0006] The model building module, connected to the data processing module, is used to build a multimodal network model and train the multimodal network model to obtain a carbon emission prediction model.

[0007] The trend prediction module is connected to the model building module and is used to predict agricultural carbon emissions based on the carbon emission prediction model to obtain carbon emission prediction results.

[0008] The control and management module is connected to the trend prediction module and is used to regulate and manage emissions based on the carbon emission prediction results.

[0009] Preferably, the data acquisition module includes a data acquisition unit and an image acquisition unit;

[0010] The data acquisition unit is used to acquire parameter data in agricultural production activities;

[0011] The parameter data includes, but is not limited to, environmental parameter data, agricultural production activity data, agricultural input data, agricultural machinery usage data, agricultural waste data, livestock and poultry breeding data, farmland management data, energy consumption data, and agricultural product supply chain data.

[0012] The image acquisition unit is used to acquire remote sensing image data in agricultural production activities;

[0013] The remote sensing image data includes, but is not limited to, farmland remote sensing images, crop growth images, agricultural facility images, livestock monitoring images, agricultural waste images, soil condition images, water resource utilization images, and agricultural activity images.

[0014] Preferably, the data processing module includes a data processing unit and an image processing unit;

[0015] The data processing unit is used to denoise the original data, convert it into text data, and then merge texts from different sources and formats.

[0016] The image processing unit is used to perform pixel preprocessing on the initial remote sensing image and then perform image correction to obtain the target remote sensing image.

[0017] Preferably, the data processing unit includes a data cleaning unit, a text processing unit, and a data integration unit;

[0018] The data cleaning unit is used to remove noise and outliers from the original data and fill in missing data.

[0019] The text processing unit is used to perform word segmentation, dictionary creation, text vectorization, and text normalization on the text data in sequence.

[0020] The data integration unit is used to unify and merge data from different sources and formats, establish relationships between data, and obtain a data view.

[0021] Preferably, the image processing unit includes a pixel processing unit and an image correction unit;

[0022] The pixel processing unit is used to sequentially perform pixel normalization, pixel centering, data enhancement, noise reduction, size adjustment and color space conversion on the initial remote sensing image to obtain the target image for subsequent use;

[0023] The image correction unit is used to sequentially perform radiometric correction, geometric correction, image registration, image fusion, light and color equalization, image stitching, and image cropping on the image processed by the pixel processing unit to obtain the target optical remote sensing image.

[0024] Preferably, the model building module includes a model building unit, a model training unit, and a fusion unit;

[0025] The model building unit is used to build convolutional neural network models and recurrent neural network models;

[0026] The model training unit is used to train the convolutional neural network model and the recurrent neural network model respectively using image data and text data;

[0027] The fusion unit is used to fuse the trained convolutional neural network model and recurrent neural network model to obtain a multimodal neural network model; it is also used to integrate image data and text data to obtain an integrated dataset.

[0028] Preferably, the model building unit includes a first network building unit and a second network building unit;

[0029] The first network building unit is used to build a convolutional neural network and generate a group of convolutional neural networks based on the convolutional neural network;

[0030] The second network building unit is used to build a recurrent neural network and introduce an attention mechanism into the recurrent neural network for optimization, thereby generating an improved recurrent neural network.

[0031] Preferably, the model training unit includes a first training unit, a second training unit, and a third training unit;

[0032] The first training unit includes an iterative computation unit, which is used to iterate the convolutional neural network group multiple times using a genetic algorithm until the optimal model is obtained and the iteration ends to obtain the genetic neural network.

[0033] The second training unit includes a pre-training unit, an iterative training unit, and a transfer learning unit;

[0034] The pre-training unit is used to pre-train the improved recurrent neural network using the Transformer model to generate a pre-trained model; the improved recurrent neural network is generated by introducing the constructed attention mechanism module and residual module into the recurrent neural network.

[0035] The iterative training unit is used to iteratively train the pre-trained model using a sample set, and then optimize the trained model using an optimizer to generate an optimized model.

[0036] The transfer learning unit is used to freeze the low-level parameters of the optimization model, optimize the high-level parameters of the optimization model, and generate the target recurrent neural network.

[0037] The third training unit is used to train the target recurrent neural network using the output of the genetic neural network to generate a multimodal neural network model.

[0038] Preferably, the fusion unit includes a model fusion unit and a data fusion unit;

[0039] The data fusion unit includes:

[0040] A weight generation unit is used to assign weights to the image data and the text data based on the importance of the data, thereby obtaining image weight data and text weight data.

[0041] The weight calculation unit is used to multiply the image weight data and text weight data by their corresponding weights to obtain a weighted value;

[0042] The summation and integration unit is used to add the weighted values ​​to obtain the integrated dataset.

[0043] Preferably, the iterative calculation unit includes:

[0044] The network structure encoding unit is used to encode the structural parameters of the convolutional neural network group into chromosomes for a genetic algorithm, thereby obtaining the network chromosome structure.

[0045] The function definition unit is used to construct a fitness function based on detection rate, false detection rate, and positioning accuracy.

[0046] An iterative unit is used to randomly generate a set of initial CNN chromosome structures based on the network chromosome structure;

[0047] A genetic unit is used to iterate the initial CNN chromosome structure based on the fitness function, select the CNN structure corresponding to the individual with the highest fitness in the last generation of the iteration, and decode it to obtain the genetic neural network.

[0048] The pre-training unit includes:

[0049] Module building units are used to build attention mechanism modules and residual modules;

[0050] An attention mechanism introduction unit is used to introduce the attention mechanism module into the first three hidden layers of a recurrent neural network to obtain an attention classification model;

[0051] A residual introduction unit is used to introduce the residual module into the attention classification model to generate the improved recurrent neural network;

[0052] The genetic unit includes:

[0053] The selection subunit is used to select the initial CNN chromosome structure through the fitness function to obtain the selected parent CNN chromosome structure;

[0054] The crossover subunit is used to perform a crossover operation on the selected parent CNN chromosome structure to generate a new individual;

[0055] The mutation subunit is used to introduce randomness and mutate the new individual to obtain a mutated individual;

[0056] The evolutionary iteration subunit is used to iterate and evolve the CNN chromosome structure multiple times through the selection subunit, crossover subunit, and mutation subunit to select the individual with the highest fitness.

