Crop disease and pest monitoring system based on image recognition

Abstract

The present application relates to the field of wisdom agriculture and image recognition technology, specifically to a crop disease and pest monitoring system based on image recognition, comprising: an environment perception module, which obtains the environmental context parameters of the monitoring area through multi-modal sensor nodes; a multi-source context analysis module, which constructs a disease and pest outbreak risk model in combination with historical disease and pest distribution spatio-temporal data and calculates the disease and pest risk level; an adaptive strategy generation module, which dynamically reconstructs the image acquisition mode and feature extraction strategy according to the risk level and the edge side resource constraint state, and generates adaptive perception instructions; an image acquisition execution module, which adjusts the physical acquisition parameters and algorithm calculation density of the visual sensor in response to the instructions, and executes image acquisition and recognition; the present application breaks the traditional rigid acquisition mode and realizes the optimal balance between monitoring efficiency and device survival period.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and image recognition technology, specifically to a crop pest and disease monitoring system based on image recognition. Background Technology

[0002] In smart agriculture and crop protection applications, monitoring systems rely on accurate environmental data and pest and disease image collection to ensure the timeliness and accuracy of prevention and control measures. Field monitoring terminals usually need to combine environmental perception parameters and visual sensor images to perceive the health status of crops in real time. For monitoring crop diseases and pests, existing solutions generally adopt a fixed-frequency, periodic data acquisition architecture. This involves deploying cameras in the monitoring area, continuously acquiring images at preset fixed time intervals, and uploading them to the cloud for identification. This high-frequency data stream covers the crop growth cycle. While this solution is feasible under ideal power and network conditions, its all-weather, high-frequency acquisition mode lacks the ability to perceive environmental disease risk levels and edge resource constraints. This causes the equipment to continue performing high-load image acquisition and transmission tasks even during periods of low disease risk or when power is critical, resulting in the accumulation of invalid data. The excessive consumption of battery power is a significant issue. Furthermore, this rigid monitoring mode makes it difficult to dynamically adjust the algorithm's computational density and acquisition accuracy according to actual needs. In situations with weak networks or limited computing power in the field, it easily leads to data transmission congestion or premature network disconnection due to battery depletion, resulting in blind spots or interruptions in the monitoring process. This makes it difficult to support unattended continuous monitoring with high endurance and low maintenance costs in large-scale farmland scenarios. Therefore, establishing an adaptive sensing mechanism driven by both environmental risk and resource status, while effectively reducing equipment energy consumption and bandwidth usage, and improving the robustness of pest and disease monitoring and the equipment's lifespan, has become an urgent technical problem to be solved. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an image recognition-based crop pest and disease monitoring system. Specifically, the technical solution of this invention includes: The cloud-based management and control platform has communication connections to an environmental perception module, a multi-source context analysis module, an adaptive strategy generation module, and an image acquisition and execution module. The environmental perception module is configured to deploy multimodal sensor nodes in the monitoring area to acquire environmental context parameters. The multi-source context analysis module is configured to acquire historical spatiotemporal data of disease distribution in the monitoring area, combine this data with environmental context parameters to construct a disease outbreak risk model, and calculate the disease risk level of the monitoring area in the current time window. The adaptive strategy generation module is configured to dynamically reconstruct the image acquisition mode and feature extraction strategy based on the disease risk level and edge-side resource constraints, generating adaptive perception commands. The edge-side resource constraints include at least remaining power parameters and network bandwidth parameters. The image acquisition and execution module is configured to respond to the adaptive perception commands by adjusting the physical acquisition parameters of the visual sensor and the computational density of the image processing algorithm, acquiring and recognizing the target image, and outputting the disease monitoring results.

[0004] Preferably, the process by which the environmental perception module acquires environmental context parameters of the monitoring area includes: acquiring meteorological data from several key monitoring outposts within the monitoring area, the meteorological data including temperature, humidity, light intensity, and wind speed; activating a low-power panoramic camera to acquire a macroscopic panoramic image of the monitoring area at a preset low-frequency sampling period; extracting texture change feature vectors and spectral reflectance feature vectors from the macroscopic panoramic image; and concatenating the meteorological data, texture change feature vectors, and spectral reflectance feature vectors to generate environmental context parameters.

