An intelligent early warning system for high-altitude monitoring of land type changes based on cloud computing

Abstract

The present invention discloses a cloud computing-based high-altitude monitoring land type change intelligent early warning system, comprising: a target detection module for detecting the monitoring area using a target detection model and statistically obtaining the number of people and vehicles per unit time; an abnormality judgment module for obtaining a carbon emission impact index based on the cultivated land area of ​​the monitoring area, the number of people and vehicles per unit time, and activating a data acquisition module when the carbon emission impact index exceeds a preset threshold; a data acquisition module for acquiring image data of the monitoring area and transmitting it to a cloud platform; a land type change detection module for acquiring historical images of the previous stage and using a cultivated land change detection model to determine whether there has been an abnormal decrease in cultivated land area; if so, an early warning signal is sent; if not, a normal signal is sent. The present invention relates to the field of land resource monitoring technology and solves the technical problems of existing cultivated land change detection methods, such as low detection efficiency, long detection time consumption, and lack of intelligence.

Description

Technical Field

[0001] The present invention belongs to the field of land resource monitoring and involves deep learning technology, specifically a cloud computing-based high-altitude monitoring land type change intelligent early warning system. Background Art

[0002] Land is a fundamental resource for human survival and development, and monitoring land use change is crucial in modern society. Land use change not only affects the balance and stability of the ecological environment but is also closely related to economic development, food security, and many other aspects. Among these, cultivated land, as a core element of agricultural production, is of paramount importance.

[0003] Existing methods for detecting cultivated land changes mainly include traditional field survey methods, detection methods based on satellite remote sensing images, and detection methods based on geographic information systems (GIS).

[0004] Traditional field surveys rely on staff visiting monitoring areas, regularly monitoring cultivated land conditions through on-site measurements and detailed record-keeping. This requires significant human, financial, and time resources. Satellite remote sensing imagery-based detection methods utilize satellites to acquire remote sensing images from different time periods, then conduct in-depth analysis and processing of these images to identify changes in cultivated land. However, their accuracy relies heavily on the resolution of satellite remote sensing images, and the acquisition cycle is long. Data processing is complex, and they lack intelligent automated analysis capabilities, making it difficult to detect non-agricultural conversion of cultivated land in a timely manner. GIS technology can integrate diverse geographic data to perform spatial analysis of cultivated land changes and visualize them. However, establishing and maintaining a GIS database requires significant initial investment in resources, including data collection, organization, and data entry, resulting in high costs. Furthermore, GIS-based methods suffer from significant lags in data updates, making it difficult to reflect dynamic changes in cultivated land in real time. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a cloud computing-based high-altitude monitoring land type change intelligent early warning system to solve the technical problems of existing cultivated land change detection methods such as low detection efficiency, long time consumption and lack of intelligence.

[0006] To achieve the above objectives, the present invention provides a cloud computing-based high-altitude monitoring and intelligent early warning system for land type changes, comprising:

[0007] Target detection module: used to detect people and vehicles in the monitoring area using a deep learning-based target detection model, and to count the number of people and vehicles per unit time;

[0008] Abnormal judgment module: used to obtain the cultivated land area in the monitoring area and obtain the carbon emission impact index based on the number of people, vehicles and cultivated land area per unit time; and

[0009] When the carbon emission impact index exceeds the preset threshold, the data collection module is activated;

[0010] Data acquisition module: used to obtain image data of the monitoring area using high-altitude measurement tools, obtain several current images, and transmit the current images to the cloud platform;

[0011] Land type change detection module: used to obtain historical images of the previous stage, and use the cultivated land change detection model to determine whether there is an abnormal decrease in the cultivated land area in the current images and the historical images of the previous stage; if so, a warning signal is sent; if not, a normal signal is sent; wherein, the cultivated land change detection model is constructed based on a deep learning algorithm, and the abnormal decrease indicates that the pixel area output by the cultivated land change detection model is greater than the preset area threshold.

[0012] Based on the above-mentioned technical modules, the present invention can use the target detection model based on deep learning to obtain the carbon emission impact index of the monitored area, thereby intelligently judging whether there may be abnormal situations of illegal use of cultivated land, and then starting the data acquisition module in a targeted manner to use high-altitude measurement tools for further detection, avoiding the frequent start-up of high-altitude measurement tools, saving resources and improving monitoring efficiency; at the same time, with the help of the cloud platform to transmit and compare cultivated land data, it facilitates data storage, retrieval and large-scale calculations, which is conducive to the efficient implementation of land change detection work and provides strong support for timely and accurate early warning.

[0013] Furthermore, the process of constructing the target detection model based on deep learning includes:

[0014] A1: Collect several images through surveillance cameras, pre-process and manually annotate them to obtain an annotation set containing human and vehicle categories;

[0015] A2: Divide the images and annotation sets into training, validation, and test sets according to a preset ratio.

[0016] A3: Build a lightweight object detection network based on a deep learning algorithm. Use the training set and validation set to iteratively train and validate the lightweight object detection network to obtain a detection model with optimal validation accuracy.

