Passion fruit disease and insect pest early warning method fused with multispectral imaging

Abstract

The invention discloses a passion fruit disease and insect pest early warning method fused with multispectral imaging, and the method comprises the steps: deploying multispectral imaging equipment and an Internet of Things environment monitoring station in a planting region, and transmitting environment data to an edge calculation terminal in real time through a LoRa protocol; collecting canopy multispectral images at intervals, and calculating the reflectivity of each wave band after correction; an edge computing terminal is used for preprocessing the image, and a disease sensitivity index DSI and a pest sensitivity index NDRE are calculated; dSI, NDRE, 426nm waveband reflectivity, temperature and humidity are combined into a multi-dimensional feature vector, the multi-dimensional feature vector is input into a two-channel convolutional neural network model of a cloud server, the model comprises independent disease recognition branches and pest recognition branches, and DSI and temperature, NDRE, 426nm and humidity data are processed respectively; and when the disease probability is higher than 65% or the pest probability is higher than 60%, an early warning instruction is automatically triggered and pushed to a farmer terminal. The method is used for early recognition and precise control of diseases and pests in the passion fruit planting area.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of agricultural information technology, and particularly relates to a guava disease and pest early warning method fusing multispectral imaging. BACKGROUND

[0002] During the planting process of guava, various diseases and pests such as anthracnose, blight, fruit fly pests and the like are faced, and these diseases and pests often have no significant visible symptom characteristics in the early stage, so that when discovered, they have entered the damage stage. In particular, in the incubation period stage, the pathogenic bacteria or pests have invaded the plant tissue, but the plant outside has not yet shown typical disease spots or worm-eaten traces, so it is difficult to achieve effective identification by relying on manual field inspection. Planters usually need to take prevention and control measures after the leaves appear obvious yellowing, wilting or the fruit surface forms disease spots, and at this time, the diseases and pests have caused substantial damage to the physiological function of the plant, and the prevention and control effect is significantly reduced.

[0003] The traditional monitoring method mainly relies on single-dimensional data collection, such as only analyzing the pathogenic bacteria quantity through soil sampling or only predicting the disease epidemic trend according to meteorological data. This kind of method cannot capture the relevance of crop physiological state and environmental factors at the same time. For example, the occurrence of guava anthracnose is not only related to the density of pathogenic bacteria, but also significantly affected by temperature and humidity changes; the oviposition behavior of fruit flies is coupled with light, temperature and humidity. However, the existing technology lacks the ability of synchronous and collaborative analysis of multispectral information (such as leaf reflectance change) and environmental parameters (such as temperature and humidity fluctuation), resulting in single input features of early warning model, which is difficult to accurately reflect the complex conditions of disease and pest occurrence.

[0004] Due to the above technical limitations, there are two difficulties in actual production: first, early identification relies on laboratory detection (such as PCR technology), which needs to culture or detect samples in vitro, and takes a long time of 24-36 hours, and the representative sampling position is easy to cause missed detection; second, the data of environmental parameters and crop physiological state are separated, so that the prediction model cannot establish key criteria such as "quantitative correlation between leaf spectrum red edge shift and anthracnose outbreak under continuous high temperature and high humidity environment". This data fragmentation problem causes the early warning result to lag behind the actual development process of diseases and pests, and farmers are often forced to increase the amount of pesticides due to the missed best prevention and control window period (such as within 24 hours before the end of the incubation period), which not only increases the cost but also affects the ecological safety.

[0005] In addition, the complex field environment further increases the difficulty of early monitoring. The guava canopy structure is dense, and the traditional visible light imaging is easily disturbed by branch and leaf shielding and light changes, making it difficult to capture the initial disease spots or eggs on the back of the leaves. Although the handheld spectrometer can obtain the local reflectivity of the leaves, the point sampling cannot cover the entire planting area, and the operation process easily damages the plants. These technical defects make it difficult for existing systems to meet the needs of large-scale, non-invasive and continuous monitoring, restricting the practical application value of early warning. SUMMARY

[0006] An object of the present application is to solve at least the above problems and provide at least the advantages to be explained later.

[0007] The present application at least solves the following technical problems:

[0008] 1. Solve the problem that the early stage of guava disease and insect pests is difficult to identify by naked eye, leading to delayed prevention and control, and the defect that the traditional monitoring method cannot synchronously fuse multispectral images (such as reflectivity changes) and environmental parameters (temperature and humidity) for early warning. The existing technology relies on single-dimensional data (such as only meteorological or soil analysis), which breaks the correlation between crop physiological state and environmental factors, cannot establish quantitative criteria such as "red edge shift of spectrum under high temperature and high humidity and outbreak of anthracnose", and causes the early warning to lag behind the actual development process of diseases and insect pests.