[0005] Preferably, the process by which the multi-source context analysis module calculates the disease risk level of the monitoring area in the current time window includes: inputting environmental context parameters into a preset disease outbreak risk model and outputting the disease occurrence probability value in the current time window; setting a high-risk threshold and a low-risk threshold, wherein the high-risk threshold value is greater than the low-risk threshold value; if the disease occurrence probability value is greater than or equal to the high-risk threshold, then the disease risk level is marked as a high-risk outbreak level; if the disease occurrence probability value is less than the high-risk threshold but greater than the low-risk threshold, then the disease risk level is marked as a potential early warning level; if the disease occurrence probability value is less than or equal to the low-risk threshold, then the disease risk level is marked as a normal safe level.

[0006] Preferably, the process of the adaptive strategy generation module generating adaptive perception instructions includes: obtaining the current remaining power parameters and network bandwidth parameters of the image acquisition execution module; if the disease risk level is high-risk outbreak level or potential early warning level, and the remaining power parameters are greater than a preset power threshold, then a high-precision perception instruction is generated, which instructs to increase the image acquisition resolution, shorten the acquisition time interval, and activate the deep neural network feature extraction channel; if the disease risk level is normal safety level, or the remaining power parameters are less than or equal to the preset power threshold, then a low-power sentinel instruction is generated, which instructs to reduce the image acquisition resolution, extend the acquisition time interval, and freeze the redundant feature extraction channels of the neural network.

[0007] Preferably, the image acquisition execution module, in response to a high-precision perception command, performs the process of acquiring and recognizing the target image, including: controlling the vision sensor to acquire the original image frame of the monitoring area at high resolution; using a lightweight saliency detection algorithm to scan the original image frame and locate the region of interest with abnormal texture or color change; cropping local image blocks according to the coordinates of the region of interest; inputting the local image blocks into a preset refined disease recognition model and outputting the disease category and confidence level of the local image blocks.

[0008] Preferably, the refined disease identification model adopts a dynamic sparse computing architecture, and its execution process includes: obtaining the computing density parameter carried in the adaptive perception instruction; dynamically adjusting the channel pruning rate of the convolutional layer in the refined disease identification model according to the computing density parameter; if the computing density parameter corresponds to a high-precision perception instruction, indicating a high computing mode, then reducing the channel pruning rate and retaining all feature extraction channels to perform identification; if the computing density parameter corresponds to a low-power sentinel instruction, indicating a low computing mode, then increasing the channel pruning rate and blocking a preset proportion of feature extraction channels to perform identification.

[0009] Preferably, the system also includes a spatiotemporal collaborative feedback module, used to update the monitoring strategy of the surrounding area based on the disease monitoring results. The process includes: determining whether the disease monitoring results output by the image acquisition execution module indicate the existence of a disease entity; if a disease entity exists, obtaining the geographical coordinates of the current monitoring area; retrieving the device IDs of adjacent monitoring areas within a preset distance range based on the geographical coordinates; sending a risk diffusion warning signal to the device IDs of adjacent monitoring areas; and the adaptive strategy generation module of the adjacent monitoring areas responding to the risk diffusion warning signal by forcibly upgrading its own disease risk level to the potential warning level or the high-risk outbreak level, and triggering the corresponding adaptive perception command.