[0017] In A4, the test set is input into the detection model with the best verification accuracy for accuracy evaluation. The test accuracy is then compared to see whether it reaches the preset accuracy threshold. If so, the detection model with the best verification accuracy is used as the output to obtain the target detection model. If not, the parameters of the lightweight target detection network are adjusted and the process proceeds to A3.

[0018] Furthermore, the carbon emission impact index is obtained according to the number of people, the number of vehicles and the area of ​​cultivated land in a unit time, including:

[0019] B1, collects the per capita carbon emission coefficient k1, the carbon emission coefficient of a bicycle k2, and the carbon emission coefficient per unit of cultivated land area k3;

[0020] B2, the maximum number of personnel Pmax and the maximum number of vehicles Vmax required per unit area of ​​cultivated land in the statistical monitoring area;

[0021] B3, according to the carbon emission impact index formula ICEI = [((P×k1)+(V×k2)+(A×k3)) / (P+V+A)]×√[(P / Pmax) 2 +(V / Vmax) 2 ]Get the carbon emission impact index ICEI; where P represents the number of people per unit time, V represents the number of vehicles per unit time, and A represents the cultivated land area.

[0022] The linear combination part of the carbon emission impact index formula focuses on the carbon emission contribution and relative proportion of each factor, while the root part focuses on the relationship between activity intensity and carrying capacity. Multiplying the two can obtain a more comprehensive and integrated indicator to measure the possible impact of various activities in the monitoring area on carbon emissions and cultivated land.

[0023] Furthermore, the method for obtaining the preset threshold includes:

[0024] C1, collect data on the number of people, vehicles, and cultivated land area over a certain period of time, and derive several carbon emission impact values ​​based on the carbon emission impact index formula;

[0025] C2, dividing the carbon emission impact values ​​according to a preset division rule to obtain a plurality of data sequences; wherein the data sequences are a plurality of carbon emission impact values ​​of the same time period on different dates;

[0026] C3, calculate the upper quartile Q3 and lower quartile Q1 of several carbon emission impact values ​​in each data series, and subtract Q3 from Q1 to obtain the interquartile range (IQR) of each data series;

[0027] C4, calculate the preset threshold θ for each data sequence according to the formula θ = Q3 + α × IQR; where α represents the adjustment coefficient;

[0028] C5, using database technology to store the preset threshold value of each data sequence to obtain a dynamic threshold value table; wherein the dynamic threshold value table contains a month field, a time period field, and a corresponding preset threshold value field, and the values ​​of the month and time period are defined according to the preset division rules;

[0029] C6, according to the acquisition time of the number of people and vehicles in the target detection module, obtain the preset threshold value of the corresponding time period from the dynamic threshold table.

[0030] Furthermore, the preset division rules include:

[0031] Divide a number of carbon emission impact values ​​into preset monthly intervals to obtain a number of data groups;

[0032] The carbon emission impact values ​​in the data groups are divided according to preset time periods to obtain a number of data sequences in each data group.

[0033] In agricultural production, because farming activities vary significantly at different times of the day and at different stages of the crop life cycle, the activities of people and vehicles, as well as carbon emissions, also vary accordingly. Therefore, setting up a dynamic threshold table divided by month and time period allows for flexible adjustment of thresholds based on actual conditions, accurately reflecting carbon emissions within the monitoring area at different times. This allows for more precise determination of anomalies, effectively avoiding missed or misjudgment issues caused by using fixed thresholds, and provides a more scientific and accurate basis for anomaly determination for the cloud-based high-altitude monitoring land change intelligent early warning system.

[0034] Furthermore, the transmitting of the current plurality of images to the cloud platform includes:

[0035] D1, dividing each of the plurality of images into a plurality of image blocks, and performing color space conversion on pixel values ​​of the plurality of image blocks to obtain a plurality of converted image blocks;

[0036] D2, performing discrete cosine transform on each color component of the converted image blocks to obtain a number of frequency domain coefficient matrices, and encoding the number of frequency domain coefficient matrices using a Huffman coding algorithm to obtain encoded data of the number of image blocks;

[0037] D3, using Logistic chaotic mapping to generate a chaotic sequence, and quantizing the chaotic sequence to obtain a quantized chaotic sequence;

[0038] D4, performing bit-wise XOR operation on the coded data of the image blocks and the quantized chaotic sequence to obtain a number of encrypted data, and transmitting the encrypted data to the cloud platform.

[0039] Encryption processing is a key component of system security. Through discrete cosine transforms and Huffman coding, image data compression and feature extraction are achieved, reducing redundancy and increasing transmission speed while ensuring that critical information is not lost. Logistic chaotic mapping is then used to generate chaotic sequences, whose randomness provides a reliable foundation for data encryption and enhances security. Finally, the encoded data is encrypted with the chaotic sequence using an XOR method before transmission, ensuring the confidentiality of the image data and preventing unauthorized access and tampering, allowing the image to be transmitted securely and efficiently to the cloud platform.