[0057] The decoding subunit is used to decode the individual with the highest fitness to obtain the genetic neural network.

[0058] Compared with the prior art, the present invention has the following advantages and technical effects:

[0059] This invention enables comprehensive monitoring of agricultural carbon emissions through Internet of Things (IoT) technology. Real-time data collection and cloud processing ensure the immediacy and accuracy of the data, avoiding the delays and errors that may occur with traditional manual data collection, and improving the comprehensiveness and real-time nature of data collection.

[0060] This invention utilizes the high-efficiency data processing and storage capabilities of cloud computing, combined with Internet of Things (IoT) technology to achieve real-time data acquisition, and uses intelligent algorithms for data analysis and decision support, thereby improving data processing speed and efficiency.

[0061] This invention can more accurately identify carbon emission sources and predict emission trends, and real-time monitoring and intelligent analysis support rapid response management decisions, improving the timeliness and effectiveness of management.

[0062] The cloud computing-based agricultural carbon emission management system of this invention achieves efficient and precise management of agricultural carbon emissions through technological innovation, providing strong support for sustainable agricultural development. Attached Figure Description

[0063] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0064] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0065] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0066] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0067] like Figure 1 As shown, this embodiment provides a cloud computing-based agricultural carbon emission management system, including:

[0068] The data acquisition module is used to collect carbon emission data generated by agricultural activities;

[0069] The data processing module, connected to the data acquisition module, is used to preprocess carbon emission data and then extract features.

[0070] The model building module, connected to the data processing module, is used to build a multimodal network model and train the multimodal network model to obtain a carbon emission prediction model.

[0071] The trend prediction module, connected to the model building module, is used to predict agricultural carbon emissions based on the carbon emission prediction model and obtain carbon emission prediction results.

[0072] The control and management module, connected to the trend prediction module, is used for regulation and management based on carbon emission prediction results.

[0073] Furthermore, the data acquisition module includes a data acquisition unit and an image acquisition unit;

[0074] The data acquisition unit is used to acquire parameter data in agricultural production activities;

[0075] Among them, the parameter data includes, but is not limited to, environmental parameter data, agricultural production activity data, agricultural input data, agricultural machinery usage data, agricultural waste data, livestock and poultry breeding data, farmland management data, energy consumption data, and agricultural product supply chain data;

[0076] Furthermore, environmental parameter data is collected through various environmental monitoring sensors deployed in key areas such as farmland and livestock farms, including those for temperature, humidity, light intensity, soil moisture, and CO2 concentration. The number of sensors deployed is determined by the size and needs of the monitoring area, using a grid-like distribution to ensure data representativeness and comprehensiveness. Data acquisition frequency varies from once per second to once per day, depending on different monitoring needs and sensor response times. Real-time collection of environmental parameters such as soil moisture, temperature, light intensity, and CO2 concentration, as well as soil and climate data including soil type, fertility level, and climatic conditions (such as temperature, humidity, and rainfall), is conducted.

[0077] Agricultural production activity data includes crop planting area, planting type, planting cycle, irrigation water volume, fertilizer application, and pesticide use.

[0078] Agricultural input data includes the usage and efficiency of agricultural production materials such as fertilizers, pesticides, seeds, and agricultural films.

[0079] Agricultural machinery usage data includes records of the type of agricultural machinery, frequency of use, energy consumption, and emissions.

[0080] Agricultural waste data includes information on the generation, treatment methods, and utilization of agricultural waste such as straw and livestock manure.

[0081] Livestock and poultry farming data includes information related to carbon emissions, such as farming scale, feed consumption, manure management, and intestinal fermentation.

[0082] Farmland management data includes farming practices, land use changes, and the use of farmland irrigation facilities.

[0083] Energy consumption data includes records of the consumption of energy sources such as electricity, fuel oil, and natural gas used in agricultural production.

[0084] Agricultural product supply chain data includes carbon emissions throughout the entire supply chain from production to consumption, including data from processing, storage, and transportation.

[0085] The image acquisition unit is used to acquire remote sensing image data in agricultural production activities;

[0086] Among them, remote sensing image data includes, but is not limited to, farmland remote sensing images, crop growth images, agricultural facility images, livestock monitoring images, agricultural waste images, soil condition images, water resource utilization images, and agricultural activity images;

[0087] Furthermore, farmland remote sensing images, which are images of farmland taken by satellites or drones, are used to monitor crop growth, land cover changes, and other factors.

[0088] Crop growth images are high-resolution images of crops inside farmland, used to assess crop health and the occurrence of pests and diseases.

[0089] Images of agricultural facilities, including greenhouses, irrigation systems, and other agricultural facilities, are used to monitor the status and efficiency of these facilities.

[0090] Livestock monitoring images are images of farms used to monitor the health status of livestock and poultry and the conditions of the farming environment.

[0091] Agricultural waste images include images of farmland waste such as straw and livestock manure, used to assess waste management and resource utilization.

[0092] Soil condition images include images of soil texture, moisture, and other conditions, which are used to analyze soil health and carbon storage capacity.

[0093] Water resource utilization images include images of farmland irrigation water bodies, reservoirs, etc., used to assess water resource utilization efficiency and its impact on carbon emissions.

[0094] Agricultural activity images are images that record agricultural production activities such as tilling, fertilizing, and harvesting, and are used to analyze the contribution of agricultural activities to carbon emissions.

[0095] By collecting these parameter data and image data, we can gain a comprehensive understanding of the impact of agricultural activities on carbon emissions, providing data support for the monitoring, analysis and management of carbon emissions.

[0096] Furthermore, the data processing module includes a data processing unit and an image processing unit;

[0097] The data processing unit is used to denoise the raw data, convert it into text data, and then merge texts from different sources and formats.