[0010] Preferably, the adaptive strategy generation module processes network bandwidth parameters by: real-time monitoring of the current network transmission rate; if the network transmission rate is less than a preset bandwidth threshold, then when generating the adaptive perception instruction, activating the edge preprocessing flag; the image acquisition execution module responds to the edge preprocessing flag, completes image feature extraction locally, and only uploads the extracted feature vector data to the cloud management platform, intercepting the upload operation of the original image data.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This system constructs an adaptive sensing mechanism based on the dual constraints of environmental risk and edge-side resources, effectively solving the problems of excessive energy consumption and data redundancy caused by the traditional fixed-frequency acquisition mode. Unlike all-weather high-frequency acquisition, this solution dynamically reconstructs the acquisition strategy according to the probability of disease outbreak and the remaining power of the equipment. During periods of high risk, the sensing accuracy is automatically improved to prevent missed detections, while during safe periods or when resources are limited, it is downgraded to a low-power sentinel mode. Thus, under the limited power and computing power budget, the optimal balance between monitoring efficiency and equipment lifespan is achieved. 2. This system introduces a multi-source context fusion and saliency region clipping strategy, which significantly reduces the invalid computational load at the edge. By fusing meteorological data, texture features and spectral reflectance features to construct a digital sentinel mechanism, the system can capture environmental stress signals before the disease appears. Combined with the saliency detection algorithm, it performs fine-grained identification only on local image patches containing abnormal textures. Compared with full image processing, it significantly reduces the amount of floating-point operations and improves the edge inference efficiency while ensuring that minor disease features are not lost. 3. This system adopts a dynamic sparse computing architecture and a slim neural network training paradigm, realizing the flexible adaptation of deep learning models to different computing power scenarios. The system can dynamically adjust the channel pruning rate of convolutional layers according to real-time computing density instructions, and use switchable batch normalization technology to ensure the stability of feature distribution of the model under different widths. This allows the same set of model weights to be seamlessly switched between high-performance mode and power-saving mode without retraining, avoiding system lag or thermal runaway caused by computing power overload. 4. This system establishes an edge computing offloading and spatiotemporal collaborative defense network for weak network environments, enhancing the system's robustness in complex field environments. Addressing insufficient network bandwidth, the system automatically activates edge-end preprocessing to replace the original image upload with feature vectors, mitigating data congestion risks. Simultaneously, it utilizes a geofencing mechanism to achieve early warning of risk diffusion between neighboring nodes, enabling devices downstream of the propagation path to enter a state of readiness in advance, significantly improving the regional-level prevention and control response speed to sudden and migratory pests and diseases. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1: Please see Figure 1 The image recognition-based crop pest and disease monitoring system includes a cloud-based management platform. This platform is communicatively connected to an environmental perception module, a multi-source context analysis module, an adaptive strategy generation module, and an image acquisition and execution module. The environmental perception module is configured to deploy multimodal sensor nodes in the monitoring area to acquire environmental context parameters. The multi-source context analysis module is configured to acquire historical spatiotemporal data of disease distribution in the monitoring area, combine this data with environmental context parameters to construct a disease outbreak risk model, and calculate the disease risk level of the monitoring area within the current time window. The adaptive strategy generation module is configured to dynamically reconstruct the image acquisition mode and feature extraction strategy based on the disease risk level and edge-side resource constraints, generating adaptive perception commands. Edge-side resource constraints include at least remaining power and network bandwidth parameters. The image acquisition and execution module is configured to respond to the adaptive perception commands by adjusting the physical acquisition parameters of the visual sensor and the computational density of the image processing algorithm, acquiring and recognizing the target image, and outputting the disease monitoring results.

[0015] This embodiment details the physical architecture and logical interaction mechanism of the above system. The system deploys multimodal sensor nodes at the physical level through an environmental perception module. This node integrates a micro-meteorological instrument and a visual sensor. In this embodiment, the node integrates a low-power panoramic camera with zoom function, which aims to unify the terminology used in this document regarding image acquisition hardware. The visual sensor will be referred to as this device in the following text. The purpose is to establish a continuous perception capability of the environmental base of the monitoring area, rather than directly performing high-energy-consuming pest and disease identification. The multi-source context analysis module, as the cognitive center of the system, accesses historical spatiotemporal data of pest and disease distribution and real-time environmental context parameters, and uses machine learning algorithms to construct a pest and disease outbreak risk model. This model aims to quantify the probability of the current environment inducing pests and diseases. The adaptive strategy generation module dynamically generates adaptive sensing commands based on the calculated pest and disease risk level, which covers the risk of integrated outbreaks of diseases and pests, and coupled with the resource constraint status on the edge side, namely the remaining power parameters and network bandwidth parameters of the device. The image acquisition and execution module, as the end effector, responds to the command by physically adjusting the optical parameters of the sensor and the computational density of the algorithm to complete the closed-loop control from environmental perception to target recognition. This embodiment breaks away from the rigid mode of all-weather, high-frequency data collection in traditional monitoring systems by being driven by the dual constraints of environmental risk and resource status. It automatically improves the sensing accuracy during high-risk periods of pest and disease outbreaks, and automatically degrades to a low-power sentinel mode during periods of limited resources or low risk. Thus, it achieves the optimal balance between monitoring efficiency and equipment lifespan within the limited power and computing power budget of edge devices.

[0016] Example 2: The process by which the environmental perception module acquires environmental context parameters of the monitoring area includes: acquiring meteorological data from several key monitoring points within the monitoring area, including temperature, humidity, light intensity, and wind speed; activating a low-power panoramic camera to acquire a macroscopic panoramic image of the monitoring area at a preset low-frequency sampling period; extracting texture change feature vectors and spectral reflectance feature vectors from the macroscopic panoramic image; and concatenating the meteorological data, texture change feature vectors, and spectral reflectance feature vectors to generate environmental context parameters.