[0040] Furthermore, the farmland change detection model is constructed based on a deep learning algorithm, including:

[0041] F1, using a high-altitude measurement tool to collect a plurality of cultivated land images; wherein the plurality of cultivated land images include a plurality of first time phase images and a plurality of second time phase images, and the first time phase images and the second time phase images correspond one to one in the shooting area;

[0042] F2, using an image annotation tool, annotating the cultivated land change areas in a number of first-phase images and a number of second-phase images to obtain a number of cultivated land change annotation sets;

[0043] F3, dividing the plurality of cultivated land images and the corresponding plurality of cultivated land change annotation sets according to a preset division ratio to obtain a change training set, a change verification set, and a change test set;

[0044] F4: Build a change detection model based on a deep learning algorithm. Input the change training set and change validation set into the change detection model for training and validation. Then, use the change test set to test the trained and validated model to obtain a change detection model that meets the preset test requirements.

[0045] F5, takes the change detection model that meets the preset test requirements as output to obtain the cultivated land change detection model.

[0046] Furthermore, the farmland change detection model includes a target detection branch, a semantic segmentation branch, and a fusion module; wherein,

[0047] The target detection branch is used to detect non-arable land objects;

[0048] The semantic segmentation branch is used to perform pixel-level land classification on the image to obtain a cultivated land distribution map; wherein, the target detection branch and the semantic segmentation branch use a shared backbone network for feature extraction;

[0049] The fusion module is used to use the attention mechanism to fuse the output results of the target detection branch with the output results of the semantic segmentation branch to obtain the cultivated land change detection results.

[0050] The cultivated land change detection model of the present invention includes a target detection branch, a semantic segmentation branch and a fusion module; among them, the target detection branch focuses on detecting non-cultivated land objects, can accurately locate non-cultivated land objects in the image, and enhances the targetedness of the detection; the semantic segmentation branch can classify the image at the pixel level, generate a detailed cultivated land distribution map, and show the actual range and distribution of cultivated land with fine pixel-level division, making up for the shortcomings of target detection in detail presentation; the fusion module uses the attention mechanism to integrate the output results of the first two. The attention mechanism enables the model to focus on the key areas and features of cultivated land changes, fully integrates information at different levels, and combines the advantages of target detection and semantic segmentation, so that the final cultivated land change detection results can more accurately and comprehensively reflect the actual situation, effectively improving the overall detection efficiency and providing technical support for cultivated land change monitoring.

[0051] Furthermore, the architecture of the target detection branch includes:

[0052] Backbone network: Use the Focus module to slice the input data to obtain several slice images, and use the CSP-Darknet53 network as a shared backbone network to extract features from the slice images to obtain multi-scale features;

[0053] Neck network: includes the fast spatial pyramid pooling layer SPPF, the feature pyramid network FPN, and the path aggregation network PAN; among them, SPPF is used to fuse multi-scale features, FPN is used to transfer semantic information of the fused multi-scale features from top to bottom, and PAN is used to transfer positioning information of the FPN output from bottom to top;

[0054] Detection head: Use the YOLOv3 detection head to output detection results.

[0055] The backbone network of the target detection branch uses the Focus module for slicing operations, enriching the image feature dimensions while retaining key information. The CSP-Darknet53 network, as a shared backbone network, has powerful multi-scale feature extraction capabilities and can accurately capture target-related features of different scales. The fast spatial pyramid pooling layer SPPF in the neck network effectively integrates multi-scale features, avoids feature fragmentation, and ensures information integrity. The feature pyramid network FPN transmits semantic information from top to bottom, enhances the semantic expression of low-level features, and the path aggregation network PAN transmits positioning information from bottom to top. The two work together to improve feature quality, enabling the detection head to work based on high-quality features. The detection head adopts the YOLOv3 detection head, which has the advantages of fast detection speed and relatively balanced detection capabilities for targets of different scales. While ensuring a certain detection accuracy, it quickly outputs the detection results of non-arable land objects, meets the needs of real-time monitoring scenarios, and improves the practicality and work efficiency of the target detection branch.

[0056] Furthermore, the land change detection module is deployed on a cloud platform, and the cloud platform includes an image database for storing historical image data of several stages; the target detection module, anomaly judgment module and data acquisition module are deployed in a monitoring center, and the monitoring center monitors the monitoring area through several monitoring cameras, and the monitoring center and the cloud platform transmit data and signals through a communication network.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] First, the target detection module is used to identify people and vehicles and other activities related to changes in cultivated land use in the monitoring area, accurately locating potential change areas. Then, the land type change detection module is activated to detect cultivated land changes in a timely manner, achieve rapid early warning, avoid indiscriminate detection of the entire monitoring area, and improve detection efficiency.

[0059] In the process of transmitting images to the cloud platform, strict data encryption operations are used to effectively compress data volume, increase transmission speed, enhance data transmission security and confidentiality, prevent data from being stolen or tampered, and ensure the security and reliability of monitoring data while ensuring data integrity.