[0098] The image processing unit is used to perform pixel preprocessing on the initial remote sensing image and then perform image correction to obtain the target remote sensing image.

[0099] Furthermore, the data processing unit includes a data cleaning unit, a text processing unit, and a data integration unit;

[0100] The data cleaning unit is used to remove noise and outliers from the raw data and fill in missing data; it ensures data quality and provides accurate input for subsequent analysis.

[0101] Specifically, data cleaning includes identifying and removing noisy data, identifying and removing outliers, and handling missing data;

[0102] Identifying and removing noisy data involves: first, performing noise analysis to understand the data's distribution characteristics and potential noise patterns; applying filtering techniques such as moving average filtering, median filtering, or Kalman filtering to smooth the data and reduce random fluctuations; using the Shapiro-Wilk test to assess the data's normality and identify noise points deviating from the normal distribution; and visually identifying outliers or abnormal fluctuations in the data by plotting scatter plots, box plots, etc.

[0103] Identifying and removing outliers includes: setting thresholds (such as mean ± 3σ) based on the statistical characteristics of the data to identify potential outliers; using boxplots to determine quartiles and outlier ranges; and using Mahalanobis distance based on the assumption of a multivariate normal distribution to identify outliers in multidimensional data.

[0104] Handling missing data includes:

[0105] Analyze the patterns and causes of missing data, such as random missing data and non-random missing data.

[0106] For randomly missing data, simple imputation can be performed using the mean, median, or mode.

[0107] Predict missing values ​​using regression, decision trees, or other machine learning models.

[0108] A multiple imputation method is used to estimate the uncertainty of missing data by generating multiple imputation datasets.

[0109] In cases of severe missing data, deleting records containing missing values ​​can be considered, but the impact on the dataset must be assessed.

[0110] The text processing unit is used to perform word segmentation, dictionary creation, text vectorization, and text normalization on text data in sequence.

[0111] Specifically, word segmentation is used to break down text into words or vocabulary units. The word segmentation tools in this embodiment include SpaCy and NLTK.

[0112] Create a dictionary: By creating a dictionary, each unique word is mapped to a unique index. This dictionary is used to convert text data into a numerical form that the model can process.

[0113] Text vectorization is used to convert text from a sequence of words into a sequence of indices, that is, to convert text into numerical vectors so that it can be input into a recurrent neural network model.

[0114] Normalize the text data to help the model learn and understand it better.

[0115] In addition to the above steps, the text processing unit of this embodiment may further include removing stop words (such as common words like "de", "shi" and other words that are common but contribute little to analysis), stemming (restoring words to their basic forms), and lemmatization (restoring words to their basic forms) to reduce noise and improve the efficiency and accuracy of the model.

[0116] The data integration unit is configured to unify and merge data from different sources and in different formats, establish associations between the data, obtain a consistent data view, and facilitate management and analysis.

[0117] Further, the image processing unit includes a pixel processing unit and an image correction unit;

[0118] The pixel processing unit is configured to sequentially perform pixel normalization, pixel centering, data augmentation, noise reduction processing, size adjustment, and color space conversion on the initial remote sensing image, to obtain a target image for subsequent use;

[0119] Specifically, pixel normalization is used to adjust the brightness and contrast of the image, so that the pixel values of all images fall within the same range, which is 0 to 1 or -1 to 1 in this embodiment, and this helps to accelerate the convergence speed of the model and improve the performance of the model.

[0120] Pixel centering centers the pixel values by subtracting the average value of the image data, which contributes to the stability and convergence of model training.

[0121] Data augmentation includes operations such as rotating, scaling, and shearing the image, which are used to increase the diversity of the dataset, prevent overfitting, and improve the generalization ability of the model.

[0122] Noise reduction processing is used to reduce random noise in the image, so as to improve image quality and make it easier for the model to capture important image features.

[0123] Size adjustment is used to adjust all images to the same size, ensuring that the network can accept input data of a uniform shape.

[0124] Color space conversion is used to convert the image from the RGB color space to other color spaces, such as HSV or LAB.

[0125] This embodiment can achieve the following effects through the above image processing process:

[0126] Improve the accuracy of feature extraction: by eliminating irrelevant information and enhancing important features, it helps to improve the reliability of subsequent feature extraction, image segmentation, matching and recognition.

[0127] Improve model performance: the preprocessed image data can enhance the performance of the machine learning model and avoid the occurrence of overfitting.

[0128] Speed ​​up processing: For example, grayscale can reduce the amount of data that needs to be processed, thereby improving the processing speed of the entire application system.

[0129] Data consistency: Ensures the consistency of input data, which facilitates effective model learning.

[0130] Reduce information redundancy: Remove unnecessary information and retain key features that are helpful for model training.

[0131] Preventing model overfitting: Increase data diversity through methods such as random pruning and noise reduction to enable the model to have better generalization ability.

[0132] The image correction unit is used to sequentially perform radiometric correction, geometric correction, image registration, image fusion, light and color homogenization, image stitching, and image cropping on the image processed by the pixel processing unit in order to remove the influence of the atmosphere, clouds, etc., and obtain accurate surface information and target optical remote sensing images to ensure the accuracy and usability of the data.

[0133] Further optimization schemes include radiation correction. Radiation correction refers to the correction of systematic and random radiation distortion or aberration caused by external factors, data acquisition and transmission systems, to eliminate or correct image distortion caused by radiation errors.

[0134] When using sensors to observe the reflected or radiated energy of a target, the difference between the measured value and the target's physical quantities such as spectral reflectance or spectral radiance is called radiometric error. Radiometric error causes distortion in remote sensing images, affecting their interpretation and calculation; therefore, it must be eliminated or reduced. It should be noted that many factors contribute to the distortion of radiometric values ​​in remote sensing images. Besides distortion caused by the sensor's sensitivity characteristics, factors such as field of view, solar angle, topographic relief, and atmospheric absorption and scattering also have a strong influence.