[0017] This embodiment further refines the logic for acquiring and fusing environmental parameters; regarding the consistency of terminology, the key monitoring outposts are the deployment examples of the multimodal sensor nodes in the key locations within the monitoring area in Embodiment 1, and both are physically the same component; the system locks the pre-marked key monitoring outposts within the monitoring area and collects meteorological data in real time through the micro-meteorological sensor array. The data covers four dimensions: temperature, humidity, light intensity, and wind speed, aiming to capture the meteorological factors that induce pest reproduction and disease spore germination. During this period, the system activates the dormant visual sensor, namely the low-power panoramic camera in this embodiment, and captures macroscopic panoramic images according to a preset low-frequency sampling period, such as every 30 minutes. Regarding the spectral reflectance feature vector extraction mentioned in this embodiment, this embodiment specifically adopts an algorithm based on pseudo-spectral reconstruction to compensate for the hardware deficiencies of ordinary cameras: a mapping matrix based on color space transformation is pre-established. The matrix was constructed as follows: Select samples of the same type of crop and simultaneously collect their RGB image data under natural light. Its matrix dimension is and The sample size is compared with the actual spectral reflectance data measured by the hyperspectral analyzer. Its matrix dimension is and To find the number of bands, the mapping matrix is ​​solved using a multiple linear regression algorithm. The specific regression equation is constructed as follows: in, Here is the error matrix; in order to solve... Construct the least squares loss function ,right Taking the derivative and setting it to zero, we obtain the formula for the analytical solution: like If the matrix is ​​singular, then a regularization term is introduced. , Value Ridge regression formula Calculations are performed to obtain the parameters of the mapping matrix; although linear mapping is difficult to fully restore complex hyperspectral details, it is sufficient here to extract the band trends that are sensitive to specific crop diseases for initial screening. During the monitoring phase, the system uses the Local Binary Pattern (LBP) operator to extract texture variation feature vectors reflecting crop canopy roughness or insect damage marks. The specific process is as follows: The image is converted to grayscale; for each pixel, its grayscale comparison value with its 8 neighboring pixels is calculated to generate an 8-bit binary code; the LBP code histogram of the entire image is calculated; and L1 normalization is performed on the histogram vector, that is, each dimension value is divided by the total number of pixels, so that the feature vector... The sum of all its components is 1, thus constraining its numerical range to 1. The interval was adjusted to eliminate the dimensional differences with meteorological data; and the RGB pixel average of the real-time acquired panoramic images was also adjusted. Input the above mapping matrix, and perform linear operations. The fitted spectral reflectance in a specific vegetation-sensitive band was calculated by inversion, and the results were normalized to... The interval is used to generate the spectral reflectance feature vector. ; The system performs multimodal data fusion operations, using the following feature concatenation and scaling formulas to generate the final environmental context parameters. : in, : Preprocessing normalization mapping function, given the raw meteorological data The values ​​of physical quantities such as light intensity and humidity in the data far exceed the effective linear range of the Sigmoid function. Directly substituting the values ​​will lead to output saturation and gradient vanishing. Therefore, this embodiment first... Perform Z-Score normalization, i.e., define ,in, The mean and standard deviation of the historical data for this monitoring area are then used to standardize the data. Substitution This effectively maps physical quantities with a wide dynamic range to... The sensitive interval; Feature scaling factor: In this specific implementation, in order to balance the contribution of different modal features, it is set as follows: ; : Vector concatenation operator, physically meaning the dimensionality expansion of multi-source features; This embodiment utilizes multimodal fusion technology to keenly capture the early spectral response of environmental stress or crop populations before the appearance of visible symptoms of pests and diseases, constructing a digital sentinel mechanism that can continuously monitor abnormal trends in large-scale farmland with extremely low computing power, providing advance data support for subsequent high-precision decision-making.

[0018] Example 3: The process by which the multi-source context analysis module calculates the disease risk level of the monitoring area in the current time window includes: inputting environmental context parameters into a preset disease outbreak risk model and outputting the disease occurrence probability value for the current time window; setting a high-risk threshold and a low-risk threshold, where the high-risk threshold value is greater than the low-risk threshold value; if the disease occurrence probability value is greater than or equal to the high-risk threshold, the disease risk level is marked as a high-risk outbreak level; if the disease occurrence probability value is less than the high-risk threshold but greater than the low-risk threshold, the disease risk level is marked as a potential early warning level; if the disease occurrence probability value is less than or equal to the low-risk threshold, the disease risk level is marked as a normal safe level.