[0060] The cultivated land change detection model adopts a dual-branch structure with a shared backbone network, which can make full use of the same underlying feature information while performing specialized feature learning and optimization on their respective tasks. Among them, the target detection branch can accurately locate non-cultivated land objects, and the semantic segmentation branch can divide the cultivated land distribution map in detail. The fusion module uses the attention mechanism to fuse the results of the two to obtain accurate cultivated land change detection results. It can promptly detect abnormal reductions in cultivated land data, provide strong support for the protection of cultivated land resources, accurately send early warning signals, and assist relevant departments in taking timely measures to deal with problems such as illegal use of cultivated land. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0062] Figure 1 A schematic diagram of the technical process of a cloud computing-based high-altitude monitoring land type change intelligent early warning system provided by the present invention;

[0063] Figure 2 The process of constructing the dynamic threshold table provided by the present invention;

[0064] Figure 3A schematic diagram of the process of transmitting image data to a cloud platform provided by the present invention;

[0065] Figure 4 This is a schematic diagram of the framework of a cloud computing-based high-altitude monitoring land type change intelligent early warning system provided by the present invention. DETAILED DESCRIPTION

[0066] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] See also Figures 1-4 The first embodiment of the present invention provides a cloud computing-based high-altitude monitoring land type change intelligent early warning system, comprising:

[0068] Target detection module: used to detect people and vehicles in the monitoring area using a deep learning-based target detection model, and to count the number of people and vehicles per unit time;

[0069] Abnormal judgment module: used to obtain the cultivated land area in the monitoring area and obtain the carbon emission impact index based on the number of people, vehicles and cultivated land area per unit time. When the carbon emission impact index exceeds the preset threshold, the data collection module is activated;

[0070] Data acquisition module: used to obtain image data of the monitoring area using high-altitude measurement tools to obtain several current images;

[0071] Land change detection module: used to transmit the current images to the cloud platform and obtain historical images from the previous stage. The cultivated land change detection model is used to compare whether there is an abnormal decrease in cultivated land data in the current images and the historical images from the previous stage. If so, an early warning signal is sent; if not, a normal signal is sent. The cultivated land change detection model is built based on a deep learning algorithm.

[0072] It should be noted that the target detection module, anomaly judgment module, data acquisition module and land change detection module of the present invention are communicatively connected, and the target detection module, anomaly judgment module and data acquisition module are deployed in a monitoring center. The monitoring center monitors the monitoring area through a number of monitoring cameras, and the monitoring center and the cloud platform transmit data and signals through a communication network.

[0073] In this embodiment, in order to efficiently and accurately conduct preliminary screening of activities in the monitoring area, a lightweight target detection model based on a deep learning algorithm is deployed in the target detection module of the monitoring center to identify the number of people and vehicles in the monitoring area, thereby indirectly reflecting the frequency of human activities in the area and the possible resource allocation involved.

[0074] The target detection model of the present invention is constructed as follows:

[0075] First, during the data collection phase, a number of images are collected through surveillance cameras, pre-processed, and manually labeled to obtain a label set containing the categories of people and vehicles.

[0076] Next, several images and annotation sets are divided according to a preset ratio to obtain training sets, validation sets, and test sets. The training set occupies a larger proportion and is used for the initial learning and parameter adjustment of the model, so that it can fully learn the various characteristic patterns of people and vehicles. The validation set is used to evaluate and verify its performance in real time during the model training process, promptly discover problems such as overfitting or underfitting that may occur in the model, and optimize and adjust the model's hyperparameters accordingly to ensure the model's generalization ability and stability. The test set is independent of the training and validation process and is used to ultimately objectively and accurately evaluate the overall performance of the model. Its data distribution is kept as consistent as possible with the data distribution in the actual monitoring scenario to truly reflect the performance of the model in actual applications.

[0077] Next, a lightweight object detection network is constructed based on a deep learning algorithm, such as the YOLO series network architecture, and is optimized and improved. The lightweight object detection network is then iteratively trained and validated using training and validation sets. During each iteration, the model parameters are fine-tuned based on a pre-defined loss function and optimizer, reducing the error between the model's predictions and the true labels in the annotation set. This gradually improves the model's detection accuracy and performance until a detection model with optimal validation accuracy is achieved.

[0078] Finally, the test set is input into the detection model with the best verification accuracy for accuracy evaluation to obtain the test accuracy, and then compared to see whether the test accuracy reaches the preset accuracy threshold (for example, the F1 value is used as the final evaluation accuracy value); if so, the detection model with the best verification accuracy is used as the output to obtain the target detection model; if not, the parameters of the lightweight target detection network are adjusted, and retraining and verification are performed.

[0079] During normal agricultural production, the movement of people and vehicles follows certain patterns and ranges. However, any movement beyond these norms may indicate unusual activity, such as illegal construction, excessive reclamation, or unreasonable land use changes. By accurately counting the number of people and vehicles in real time and using this data as key input for subsequent anomaly detection modules, the entire monitoring system can be provided with timely and valuable clues, thereby improving the system's early warning capabilities and response speed for cultivated land change monitoring. This ensures the rational use and effective protection of land resources, protects them from illegal or inappropriate activities, and further maintains the stable order of agricultural production and the balanced development of the ecological environment.