[0135] Radiometric correction of remote sensing images mainly includes three aspects:

[0136] Radiation errors caused by the sensitivity characteristics of sensors, such as edge light reduction caused by the non-uniformity of optical lenses, and radiation distortion caused by the sensitivity characteristics of photoelectric conversion systems;

[0137] Radiation errors caused by differences in lighting conditions, such as radiation distortion correction caused by different solar altitude angles, and radiation distortion correction caused by ground tilt and undulation;

[0138] Correction for radiation errors caused by atmospheric scattering and absorption.

[0139] Radiation correction is divided into two parts: radiation calibration and atmospheric correction.

[0140] Radiometric calibration is a process where users need to calculate the spectral reflectance or spectral radiance of ground objects, or compare images acquired at different times and by different sensors. This process involves converting the grayscale values ​​of an image into absolute radiance.

[0141] This invention is based on an on-orbit calibration method for sub-pixel targets. It uses a reflection point source with a sufficiently small spatial scale and sufficiently bright radiance as the sub-pixel reference target, and relies mainly on ground synchronous measurements, supplemented by simplified radiative transfer calculations, to achieve high-precision calibration of high-resolution optical remote sensing satellites across the entire dynamic range.

[0142] High-resolution optical remote sensing satellites perform on-orbit PSF detection of sub-pixel targets using a direct detection method based on the physical definition of the point spread function. A parametric point spread model is employed to fit the response data of each sub-pixel target to a surface, determining the peak position of each target's response. Then, sub-pixel interpolated point spread responses are obtained by positional registration of the sub-pixel target response values ​​or count values ​​(DN values) of a 3x3 array with non-integer pixel intervals, reducing the influence of system sampling effects and random noise. As shown in the figure, X and Y represent the horizontal and vertical coordinate axes of the sub-pixel target image. The point spread function of the optical remote sensing satellite imaging system is obtained by fitting a parametric Gaussian model, separating the response signals of sub-pixel target reflection from background radiation such as path radiation, Earth-atmosphere coupling radiation, and dark current. This allows for statistical analysis of the optical remote sensing satellite's radiative response to sub-pixel targets. Performing a Discrete Fourier Transform on the obtained system point spread function, taking the modulus, and normalizing it yields the modulation transfer function (MTF) of the optical remote sensing satellite imaging system, which is used to evaluate the imaging quality of the optical remote sensor system.

[0143] Atmospheric correction for visible and near-infrared satellite imagery includes atmospheric absorption correction, atmospheric path radiation correction, and proximity effect correction.

[0144] This invention develops an adaptive atmospheric correction algorithm based on equivalent average background reflectance, which can effectively remove proximity effects in sub-meter level optical satellite imagery. The adaptive feature of the algorithm lies in its ability to adjust the contribution weight of each background pixel to the proximity effect based on atmospheric molecular optical thickness, aerosol optical thickness, the spatial distance between background and target pixels, and the difference between the reflectance of the background and target pixels. The adaptive atmospheric correction algorithm can be implemented through the following four steps:

[0145] Step 1: Select cloudless or lightly cloudless satellite imagery. Currently, this is done manually by selecting cloudless or lightly cloudless satellite imagery from blue-band and near-infrared satellite images. Blue-band and near-infrared satellite imagery are used to detect clouds over land and ocean, respectively.

[0146] The second step is to perform radiometric calibration on the original satellite image to obtain the apparent radiance satellite image, and then calculate the apparent reflectance of each pixel in the satellite image. The spatial coordinates of each pixel can be omitted here.

[0147] Step 3: Perform atmospheric absorption and intrinsic atmospheric reflectivity correction on satellite images that reflect reflectivity.

[0148] Step 4: Calculate the equivalent average background reflectance of each pixel in the satellite image and perform proximity effect correction on a pixel-by-pixel basis.

[0149] Geometric correction: When a remote sensing image undergoes geometrical changes, resulting in distortions such as uneven rows and columns, inaccurate pixel size correspondence to ground size, or irregular changes in the shape of ground features, it indicates that the remote sensing image has experienced geometric distortion. The overall deformation of a remote sensing image (relative to the true shape of the ground) is the result of a combination of translation, scaling, rotation, twisting, bending, and other deformations. Distorted images pose difficulties for quantitative analysis and positional registration. Therefore, after receiving remote sensing data, the receiving department first performs correction, which is often based on various parameters of the remote sensing platform, the Earth, and the sensor. However, after receiving the product, users may require further geometric correction due to different usage purposes, projections, or scales. Geometric correction generally includes fine correction and orthorectification.

[0150] Fine calibration: This method uses ground control points to correct geometric distortions in remote sensing images caused by various factors. Simply put, it's similar to topographic map calibration, providing accurate latitude and longitude information after correction. Fine calibration is suitable for flat terrain where elevation information is not required, or for terrain with significant undulations and no elevation information available. Sometimes, even after initial calibration based on various parameters of the remote sensing platform, the requirements may still not be met. In such cases, this method can be used for registration and correction of remote sensing images relative to ground coordinates, registration and correction of remote sensing images relative to map projection coordinate systems, and geometric registration and composite analysis between different types or time-phase remote sensing data to obtain more accurate results.

[0151] Orthorectification: Orthophoto creation typically involves selecting ground control points on the image and using previously acquired digital elevation model (DEM) data for that area to simultaneously correct tilt and projection difference, resampling the image into an orthophoto. Multiple orthophotos are then stitched together, color balanced, and cropped within a defined range to create an orthophoto map. Orthophotos possess both topographic and imagery characteristics, are rich in information, and can serve as a data source for Geographic Information Systems (GIS), thus enriching the representation of GIS data.

[0152] Further optimization schemes suggest that the image registration problem can be divided into two parts: radiosity transformation (i.e., grayscale transformation) and spatial transformation. Radiosity transformation addresses the mapping relationship between radiosities of images from different modalities, while spatial transformation addresses the geometric transformations between images. Radiosity transformation is less discussed in feature-based registration methods but has a significant impact on pixel-based methods. Solving the spatial transformation problem is the core issue in image registration, typically involving two parts: first, determining the geometric transformation model based on prior knowledge; and second, obtaining the parameters of this transformation model using an optimization algorithm.