[0019] This embodiment details the quantification and discretization process of risk levels; it constructs a time-series feature sequence and sets the sliding window length. ,For example Representing the past 48 sampling periods, including the current moment and historical data. Instantaneous environmental context parameters at each moment Combining into temporal input tensors The time-series input tensor is fed into a pest and disease outbreak risk model built on a Long Short-Term Memory (LSTM) network. This model is trained under supervised instruction using historical pest and disease data and outputs a normalized probability value for pest and disease occurrence. ; The system introduces a dual-threshold decision mechanism, setting a high-risk threshold. With low risk threshold ,satisfy These two thresholds aim to define the critical point of the system response; the specific threshold determination method is as follows: optimization is performed based on the receiver operating characteristic (ROC) curve of historical validation set data; high-risk threshold. The probability value is set to the specificity of 0.95 on the ROC curve, for example, 0.85 in the specific parameter settings of this embodiment, to ensure the accuracy of high-risk alarms and reduce the waste of resources caused by false alarms; low-risk threshold The value is set as the Youden index on the ROC curve, which is Sensitivity + Specificity - 1, and the probability value corresponding to the maximum point. For example, in the specific parameter settings of this embodiment, the value is 0.4 to balance the false negative rate and the false positive rate and cover the potential risk range. The system executes the hierarchical judgment logic: in response to The system determined that the environment was extremely conducive to the spread of pests and diseases, and marked it as a high-risk outbreak level; in response to The system determined that there was a potential outbreak risk and marked it as a potential warning level; in response to The system determines that the environment is safe and marks it as a normal safety level. This embodiment effectively filters out random noise caused by minor fluctuations in environmental parameters by introducing a dual-threshold, three-level judgment mechanism. It discretizes continuous probability values ​​into explicit operation instructions, avoiding frequent mode switching (i.e., jitter) in the system under critical conditions, and significantly improves the robustness and stability of the decision-making system.

[0020] Example 4: The process of generating adaptive perception instructions by the adaptive strategy generation module includes: obtaining the current remaining power parameters and network bandwidth parameters of the image acquisition execution module; if the disease risk level is high-risk outbreak level or potential early warning level, and the remaining power parameters are greater than the preset power threshold, then a high-precision perception instruction is generated, which instructs to increase the image acquisition resolution, shorten the acquisition time interval, and activate the deep neural network feature extraction channel; if the disease risk level is normal safety level, or the remaining power parameters are less than or equal to the preset power threshold, then a low-power sentinel instruction is generated, which instructs to reduce the image acquisition resolution, extend the acquisition time interval, and freeze the redundant feature extraction channels of the neural network.

[0021] This embodiment illustrates the resource-aware strategy generation logic; the system reads the remaining power parameters of edge devices in real time through the underlying driver interface. With network bandwidth parameters The system executes joint decision-making logic: responding to a disease risk level that is high-risk or in a warning state, and Greater than the preset power threshold In this embodiment, the specific setting is 20% of the total battery capacity. That is, when the battery level is lower than 20%, the protection logic is triggered. The system determines that the conditions for detailed investigation are met and generates a high-precision perception command. This command forcibly increases the acquisition resolution, such as to 4K, shortens the acquisition interval, and fully activates the DNN model channels. It should be noted here that the network bandwidth parameter... Potential shortcomings, namely weak network conditions, are not directly addressed as triggers for downgrading to low-power sentinel mode in this step. Instead, the aim is to ensure that the front end acquires high-resolution raw images. Specific transmission optimization is handled by the edge preprocessing mechanism of the implementation, essentially trading computing power for bandwidth to avoid high-risk missed detections due to poor network conditions. Conversely, in response to a risk level of normal security, or even if the risk is high but... If there is a critical power shortage, the system determines that the survival of the device should be prioritized and generates a low-power sentinel command. This command forcibly reduces the resolution, such as to 480p, extends the acquisition interval, and performs model pruning. This embodiment implements a resource-aware dynamic control strategy. During periods of high disease incidence, concentrated firepower ensures no missed detections, while during safe periods or periods of low battery, it automatically enters a dormant or reduced-dimensional operation state. This non-linear resource scheduling mechanism solves the problem of premature power outages caused by continuous high-load operation of field equipment, greatly extending the effective working time of unattended operation. Field tests have verified that, under the same hardware conditions with a 20000mAh battery pack, the equipment using the adaptive strategy of this embodiment has an average continuous working time of 42 days, while the control group equipment using the traditional fixed-frequency, 30-minute data collection mode has an average working time of only 8.5 days. This embodiment extends the equipment lifespan by approximately 394%, effectively reducing the frequency of manual maintenance.

[0022] Example 5: The image acquisition and execution module responds to high-precision perception commands to perform the process of acquiring and recognizing target images, including: controlling the vision sensor to acquire original image frames of the monitoring area at high resolution; using a lightweight saliency detection algorithm to scan the original image frames and locate regions of interest with abnormal textures or color changes; cropping local image blocks according to the coordinates of the regions of interest; inputting the local image blocks into a preset refined disease recognition model and outputting the disease category and confidence level of the local image blocks.