[0080] Since starting high-altitude measurement equipment to acquire farmland images and conducting subsequent complex algorithm analysis and data processing requires a large amount of resources, it not only places extremely high demands on the hardware equipment performance of the monitoring system, but also increases the system's operating costs and energy consumption.

[0081] Therefore, in order to optimize the intelligent early warning process of cultivated land changes and reduce the early warning cost, an abnormality judgment module is set up in the present invention, which is used to obtain the carbon emission impact index based on the cultivated land area of ​​the monitoring area obtained and the number of people and vehicles per unit time counted by the target detection module. The carbon emission impact index is used to judge whether abnormal behavioral activities occur in the monitoring area, providing a decision-making basis for the startup of the data acquisition module.

[0082] In the abnormality judgment module, the carbon emission impact index is obtained based on the number of people, vehicles, and cultivated land area per unit time, including:

[0083] B1, using scientific research data and agricultural statistics, collect per capita carbon emission coefficient k1, carbon emission coefficient per vehicle k2, and carbon emission coefficient per unit of cultivated land k3 that are suitable for the characteristics of this monitoring area;

[0084] B2, the maximum number of personnel Pmax and the maximum number of vehicles Vmax required per unit area of ​​cultivated land in the statistical monitoring area;

[0085] B3, according to the carbon emission impact index formula ICEI = [((P×k1)+(V×k2)+(A×k3)) / (P+V+A)]×√[(P / Pmax) 2 +(V / Vmax) 2 ]Get the carbon emission impact index ICEI; where P represents the number of people per unit time, V represents the number of vehicles per unit time, and A represents the cultivated land area.

[0086] When the carbon emission impact index exceeds the preset threshold, it indicates that there are abnormal behaviors in the monitoring area, such as excessive non-agricultural construction activities, land compaction and damage caused by frequent entry and exit of large vehicles, and illegal reclamation or excavation caused by the gathering of people beyond the normal range.

[0087] In this embodiment, whether the carbon emission impact index exceeds a preset threshold is determined by comparing the corresponding thresholds in the dynamic threshold table obtained according to the current time period.

[0088] Specifically, we first need to collect data on the number of people, vehicles, and cultivated land area over a certain period of time (over a year) in history, and then derive several carbon emission impact values ​​based on the carbon emission impact index formula;

[0089] Then, several carbon emission impact values ​​are divided according to the preset division rules:

[0090] Divide a number of carbon emission impact values ​​into preset monthly intervals, for example, divide a year into quarters to obtain four data groups;

[0091] Then, the carbon emission impact values ​​in the data group are divided according to preset time periods. For example, a day is divided into a daytime interval and a nighttime interval. The daytime interval is from 7:00-19:00, with each interval being every two hours. The nighttime interval is from 19:00 to 7:00 the next day, which is a time period (there are usually fewer activities at night). This yields several data series in each data group. The data series are the carbon emission impact values ​​of the same time period on different dates. For example, the carbon emission impact values ​​from 7:00-9:00 every day in the first quarter are considered as a data series.

[0092] Next, calculate the upper quartile Q3 and lower quartile Q1 of several carbon emission impact values ​​in each data series, and subtract Q3 from Q1 to obtain the interquartile range (IQR) of each data series.

[0093] The preset threshold θ for each data sequence is calculated according to the formula θ = Q3 + α × IQR; where α represents the adjustment coefficient, which is obtained based on practical experience;

[0094] Finally, database technology is used to store the preset thresholds of each data series to obtain a dynamic threshold table. The dynamic threshold table contains a month field, a time period field, and a corresponding preset threshold field. The values ​​of the month and time period are defined according to the preset division rules.

[0095] For example, the month field in the dynamic threshold table is January-March, April-June, etc., and the time period field is 7:00-9:00, 9:00-11:00, etc., and then the preset thresholds in the corresponding time periods are stored to obtain the dynamic threshold table.

[0096] In this embodiment, by dividing the data in detail according to the time dimension, from months to specific time periods, different data sequences are formed. Each data sequence reflects the distribution pattern of carbon emission impact values ​​within a specific time interval, and can capture activity patterns and carbon emission characteristics that change over time, and then more accurately judge whether anomalies occur based on the time background of the real-time monitoring data, which is in line with the actual time patterns and helps to improve the accuracy and timeliness of early warnings.

[0097] When the carbon emission impact index exceeds the preset threshold, the data acquisition module will be automatically activated, and then the data acquisition module will start the high-altitude measurement tool (such as a drone) to obtain image data of the monitoring area and obtain the current images;

[0098] Then, in the land type change detection module, the current images are transmitted to the cloud platform, and the historical images of the previous stage are obtained from the cloud platform's image database. These images are input into the cultivated land change detection model for cloud computing to compare whether there is an abnormal decrease in cultivated land data. If so, an early warning signal is sent; if not, a normal signal is sent. Among them, the cloud platform's image database is used to store historical image data of several stages and deploy the land type change detection module.