[0153] The image registration methods of the present invention include four types: manual registration, Fourier transform and transform domain methods, pixel-based or distribution-based methods, and feature-based methods.

[0154] Image fusion comprises three stages: image data preprocessing, image fusion, and image application. This invention primarily categorizes image data fusion into three levels: pixel-level, feature-level, and decision-level fusion.

[0155] Pixel-level image fusion directly processes registered images using a specific fusion algorithm to generate a new image, and then extracts and represents features from the fused image. The process involves directly processing the pixel grayscale values ​​of the original image data to retain most of the original information about the scene and its features.

[0156] Feature-level image fusion involves extracting and analyzing target feature information from various source data. This includes: first, extracting prominent features of ground objects in the image (such as contours, textures, and edges), synthesizing them into a feature information set; then, performing feature vector analysis on this feature information set, and training a training subset based on the analysis results; finally, fusing the feature training subset to obtain the fused image. Feature-level image fusion reduces computational load and improves efficiency, and compared to pixel-level fusion, it does not require higher image registration accuracy. Fusion methods include Bayesian estimation, cluster analysis, and entropy methods.

[0157] Decision-level image fusion, based on the extracted feature information, processes it using relevant fusion methods to provide a basis for decision-making. Its fusion algorithms mainly include Bayesian estimation, neural network methods, and expert systems.

[0158] Further optimization measures address several issues that arise during image capture. These include lens distortion, variations in ground reflection characteristics due to different shooting times, varying lighting conditions, and terrain, all contributing to image quality problems such as uneven illumination, fog, shadows, and color casts. These issues significantly impact the interpretation of ground features. Therefore, to ensure color and tone consistency in the orthorectified remote sensing images and achieve seamless stitching across large areas, it is essential to perform color and tone equalization on individual orthorectified images and to balance the tone and brightness across multiple orthorectified images.

[0159] The specific solution is as follows: First, the original image is appropriately resampled, which is the orthorectification process of a single image, to ensure that the image is corrected to a suitable resolution. Second, while ensuring that the image itself is not distorted, the quality issues of the single image are addressed, including removing haze, processing shadows, and correcting color cast. Then, based on the quality issues of the single images, the consistency of multiple images is processed to ensure that all images maintain a consistent tone and brightness overall. As a result, after all the orthorectified images are mosaicked, there will be no obvious differences in tone and brightness, ensuring the overall visual effect. Finally, when mosaicking the edges, the trend and feathering value of the edge lines are adjusted to ensure that there are no obvious mosaicking lines on the finished image.

[0160] The main process of image stitching can be summarized as follows:

[0161] 1) Image preprocessing and feature point extraction

[0162] Typically, before extracting image feature points, the image needs to be processed in some way, such as removing noise or performing certain transformations. This step is included in the feature point extraction algorithm. Feature points, as a type of local feature of an image, are the most commonly used and best represent the attributes and uniqueness of an image, such as corner points, inflection points, points with high curvature, and intersection points.

[0163] 2) Matching of points with the same name

[0164] By employing a specific matching strategy, the corresponding position of the target point in the image to be stitched is found in the reference image from a large number of feature points. A good matching algorithm should have high accuracy, high precision, and fast computation speed.

[0165] 3) Image registration and transformation model establishment

[0166] This involves employing a more precise matching strategy to eliminate inaccurate and incorrectly matched point pairs from the already matched pairs. Based on this, a transformation relationship model between the two images is determined, and the parameter values ​​in the model are calculated. This step is often referred to as fine matching. This step is the most crucial in image stitching. The images to be stitched may undergo various transformations such as translation, rotation, and scaling. If the most accurate pairs of corresponding points are not retained and inaccurate pairs are eliminated, even a small error in a single point pair can lead to various unpredictable deformations in the stitched image, or even prevent the formation of a complete image.

[0167] 4) Unify coordinate system and reconstruct the system

[0168] Based on the established transformation relationship model, the pixels in the images to be stitched are transformed to the coordinate system of the reference plane, completing a unified coordinate transformation. Image fusion mainly smooths the boundaries of the image stitching, merging the overlapping areas of the stitched images to obtain a smooth and seamless panoramic image. Since no two adjacent images can be exactly the same under identical acquisition conditions, some image characteristics that should be identical, such as the lighting characteristics, will not be exactly the same in the two images. Therefore, some noticeable seams will appear when images are stitched together, so fusion technology is needed to make the seams between images less noticeable and the stitching more natural.

[0169] Specifically, this is accomplished using a minimum grayscale difference image stitching algorithm, and its main steps include:

[0170] 1) Each pixel in the first row of the difference image corresponds to the starting point of each mosaic line, and the energy value of the first row of the difference image is recorded as the starting energy value.

[0171] 2) Starting from the starting point, each point on the tessellation line is added to the gray value of the three nearest neighbor pixels in the next row. By comparing the sum of the three pixel values, the column containing the minimum value is taken as the direction of the next tessellation point of the tessellation line. At the same time, the energy value of this tessellation line is updated to be the sum of the energy values ​​of the upper and lower rows.

[0172] 3) Iterate through step 2 until the last row of mosaic lines on the difference image is found. Select the minimum energy value from all mosaic lines as the minimum gray-level difference mosaic line.

[0173] 4) Subsequently, a seam elimination algorithm was used to eliminate the color difference on both sides of the inlay line.

[0174] Furthermore, the model building module includes a model building unit, a model training unit, and a fusion unit;

[0175] Model building units are used to build convolutional neural network models and recurrent neural network models;

[0176] The model training unit is used to train the convolutional neural network model and the recurrent neural network model respectively using image data and text data;

[0177] The fusion unit is used to fuse trained convolutional neural network models and recurrent neural network models to obtain a multimodal neural network model; it is also used to integrate image data and text data to obtain an integrated dataset.