[0023] This embodiment details the coarse-to-fine processing flow in high-precision mode. In response to the high-precision perception command, the visual sensor, namely the aforementioned low-power panoramic camera, is locked in high-resolution mode to acquire original image frames to preserve the optical details of tiny lesions. A lightweight saliency detection algorithm with extremely low computational overhead, such as SpectralResidual, is used to quickly scan the high-resolution image. This algorithm aims to locate the Region of Interest (ROI) containing abnormal textures or color abrupt changes, thereby filtering out invalid backgrounds such as the sky and soil. The system crops local image patches from the original frame based on the ROI coordinates. Only these local image patches containing potential disease features are input into a computationally intensive fine-grained disease identification model, which outputs specific disease categories and confidence levels. This embodiment cleverly resolves the contradiction between high-resolution image processing and edge computing bottlenecks by cascading a saliency detection and local cropping mechanism. Compared to full-image convolution, this scheme reduces invalid computation by more than 70%, achieving efficient edge inference while ensuring that minor defects are not lost. Specific performance comparison tests show that for an input image with a resolution of 3840×2160, if convolution is performed directly without the cropping mechanism of this embodiment, the number of floating-point operations per frame (FLOPs) reaches as high as 120G, taking about 1.5 seconds. However, with the scheme of this embodiment, only about 18% of the image area needs to be processed on average, reducing the number of operations per frame to about 25G, a reduction of 79.1%, and shortening the processing time to 0.35 seconds, fully demonstrating the significant effect of this scheme in reducing invalid computation.

[0024] Example 6: The refined disease identification model adopts a dynamic sparse computing architecture. Its execution process includes: obtaining the computing density parameter carried in the adaptive perception instruction; dynamically adjusting the channel pruning rate of the convolutional layer in the refined disease identification model according to the computing density parameter; if the computing density parameter corresponds to a high-precision perception instruction, indicating a high computing mode, then the channel pruning rate is reduced, and all feature extraction channels are retained to perform identification; if the computing density parameter corresponds to a low-power sentinel instruction, indicating a low computing mode, then the channel pruning rate is increased, and a preset proportion of feature extraction channels are blocked to perform identification.

[0025] This embodiment describes the model-level elastic scaling mechanism; the recognition engine parses the computational density parameters carried in the adaptive perception instructions. This parameter directly maps to the computing power requirement of the current task; the system based on Dynamically adjust the channel pruning rate of the convolutional neural network The two are negatively correlated; specifically, in response to Indicates high computing mode, the system will Set to a value close to zero to retain all feature extraction channels to maximize feature representation; in response to Indicating low computation mode, the system significantly improves If set to 0.5, the feature extraction channels of the preset ratio will be blocked, that is, according to the preset channel sorting rules, the channels at the end of the index will be truncated, and only the continuous channels at the beginning of the index will be retained. To prevent tensor dimension mismatch caused by changes in the number of channels, this embodiment specifically employs dynamic weighted slicing technology: the system does not need to maintain complex sparse masks, but directly performs slicing operations based on the width ratio coefficient; specifically, let the first... Layer weight shape is If the current width multiplier is For example, 0.5, then the first The weighted subtensors that the layer actually participates in the computation are: This structured dynamic reorganization mechanism ensures that the tensor shapes of the data streams between layers are always aligned. To ensure that the model maintains high recognition accuracy after direct slicing and avoids feature distribution drift, this embodiment adopts SlimmableNetworks in the model construction stage, which can slim down the neural network training paradigm. Specifically, the model integrates SwitchableBatchNormalization, which allows switching batch normalization S-BN layers. This layer independently maintains specific mean and variance statistics for different channel pruning rates. During training, the system randomly samples subnetworks of different widths for joint optimization, forcing the convolutional filters to learn robustness to different input widths. Therefore, when performing dynamic weight slicing during the inference phase, the system synchronously switches to the corresponding S-BN statistical parameters, thereby ensuring that high-performance recognition can be maintained without retraining. In addition, in order to substantially reduce FLOPs, the underlying inference engine performs operations based on the compact weight matrix after physical slicing. Since dynamic weight slicing has physically reorganized the large matrix into a denser matrix with smaller dimensions, the system does not need to rely on the sparse matrix multiplication instruction set SparseGEMM to process zero elements. Instead, it can directly use the standard GEMM general matrix multiplication instruction set to efficiently process the dimensionality-reduced data stream, thereby eliminating the additional indexing overhead caused by sparse computation, ensuring the consistency between computational acceleration and tensor shape alignment logic, and transforming the theoretical reduction in computational load into actual inference acceleration and energy consumption reduction. This embodiment endows deep learning models with dynamic adaptability, enabling the same set of model weight files to adapt to computing power constraints under different power conditions, achieving seamless switching from high-performance mode to power-saving mode, and avoiding system lag or thermal runaway caused by computing power overload.