[0099] During the image data transmission to the cloud platform, a rigorous data encryption process is adopted to prevent data theft. The specific operations are as follows:

[0100] First, unique identifiers are assigned to several images to obtain several image IDs;

[0101] Then, each image is divided into several image blocks, and the pixel values ​​of the several image blocks are converted into color space, for example, from the original color space to the YUV color space, so as to better extract the feature information of the image blocks and obtain the converted image blocks, thereby reducing the complexity of data processing;

[0102] Next, discrete cosine transform is performed on each color component of the converted image blocks to obtain a number of frequency domain coefficient matrices, and the frequency domain coefficient matrices are encoded using a Huffman coding algorithm to obtain encoded data of the image blocks;

[0103] The function of discrete cosine transform (DCT) is to convert image blocks from the spatial domain to the frequency domain. In the frequency domain, the energy of an image is often concentrated in a few low-frequency coefficients. This can effectively remove redundant information from the image data, achieve data compression, and highlight the main features of the image. Afterwards, the Huffman coding algorithm is used to encode the obtained frequency domain coefficient matrices. Huffman coding is a variable-length coding method based on the probability of data occurrence. It assigns codewords of different lengths according to the probability of occurrence of each element in the frequency domain coefficient matrix. Elements with high probability of occurrence are assigned shorter codewords, and vice versa. This further compresses the amount of data, reduces the scale of data that needs to be processed during transmission, and improves transmission efficiency. The final encoded data contains the key information of the image block after compression and feature extraction, which becomes the basis for subsequent encryption operations.

[0104] Afterwards, the chaotic sequence is generated using the Logistic chaotic map, and the chaotic sequence is quantized to obtain the quantized chaotic sequence;

[0105] The Logistic Chaos Map is a typical chaotic system that has good randomness, ergodicity, and high sensitivity to initial conditions. Based on specific initial values ​​and parameters, iterative operations can generate a chaotic sequence that appears disordered but actually has inherent rules. The randomness of this chaotic sequence makes it difficult to predict and crack, providing an ideal encryption material for data encryption. The generated chaotic sequence is then quantized, and its numerical range is mapped to a discrete numerical range suitable for subsequent operations with the image block encoded data. This ensures that the chaotic sequence can accurately cooperate with the encoded data in the subsequent bitwise XOR operation, further enhancing the encryption effect and ensuring that the encrypted data has higher security.

[0106] Finally, the encoded data of the image blocks and the quantized chaotic sequence are subjected to a bitwise XOR operation to generate encrypted data, which is then transmitted to the cloud platform. Bitwise XOR is a simple yet effective encryption method. It performs an XOR operation on the binary values ​​of the corresponding bits of the encoded data and the quantized chaotic sequence (where identical values ​​are 0 and different values ​​are 1). This "masks" the original encoded data with the chaotic sequence. Without the correct chaotic sequence (i.e., the decryption key), the generated encrypted data is difficult to restore to the original image block information, thus achieving encryption protection for the image data.

[0107] After the image data of the monitored area is transmitted to the cloud platform for decryption, the cloud platform will retrieve the image data taken in the previous stage, and then match the current image with the image of the previous stage one by one according to the image ID. The data is then input into the change detection model and outputs the cultivated land change map for the two time periods. When the pixel area in the cultivated land change map exceeds the preset threshold, an early warning signal will be immediately activated.

[0108] In the present invention, the farmland change detection model consists of a target detection branch, a semantic segmentation branch, and a fusion module; wherein,

[0109] The target detection branch is used to detect non-arable land objects and consists of a backbone network, a neck network, and a detection head:

[0110] In the backbone network, the Focus module is first used to slice the input data to expand the number of input channels and reduce the spatial dimension of the feature map, thereby obtaining several slice images. Next, the CSP-Darknet53 network is used as a shared backbone network to extract features from the slice images. The CSP structure of the CSP-Darknet53 network has a unique advantage. It divides the feature map into two parts for processing. One part performs the convolution operation, and the other part is directly spliced ​​with the result of the convolution operation. In this way, the diversity of feature information can be retained, allowing the network to capture richer image features, while significantly reducing the amount of calculation, improving computing efficiency, and ultimately obtaining multi-scale features.

[0111] Then, in the neck network, several important components are included: the fast spatial pyramid pooling layer SPPF, the feature pyramid network FPN, and the path aggregation network PAN. Among them, SPPF is responsible for fusing multi-scale features. It efficiently fuses features of different scales by serially passing through multiple 5x5 maxpooling layers, greatly improving computational efficiency compared to traditional correlation structures. FPN is responsible for top-down semantic information transmission, propagating high-level features with rich semantic information downward, so that features at different levels can obtain more accurate semantic guidance, which helps to better identify targets. PAN performs bottom-up positioning information transmission, further strengthening the feature fusion capability and allowing the network to locate targets more accurately.

[0112] Finally, the YOLOv3 detection head is used to output the detection results. Based on DIoULoss, the YOLOv3 detection head adds a loss for the detection box scale, making the predicted box more consistent with the ground truth box. This improves the accuracy of object detection and can more accurately detect non-arable land objects in the image.