[0178] Furthermore, the model building unit includes a first network building unit and a second network building unit;

[0179] The first network building unit is used to construct a convolutional neural network and generate a group of convolutional neural networks based on the convolutional neural network. The structure of the convolutional neural network includes a one-dimensional dilated convolutional module, an LSTM module, and a fully connected neural network module. The training process uses cross-entropy (CE) as the loss function and uses the backpropagation algorithm to optimize the model.

[0180] Further optimization of the scheme involves a one-dimensional dilated convolutional module consisting of six convolutional layers. The base kernel size is 5, and the dilation coefficients of each layer are 2, 4, 6, 2, 4, and 6, respectively. Skip connections are added to each layer to improve network performance and the stability of training gradients. The entire module performs convolutions along the time dimension. First, the 66 channels are transformed into 128 channels through a one-dimensional convolution, which serves as an upsampling function. The subsequent six one-dimensional dilated convolutional layers maintain the 128 channels, continuously expanding the receptive field through convolution to obtain preliminary feature maps. Input data: a T×66 sequence with a time length of T and 66 feature vectors; Upsampling layer: 128 convolutional kernels of length 5, with a dilation coefficient of 1, input T×66, output T×128;

[0181] The first layer of dilated convolution: 128 convolution kernels of length 5, with a dilation coefficient of 2, input of T×128, and output of T×128;

[0182] The second layer of dilated convolution: 128 convolution kernels of length 5, with a dilation factor of 4, input of T×128, and output of T×128;

[0183] The third layer of dilated convolution: 128 convolution kernels of length 5, with a dilation coefficient of 6, input of T×128, and output of T×128;

[0184] The fourth layer of dilated convolution has 128 kernels of length 5, with a dilation factor of 2. The input is T×128 and the output is T×128.

[0185] The fifth layer of dilated convolution has 128 kernels of length 5, a dilation factor of 4, an input of T×128, and an output of T×128.

[0186] The sixth layer of dilated convolution has 128 kernels of length 5 and a dilation factor of 6. The input is T×128 and the output is T×128.

[0187] The LSTM module further extracts temporal features. Compared to RNNs, it not only effectively addresses the issue of reliance on long sequence processing capabilities but also learns temporal features autonomously based on the characteristics of Long Short-Term Memory (LSTM) networks, determining which features require short-term memory and which require long-term memory, thus enabling better extraction of different features in temporal contexts. It is a unidirectional LSTM with an input dimension of 128 and an output dimension of 128, using the leaky-ReLU activation function with a coefficient of 0.2. The last unit of the computation result is extracted and fed into the fully connected neural network module for further processing. LSTM layer: input T×128, output T×128; First fully connected layer: input 1×128, output dimension 1×64; Second fully connected layer: input 1×64, output 1×32; Third fully connected layer: input 1×32, output 1×16; Fourth fully connected layer: input 1×16, output 1×8, where 8 represents the number of categories.

[0188] Before computation, the fully connected neural network module resizes the Hidden Embedding into a long vector of (1, T×hidden_size), then feeds it into the fully connected neural network, and finally passes it through a softmax layer to obtain the final image detection and recognition result.

[0189] Furthermore, before generating a swarm of convolutional neural networks (CNNs) based on CNNs, a BatchNorm2d layer, a non-linear ReLU layer, and an attention mechanism module are added after the two-dimensional convolutional layers of two identical CNN models to generate a Siamese adversarial network (SAN) model. This SAN model is then iteratively replicated to generate a swarm of CNNs. A genetic algorithm is then used to iterate through the swarm multiple times until the optimal model is obtained, at which point the iteration ends, resulting in a genetic neural network.

[0190] The second network building unit is used to construct recurrent neural networks and introduces the attention mechanism into the recurrent neural network for optimization, generating an improved recurrent neural network.

[0191] Furthermore, the model training unit includes a first training unit, a second training unit, and a third training unit;

[0192] The first training unit includes an iterative computation unit, which is used to iterate the convolutional neural network group multiple times using a genetic algorithm until the optimal model is obtained and the iteration ends to obtain the genetic neural network.

[0193] The second training unit includes a pre-training unit, an iterative training unit, and a transfer learning unit;

[0194] The pre-training unit is used to pre-train the improved recurrent neural network using the Transformer model to generate a pre-trained model; the improved recurrent neural network is generated by introducing the constructed attention mechanism module and residual module into the recurrent neural network.

[0195] The iterative training unit is used to iteratively train the pre-trained model using a sample set, and then optimize the trained model using an optimizer to generate an optimized model.

[0196] The transfer learning unit is used to freeze the low-level parameters of the optimization model and optimize the high-level parameters of the optimization model to generate the target recurrent neural network.

[0197] The third training unit is used to train the target recurrent neural network using the output of the genetic neural network, thereby generating a multimodal neural network model.

[0198] Furthermore, the fusion unit includes a model fusion unit and a data fusion unit;

[0199] The data fusion unit includes:

[0200] The weight generation unit is used to assign weights to image data and text data based on the importance of the data, and to obtain image weight data and text weight data.

[0201] The weight calculation unit is used to multiply the image weight data and text weight data with their corresponding weights to obtain a weighted value;

[0202] The summation unit is used to add up the weighted values ​​to obtain the integrated dataset.

[0203] Furthermore, the iterative computation unit includes:

[0204] The network structure encoding unit is used to encode the structural parameters of the convolutional neural network group into chromosomes for a genetic algorithm, thereby obtaining the network chromosome structure.

[0205] The function definition unit is used to construct a fitness function based on detection rate, false detection rate, and positioning accuracy.

[0206] The iterative unit is used to randomly generate a set of initial CNN chromosome structures based on the network chromosome structure; the genetic unit is used to iterate the initial CNN chromosome structures based on the fitness function, select the CNN structure corresponding to the individual with the highest fitness in the last generation of the iteration and decode it to obtain the genetic neural network.