[0026] Example 7: The system also includes a spatiotemporal collaborative feedback module, which updates the monitoring strategy of the surrounding area based on the disease monitoring results. The process includes: determining whether the disease monitoring results output by the image acquisition execution module indicate the existence of a disease entity; if a disease entity exists, obtaining the geographical coordinates of the current monitoring area; retrieving the device IDs of adjacent monitoring areas within a preset distance range based on the geographical coordinates; sending a risk diffusion warning signal to the device IDs of adjacent monitoring areas; and the adaptive strategy generation module of adjacent monitoring areas responding to the risk diffusion warning signal by forcibly upgrading its own disease risk level to the potential warning level or the high-risk outbreak level, and triggering the corresponding adaptive perception command.

[0027] This embodiment constructs a joint prevention and control mechanism based on geofencing; the system performs semantic parsing on the output results to determine whether a definite disease entity has been detected; in response to a positive detection result, the system immediately reads the geographical coordinates of the current node. Using these coordinates as the center, a preset distance range is retrieved from the cloud topology map. The system collects the device IDs of all adjacent monitoring areas within the system; broadcasts risk spread warning signals to these neighboring nodes; upon receiving the signal, the neighboring nodes, through their interruption mechanism, forcibly ignore the locally calculated low-risk level, upgrade their status to the potential warning level or high-risk outbreak level, and immediately trigger high-frequency data collection commands. This embodiment utilizes the principle of spatial proximity in the spread of pests and diseases to establish a collaborative defense network where one party detects the disease and the surrounding areas issue early warnings. This allows monitoring nodes downstream of the transmission path to enter a state of readiness in advance, greatly improving the regional control response speed to sudden and migratory pests.

[0028] Example 8: The adaptive strategy generation module processes network bandwidth parameters by: real-time monitoring of the current network transmission rate; if the network transmission rate is less than the preset bandwidth threshold, the edge preprocessing flag is activated when generating the adaptive perception command; the image acquisition execution module responds to the edge preprocessing flag, completes image feature extraction locally, and only uploads the extracted feature vector data to the cloud management platform, intercepting the upload operation of the original image data.

[0029] This embodiment details the data transmission optimization strategy in a weak network environment; the communication module monitors the current network transmission rate in real time. The system executes bandwidth adaptive logic in response to... Less than the preset bandwidth threshold The system indicates network congestion or weak signal. The strategy generation module sets the edge preprocessing flag in the instruction. In response to the flag, the image acquisition execution module cuts off the big data stream upload channel of the original image and performs feature extraction on the local embedded chip instead, uploading only the compressed feature vector data to the cloud database. This embodiment adopts a strategy of trading computing power for bandwidth. In the farmland environment with poor network conditions, edge computing significantly reduces the amount of data transmission, ensures the continuity and integrity of key monitoring data, effectively avoids the risk of data packet loss due to network timeout, and reduces the concurrent I / O pressure of the cloud server.

[0030] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A crop disease and pest monitoring system based on image recognition, characterized in that, It includes a cloud-based management and control platform, which is connected to an environmental perception module, a multi-source context analysis module, an adaptive strategy generation module, and an image acquisition and execution module. The environmental perception module is configured to deploy multi-modal sensor nodes in the monitoring area and obtain environmental context parameters of the monitoring area through the multi-modal sensor nodes. The multi-source context analysis module is configured to acquire historical spatiotemporal data of disease distribution in the monitoring area, combine environmental context parameters, construct a disease outbreak risk model, and calculate the disease risk level of the monitoring area in the current time window. The adaptive strategy generation module is configured to dynamically reconstruct the image acquisition mode and feature extraction strategy based on the disease risk level and the edge-side resource constraint status, and generate adaptive perception instructions. The edge-side resource constraint status includes at least the remaining power parameters and network bandwidth parameters. The image acquisition execution module is configured to respond to the adaptive perception instructions, adjust the physical acquisition parameters of the visual sensor and the computational density of the image processing algorithm, perform target image acquisition and recognition, and output disease monitoring results.