[0113] The semantic segmentation branch is used to classify land types at the pixel level of the image to obtain a cultivated land distribution map. The backbone network of the semantic segmentation branch uses the same backbone network as the object detection branch to obtain multi-scale features. The multi-scale features are then upsampled in the decoder through multiple serial upsampling modules. After each upsampling, skip connections are added to fuse the multi-scale features to obtain fused features. Multiple convolutional layers and activation functions are then used as classifiers to perform pixel-level classification on the fused features. Finally, a semantic segmentation map with the same size as the input image is output to obtain a cultivated land distribution map.

[0114] Finally, a fusion module based on the attention mechanism is used to fuse the output of the target detection branch and the output of the semantic segmentation branch to obtain a more accurate map of cultivated land type changes.

[0115] It should be noted that the input of the cultivated land change detection model is a dual-phase image, namely the current image and the image of the previous stage. Therefore, the target detection branch and the semantic segmentation branch both process the two image data at the same time, and then obtain the non-cultivated land increase result of the target detection branch and the cultivated land decrease result of the semantic segmentation branch; further fusing the two using the fusion module can more accurately identify the cultivated land changes.

[0116] The construction process of the cultivated land change detection model in this embodiment specifically includes:

[0117] F1, using a high-altitude measurement tool to collect a plurality of cultivated land images; wherein the plurality of cultivated land images include a plurality of first-phase images and a plurality of second-phase images, and the first-phase images and the second-phase images correspond one-to-one to each other in the shooting area;

[0118] F2, using an image annotation tool, annotating the cultivated land change areas in a number of first-phase images and a number of second-phase images to obtain a number of cultivated land change annotation sets;

[0119] F3, dividing the plurality of cultivated land images and the corresponding plurality of cultivated land change annotation sets according to a preset division ratio to obtain a change training set, a change verification set, and a change test set;

[0120] F4: Build a change detection model based on a deep learning algorithm. Input the change training set and change validation set into the change detection model for training and validation. Then, use the change test set to test the trained and validated model to obtain a change detection model that meets the preset test requirements.

[0121] F5, takes the change detection model that meets the preset test requirements as output to obtain the cultivated land change detection model.

[0122] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0123] Working principle of the present invention:

[0124] A deep learning-based target detection model is used to identify the activities of people and vehicles within the monitoring area. By calculating the carbon emission impact index and comparing it with the dynamic threshold, it is determined whether there may be abnormal non-agricultural activities in the monitoring area.

[0125] During the image data transmission to the cloud platform, a variety of data processing and encryption technologies are used to achieve data compression, feature extraction and encryption protection, ensuring data integrity and confidentiality, preventing data theft or tampering, and ensuring the security and reliability of monitoring data;

[0126] The cultivated land change detection model utilizes the efficient processing capabilities of cloud computing and adopts a dual-branch structure with a shared backbone network to process image data from different angles. It then uses the attention mechanism through a fusion module to fuse the two results, making full use of multi-scale feature information to improve the accuracy and comprehensiveness of cultivated land change detection, timely discover abnormal changes in cultivated land and issue warnings, and assist in protecting cultivated land resources.

[0127] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A cloud computing-based high-altitude monitoring land change intelligent early warning system, characterized by: include: Target detection module: used to detect people and vehicles in the monitoring area using a deep learning-based target detection model, and to count the number of people and vehicles per unit time; Abnormal judgment module: used to obtain the cultivated land area in the monitoring area and obtain the carbon emission impact index based on the number of people, vehicles and cultivated land area per unit time; and When the carbon emission impact index exceeds the preset threshold, the data collection module is activated; Data acquisition module: used to obtain image data of the monitoring area using high-altitude measurement tools, obtain several current images, and transmit the current images to the cloud platform; Land change detection module: used to obtain historical images from the previous stage and use the cultivated land change detection model to determine whether there is an abnormal decrease in the cultivated land area between the current images and the historical images from the previous stage; If yes, an early warning signal is sent; If not, a normal signal is sent; wherein the cultivated land change detection model is constructed based on a deep learning algorithm, and the abnormal reduction indicates that the pixel area output by the cultivated land change detection model is greater than a preset area threshold; The carbon emission impact index is obtained based on the number of people, vehicles and cultivated land area per unit time, including: B1, collects the per capita carbon emission coefficient k1, the carbon emission coefficient of a bicycle k2, and the carbon emission coefficient per unit of cultivated land area k3; B2, the maximum number of personnel Pmax and the maximum number of vehicles Vmax required per unit area of ​​cultivated land in the statistical monitoring area; B3. The carbon emission impact index ICEI is obtained according to the formula ICEI=[((P×k1)+(V×k2)+(A×k3)) / (P+V+A)]×√[(P / Pmax)2+(V / Vmax)2]; where P represents the number of people per unit time, V represents the number of vehicles per unit time, and A represents the area of ​​cultivated land.

2. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 1 is characterized in that: The process of building the deep learning-based target detection model includes: A1: Collect several images through surveillance cameras, pre-process and manually annotate them to obtain an annotation set containing human and vehicle categories; A2: Divide the images and annotation sets into training, validation, and test sets according to a preset ratio. A3: Build a lightweight object detection network based on a deep learning algorithm. Use the training set and validation set to iteratively train and validate the lightweight object detection network to obtain a detection model with optimal validation accuracy. In A4, the test set is input into the detection model with the best verification accuracy for accuracy evaluation. The test accuracy is then compared to see whether it reaches the preset accuracy threshold. If so, the detection model with the best verification accuracy is used as the output to obtain the target detection model. If not, the parameters of the lightweight target detection network are adjusted and the process proceeds to A3.

3. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 1 is characterized in that: The method for obtaining the preset threshold includes: C1, collect data on the number of people, vehicles, and cultivated land area over a certain period of time, and derive several carbon emission impact values ​​based on the carbon emission impact index formula; C2, dividing the carbon emission impact values ​​according to a preset division rule to obtain a plurality of data sequences; wherein the data sequences are a plurality of carbon emission impact values ​​of the same time period on different dates; C3, calculate the upper quartile Q3 and lower quartile Q1 of several carbon emission impact values ​​in each data series, and subtract Q3 from Q1 to obtain the interquartile range (IQR) of each data series; C4, calculate the preset threshold θ for each data sequence according to the formula θ = Q3 + α × IQR; where α represents the adjustment coefficient; C5, using database technology to store the preset threshold value of each data sequence to obtain a dynamic threshold value table; wherein the dynamic threshold value table contains a month field, a time period field, and a corresponding preset threshold value field, and the values ​​of the month and time period are defined according to the preset division rules; C6, according to the acquisition time of the number of people and vehicles in the target detection module, obtain the preset threshold value of the corresponding time period from the dynamic threshold table.

4. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 3 is characterized in that: The preset division rules include: Divide a number of carbon emission impact values ​​into preset monthly intervals to obtain a number of data groups; The carbon emission impact values ​​in the data groups are divided according to preset time periods to obtain a number of data sequences in each data group.

5. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 1 is characterized in that: The transmitting of the current plurality of images to the cloud platform includes: D1, dividing each of the plurality of images into a plurality of image blocks, and performing color space conversion on pixel values ​​of the plurality of image blocks to obtain a plurality of converted image blocks; D2, performing discrete cosine transform on each color component of the converted image blocks to obtain a number of frequency domain coefficient matrices, and encoding the number of frequency domain coefficient matrices using a Huffman coding algorithm to obtain encoded data of the number of image blocks; D3, using Logistic chaotic mapping to generate a chaotic sequence, and quantizing the chaotic sequence to obtain a quantized chaotic sequence; D4, performing bit-wise XOR operation on the coded data of the image blocks and the quantized chaotic sequence to obtain a number of encrypted data, and transmitting the encrypted data to the cloud platform.

6. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 1 is characterized in that: The farmland change detection model includes a target detection branch, a semantic segmentation branch and a fusion module; wherein, The target detection branch is used to detect non-arable land objects; The semantic segmentation branch is used to perform pixel-level land classification on the image to obtain a cultivated land distribution map; wherein, the target detection branch and the semantic segmentation branch use a shared backbone network for feature extraction; The fusion module is used to use the attention mechanism to fuse the output results of the target detection branch with the output results of the semantic segmentation branch to obtain the cultivated land change detection results.

7. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 6 is characterized in that: The architecture of the target detection branch includes: Backbone network: Use the Focus module to slice the input data to obtain several slice images, and use the CSP-Darknet53 network as a shared backbone network to extract features from the slice images to obtain multi-scale features; Neck network: includes the fast spatial pyramid pooling layer SPPF, the feature pyramid network FPN, and the path aggregation network PAN; among them, SPPF is used to fuse multi-scale features, FPN is used to transfer semantic information of the fused multi-scale features from top to bottom, and PAN is used to transfer positioning information of the FPN output from bottom to top; Detection head: Use the YOLOv3 detection head to output detection results.

8. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 1 is characterized in that: The farmland change detection model is built based on a deep learning algorithm and includes: F1, using a high-altitude measurement tool to collect a plurality of cultivated land images; wherein the plurality of cultivated land images include a plurality of first time phase images and a plurality of second time phase images, and the first time phase images and the second time phase images correspond one to one in the shooting area; F2, using an image annotation tool, annotating the cultivated land change areas in a number of first-phase images and a number of second-phase images to obtain a number of cultivated land change annotation sets; F3, dividing the plurality of cultivated land images and the corresponding plurality of cultivated land change annotation sets according to a preset division ratio to obtain a change training set, a change verification set, and a change test set; F4: Build a change detection model based on a deep learning algorithm. Input the change training set and change validation set into the change detection model for training and validation. Then, use the change test set to test the trained and validated model to obtain a change detection model that meets the preset test requirements. F5, takes the change detection model that meets the preset test requirements as output to obtain the cultivated land change detection model.

9. The cloud computing-based high-altitude monitoring land type change intelligent early warning system according to claim 1 is characterized in that: The land change detection module is deployed on a cloud platform, and the cloud platform includes an image database for storing historical image data of several stages; the target detection module, anomaly judgment module and data acquisition module are deployed in a monitoring center, which monitors the monitoring area through several monitoring cameras, and the monitoring center and the cloud platform transmit data and signals through a communication network.

Images