[0207] The pre-trained units include:

[0208] Module building units are used to build attention mechanism modules and residual modules;

[0209] The attention mechanism introduction unit is used to introduce the attention mechanism module into the first three hidden layers of the recurrent neural network to obtain an attention classification model;

[0210] The residual introduction unit is used to introduce residual modules into the attention classification model to generate an improved recurrent neural network.

[0211] Genetic units include:

[0212] The selection subunit is used to select the initial CNN chromosome structure through the fitness function to obtain the selected parent CNN chromosome structure.

[0213] The crossover subunit is used to perform a crossover operation on the selected parent CNN chromosome structure to generate a new individual;

[0214] Mutant subunits are used to introduce randomness, mutating new individuals to obtain mutated individuals;

[0215] Evolutionary iteration subunits are used to iterate and evolve the CNN chromosome structure multiple times through selection subunits, crossover subunits, and mutation subunits to select the individual with the highest fitness.

[0216] The decoding subunit is used to decode the individual with the highest fitness to obtain the genetic neural network.

[0217] Furthermore, the fusion unit includes a model fusion unit and a data fusion unit;

[0218] The model fusion unit trains the convolutional neural network using the image dataset from the historical dataset and the recurrent neural network model using the text dataset from the historical dataset. Finally, the trained convolutional neural network and recurrent neural network models are connected through a fully connected layer to obtain a multimodal neural network model.

[0219] The process of obtaining the structure of a multimodal neural network also includes:

[0220] Construct an SE module and adaptively adjust the weight learning process of the SE module in a timely manner to obtain an SE module with generalization ability; insert the SE module with generalization ability after the activation function of each convolution in the convolutional neural network of the multimodal neural network structure.

[0221] The data fusion unit includes:

[0222] The weight generation unit is used to assign weights to image data and text data based on the importance of the data, and to obtain image weight data and text weight data.

[0223] The weight calculation unit is used to multiply the image weight data and text weight data with their corresponding weights to obtain a weighted value;

[0224] The summation unit is used to add up the weighted values ​​to obtain the integrated dataset.

[0225] Furthermore, the trend prediction module is used to predict agricultural carbon emissions based on the carbon emission prediction model, and the process of obtaining the carbon emission prediction results includes:

[0226] Carbon emission calculations are performed based on the carbon emission calculation model in the carbon emission prediction model to obtain carbon emission calculation results.

[0227] Carbon absorption calculations are performed based on the carbon emission reduction model in the carbon emission prediction model to obtain carbon absorption results.

[0228] A carbon neutrality assessment calculation model is constructed based on the carbon emission calculation results of the carbon emission calculation model and the carbon absorption results of the carbon emission reduction model. The carbon neutrality assessment calculation model is used to evaluate the carbon neutrality potential of agricultural carbon emissions and obtain carbon neutrality assessment results.

[0229] Furthermore, this system also includes an intelligent decision support module. Based on this module, decision-making algorithms based on data mining and machine learning are developed to analyze the optimal ratio of soil salinity, moisture, and fertilizer. Irrigation and fertilization plans are recommended according to crop type, growth cycle, and soil condition. Users can also customize parameters based on experience to adjust the recommended plans of the decision model.

[0230] Furthermore, the system also includes a precision irrigation module, which controls the irrigation amount based on the soil moisture and salinity status, and automatically adjusts the irrigation amount to avoid excessive irrigation that could lead to salt accumulation.

[0231] Controlling irrigation volume specifically includes selecting irrigation methods and optimizing irrigation time;

[0232] Among them, the irrigation method selection supports multiple irrigation methods such as drip irrigation and sprinkler irrigation, and the best irrigation method can be selected according to crop needs and soil conditions.

[0233] Optimizing irrigation time involves analyzing meteorological data to select the best irrigation time and reduce water evaporation.

[0234] Furthermore, the system also includes a precision fertilization module, which controls fertilizer application. Specifically, it formulates a reasonable fertilization plan based on the crop growth cycle and soil fertility, automatically adjusts the fertilizer application amount according to the soil fertilizer concentration and crop needs, and uses mechanical or pneumatic fertilization equipment to ensure that the fertilizer is evenly distributed in the soil.

[0235] Furthermore, the system also includes a user interaction module, which includes a data visualization unit, an operation interface unit, a report generation unit, and a fault alarm unit.

[0236] Data visualization visually displays the distribution of carbon emissions through charts and maps.

[0237] The user interface provides a simple interface that allows users to set parameters and start irrigation and fertilization tasks.

[0238] The report generation function automatically generates monitoring reports and fertilization recommendations, and supports printing and exporting.

[0239] The fault alarm unit is used to automatically send alarm information to the user when the system detects abnormal carbon emissions.

[0240] This invention employs Internet of Things (IoT) technology and remote sensing images for data acquisition, which is more efficient and accurate than traditional manual acquisition methods.

[0241] This invention employs cloud computing technology for data processing, which is more efficient and reliable than traditional local processing methods.

[0242] This invention employs a multimodal network model for carbon emission prediction, which is more accurate and comprehensive than the traditional single-modal model.

[0243] By implementing the cloud-based agricultural carbon emission management system of this invention, the efficiency of agricultural carbon emission data collection and analysis can be significantly improved, helping to reduce the carbon footprint. Real-time visualization of carbon emission data enables users to better understand their carbon emission situation and make informed management decisions. Furthermore, the widespread application of this system helps improve the level of sustainable agricultural development and is of great significance for environmental protection and climate change mitigation.

[0244] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cloud computing-based agricultural carbon emission management system, characterized in that, include: The data acquisition module is used to collect carbon emission data generated by agricultural activities; A data processing module, connected to the data acquisition module, is used to preprocess the carbon emission data and then extract its features. The model building module, connected to the data processing module, is used to build a multimodal network model and train the multimodal network model to obtain a carbon emission prediction model. The trend prediction module is connected to the model building module and is used to predict agricultural carbon emissions based on the carbon emission prediction model to obtain carbon emission prediction results. The control and management module is connected to the trend prediction module and is used to regulate and manage emissions based on the carbon emission prediction results.