2. The crop disease and pest monitoring system based on image recognition according to claim 1, characterized in that, The process by which the environmental perception module acquires environmental context parameters of the monitoring area includes: acquiring meteorological data from several key monitoring outposts within the monitoring area, including temperature, humidity, light intensity, and wind speed; activating a low-power panoramic camera to acquire a macroscopic panoramic image of the monitoring area at a preset low-frequency sampling period; extracting texture change feature vectors and spectral reflectance feature vectors from the macroscopic panoramic image; and concatenating the meteorological data, texture change feature vectors, and spectral reflectance feature vectors to generate environmental context parameters.

3. The crop disease and pest monitoring system based on image recognition according to claim 2, characterized in that, The process by which the multi-source context analysis module calculates the disease risk level of the monitoring area in the current time window includes: inputting environmental context parameters into a preset disease outbreak risk model and outputting the disease occurrence probability value in the current time window; setting a high-risk threshold and a low-risk threshold, wherein the high-risk threshold value is greater than the low-risk threshold value; if the disease occurrence probability value is greater than or equal to the high-risk threshold, the disease risk level is marked as a high-risk outbreak level; if the disease occurrence probability value is less than the high-risk threshold but greater than the low-risk threshold, the disease risk level is marked as a potential early warning level; if the disease occurrence probability value is less than or equal to the low-risk threshold, the disease risk level is marked as a normal safe level.

4. The crop disease and pest monitoring system based on image recognition according to claim 3, characterized in that, The process of generating adaptive perception instructions by the adaptive strategy generation module includes: obtaining the current remaining power parameters and network bandwidth parameters of the image acquisition execution module; if the disease risk level is high-risk outbreak level or potential early warning level, and the remaining power parameters are greater than the preset power threshold, then a high-precision perception instruction is generated, which instructs to increase the image acquisition resolution, shorten the acquisition time interval, and activate the deep neural network feature extraction channel; if the disease risk level is normal safety level, or the remaining power parameters are less than or equal to the preset power threshold, then a low-power sentinel instruction is generated, which instructs to reduce the image acquisition resolution, extend the acquisition time interval, and freeze the redundant feature extraction channels of the neural network.

5. The crop disease and pest monitoring system based on image recognition according to claim 4, characterized in that, The image acquisition and execution module responds to high-precision perception commands to perform the process of acquiring and recognizing target images, including: controlling the vision sensor to acquire original image frames of the monitoring area at high resolution; using a lightweight saliency detection algorithm to scan the original image frames and locate regions of interest with abnormal textures or color changes; cropping local image blocks according to the coordinates of the regions of interest; inputting the local image blocks into a preset refined disease recognition model and outputting the disease category and confidence level of the local image blocks.

6. The crop disease and pest monitoring system based on image recognition according to claim 5, characterized in that, The refined disease identification model adopts a dynamic sparse computing architecture. Its execution process includes: obtaining the computing density parameter carried in the adaptive perception instruction; dynamically adjusting the channel pruning rate of the convolutional layer in the refined disease identification model according to the computing density parameter; if the computing density parameter corresponds to a high-precision perception instruction, indicating a high computing mode, then reducing the channel pruning rate and retaining all feature extraction channels to perform identification; if the computing density parameter corresponds to a low-power sentinel instruction, indicating a low computing mode, then increasing the channel pruning rate and blocking a preset proportion of feature extraction channels to perform identification.

7. The crop disease and pest monitoring system based on image recognition according to claim 6, characterized in that, The system also includes a spatiotemporal collaborative feedback module, which updates the monitoring strategy of the surrounding area based on the disease monitoring results. The process includes: determining whether the disease monitoring results output by the image acquisition execution module indicate the existence of a disease entity; if a disease entity exists, obtaining the geographical coordinates of the current monitoring area; retrieving the device IDs of adjacent monitoring areas within a preset distance range based on the geographical coordinates; sending a risk diffusion warning signal to the device IDs of adjacent monitoring areas; and the adaptive strategy generation module of adjacent monitoring areas responding to the risk diffusion warning signal by forcibly upgrading its own disease risk level to the potential warning level or the high-risk outbreak level, and triggering the corresponding adaptive perception command.

8. The crop disease and pest monitoring system based on image recognition according to claim 4, characterized in that, The adaptive strategy generation module processes network bandwidth parameters by: real-time monitoring of the current network transmission rate; if the network transmission rate is less than the preset bandwidth threshold, the edge preprocessing flag is activated when generating the adaptive perception command; the image acquisition execution module responds to the edge preprocessing flag, completes image feature extraction locally, and only uploads the extracted feature vector data to the cloud management platform, intercepting the upload operation of the original image data.

Images