2. The cloud computing-based agricultural carbon emission management system according to claim 1, characterized in that, The data acquisition module includes a data acquisition unit and an image acquisition unit; The data acquisition unit is used to acquire parameter data in agricultural production activities; The parameter data includes, but is not limited to, environmental parameter data, agricultural production activity data, agricultural input data, agricultural machinery usage data, agricultural waste data, livestock and poultry breeding data, farmland management data, energy consumption data, and agricultural product supply chain data. The image acquisition unit is used to acquire remote sensing image data in agricultural production activities; The remote sensing image data includes, but is not limited to, farmland remote sensing images, crop growth images, agricultural facility images, livestock monitoring images, agricultural waste images, soil condition images, water resource utilization images, and agricultural activity images.

3. The cloud computing-based agricultural carbon emission management system according to claim 1, characterized in that, The data processing module includes a data processing unit and an image processing unit; The data processing unit is used to denoise the original data, convert it into text data, and then merge texts from different sources and formats. The image processing unit is used to perform pixel preprocessing on the initial remote sensing image and then perform image correction to obtain the target remote sensing image.

4. The cloud computing-based agricultural carbon emission management system according to claim 3, characterized in that, The data processing unit includes a data cleaning unit, a text processing unit, and a data integration unit. The data cleaning unit is used to remove noise and outliers from the original data and fill in missing data. The text processing unit is used to perform word segmentation, dictionary creation, text vectorization, and text normalization on the text data in sequence. The data integration unit is used to unify and merge data from different sources and formats, establish relationships between data, and obtain a data view.

5. The cloud computing-based agricultural carbon emission management system according to claim 3, characterized in that, The image processing unit includes a pixel processing unit and an image correction unit; The pixel processing unit is used to sequentially perform pixel normalization, pixel centering, data enhancement, noise reduction, size adjustment and color space conversion on the initial remote sensing image to obtain the target image for subsequent use; The image correction unit is used to sequentially perform radiometric correction, geometric correction, image registration, image fusion, light and color equalization, image stitching, and image cropping on the image processed by the pixel processing unit to obtain the target optical remote sensing image.

6. The cloud computing-based agricultural carbon emission management system according to claim 1, characterized in that, The model building module includes a model building unit, a model training unit, and a fusion unit; The model building unit is used to build convolutional neural network models and recurrent neural network models; The model training unit is used to train the convolutional neural network model and the recurrent neural network model respectively using image data and text data; The fusion unit is used to fuse the trained convolutional neural network model and recurrent neural network model to obtain a multimodal neural network model; it is also used to integrate image data and text data to obtain an integrated dataset.

7. The cloud computing-based agricultural carbon emission management system according to claim 6, characterized in that, The model building unit includes a first network building unit and a second network building unit; The first network building unit is used to build a convolutional neural network and generate a group of convolutional neural networks based on the convolutional neural network; The second network building unit is used to build a recurrent neural network and introduce an attention mechanism into the recurrent neural network for optimization, thereby generating an improved recurrent neural network.

8. The cloud computing-based agricultural carbon emission management system according to claim 6, characterized in that, The model training unit includes a first training unit, a second training unit, and a third training unit; The first training unit includes an iterative computation unit, which is used to iterate the convolutional neural network group multiple times using a genetic algorithm until the optimal model is obtained and the iteration ends to obtain the genetic neural network. The second training unit includes a pre-training unit, an iterative training unit, and a transfer learning unit; The pre-training unit is used to pre-train the improved recurrent neural network using the Transformer model to generate a pre-trained model; the improved recurrent neural network is generated by introducing the constructed attention mechanism module and residual module into the recurrent neural network. The iterative training unit is used to iteratively train the pre-trained model using a sample set, and then optimize the trained model using an optimizer to generate an optimized model. The transfer learning unit is used to freeze the low-level parameters of the optimization model, optimize the high-level parameters of the optimization model, and generate the target recurrent neural network. The third training unit is used to train the target recurrent neural network using the output of the genetic neural network to generate a multimodal neural network model.

9. The cloud computing-based agricultural carbon emission management system according to claim 6, characterized in that, The fusion unit includes a model fusion unit and a data fusion unit; The data fusion unit includes: A weight generation unit is used to assign weights to the image data and the text data based on the importance of the data, thereby obtaining image weight data and text weight data. The weight calculation unit is used to multiply the image weight data and text weight data by their corresponding weights to obtain a weighted value; The summation and integration unit is used to add the weighted values ​​to obtain the integrated dataset.

10. The cloud computing-based agricultural carbon emission management system according to claim 8, characterized in that, The iterative calculation unit includes: The network structure encoding unit is used to encode the structural parameters of the convolutional neural network group into chromosomes for a genetic algorithm, thereby obtaining the network chromosome structure. The function definition unit is used to construct a fitness function based on detection rate, false detection rate, and positioning accuracy. An iterative unit is used to randomly generate a set of initial CNN chromosome structures based on the network chromosome structure; A genetic unit is used to iterate the initial CNN chromosome structure based on the fitness function, select the CNN structure corresponding to the individual with the highest fitness in the last generation of the iteration, and decode it to obtain the genetic neural network. The pre-training unit includes: Module building units are used to build attention mechanism modules and residual modules; An attention mechanism introduction unit is used to introduce the attention mechanism module into the first three hidden layers of a recurrent neural network to obtain an attention classification model; A residual introduction unit is used to introduce the residual module into the attention classification model to generate the improved recurrent neural network; The genetic unit includes: The selection subunit is used to select the initial CNN chromosome structure through the fitness function to obtain the selected parent CNN chromosome structure; The crossover subunit is used to perform a crossover operation on the selected parent CNN chromosome structure to generate a new individual; The mutation subunit is used to introduce randomness and mutate the new individual to obtain a mutated individual; The evolutionary iteration subunit is used to iterate and evolve the CNN chromosome structure multiple times through the selection subunit, crossover subunit, and mutation subunit to select the individual with the highest fitness. The decoding subunit is used to decode the individual with the highest fitness to obtain the genetic neural network.

Images