[0009] 2. Solve the problem of unstable image quality and incomplete canopy coverage caused by inaccurate deployment of multispectral imaging equipment. Traditional fixed height or angle equipment is easily shielded by dense canopy structure, cannot adaptively capture disease spots or eggs on the back of the leaves, and the deviation of waveband range will reduce the reliability of reflectivity data, affecting the accuracy of subsequent feature extraction.

[0010] 3. Solve the problem of decline in effectiveness of feature vectors caused by fuzzy data collection conditions for environmental monitoring. The traditional method does not limit the applicable range of temperature and humidity, and when the data exceeds the sensitive interval of diseases and insect pests (such as low temperature and low humidity), redundant information will interfere with the model input, and the correlation between environmental fluctuations and spectral changes (such as the coupling effect of humidity on the oviposition behavior of fruit flies) cannot be accurately related.

[0011] 4. Solve the problem of decline in generalization ability caused by environmental differences (such as light and soil in South China and the target area) when the model is deployed across regions, and the problem of baseline drift of healthy leaf spectrum caused by seasonal change (rainy season and dry season). The static template library cannot adapt to regional or seasonal changes, causing feature discrimination errors and reducing the specificity of early warning.

[0012] 5. Solve the problem of resource overload and storage pressure surge of edge computing terminal caused by dynamic feature library update strategy. The traditional update mechanism (such as full-time SAD calculation) occupies a large amount of computing power, squeezing real-time monitoring tasks, and the uncompressed spectral data causes storage space shortage, affecting the continuous operation stability of the system.

[0013] 6. Solve the problem of pre-screening module accuracy decay caused by the fixation of health benchmark range. Due to the natural change of leaf spectrum with growth cycle, the fixed benchmark cannot dynamically reflect the real fluctuation of health status, which is easy to misjudge (such as identifying normal change as abnormal) and interfere with the efficiency of subsequent template matching.

[0014] 7. Solve the problem of label timeliness deviation caused by PCR detection delay (ΔT) in training data construction. The traditional method directly uses the timeout sample label, ignoring the nonlinear change of pathogen proliferation over time (such as anthracnose reproduction within ΔT), so that the model learns the outdated or distorted disease and pest status related features.

[0015] 8. Solve the problem of nonlinear deviation of PCR detection results and actual collection time disease and pest status caused by sample storage environment temperature and humidity fluctuation. Linear decay model cannot reflect the dynamic proliferation law of microorganisms (such as high temperature accelerating pathogen activity), resulting in distorted label correction and reduced model training accuracy.

[0016] To achieve these objects and other advantages and in accordance with the purpose of the application, a guava disease and pest early warning method is provided, which comprises the following steps:

[0017] S1: Deploying a multi-spectral imaging device and an Internet of Things environment monitoring station in the guava planting area, the environment monitoring station transmits data to the edge computing terminal in real time through the LoRa wireless transmission protocol;

[0018] S2: Intervals of guava plant canopy multi-spectral images are collected by the multi-spectral imaging device, and the original image is corrected to eliminate environmental light interference by diffuse reflection white board correction, and the correction formula is: I original is the original image, I dark is the dark current calibration image, I white is the total reflection calibration image, I corrected is the corrected image; and the reflectivity of each waveband is calculated

[0019] S3: The edge computing terminal is used to preprocess the corrected multi-spectral image, and the preprocessing includes: extracting a rectangular region of interest of 200 pixels x 200 pixels on both sides of the main vein of the leaf; calculating the reflectivity of each waveband, and generating a disease sensitive index DSI and a pest sensitive index NDRE, RE is the red edge band reflectivity, NIR is the near-infrared band reflectivity, and RED is the red light band reflectivity;

[0020] S4: combine DSI, NDRE, 426 nm band reflectance, 840 nm reflectance, temperature, humidity data into a multi-dimensional feature vector to construct training data, the training data is input into a dual-channel convolutional neural network model deployed on a cloud server, the dual-channel convolutional neural network includes independent disease identification branch and pest identification branch, wherein the disease branch takes DSI, 840 nm reflectance and temperature as input; the pest branch takes NDRE, 426 nm reflectance and humidity as input, the outputs of the two branches are fused through a fully connected layer;

[0021] S5: The early warning threshold is determined based on the model validation set: when the model output potential period disease probability is higher than 65% or pest infestation probability is higher than 60%, an early warning instruction is automatically triggered, which includes the type of disease and insect pest, the coordinate position of occurrence and the matched biological control agent type; the early warning instruction is pushed to the farmer terminal through 4G / 5G mobile communication network to perform precise pesticide application operation.

[0022] Preferably, the multispectral imaging device is fixed on a height of 1.8m to 2.2m from the crop canopy by an adjustable support, and an automatic rotating pan-tilt is configured to realize full-coverage of the canopy, the multispectral imaging device includes imaging channels of blue band 450nm±16nm, green band 560nm±16nm, red band 650nm±16nm, red edge band 730nm±16nm, near-infrared band 840nm±26nm and pest-sensitive band 426nm±10nm, and the multispectral imaging device collects multispectral images of the guava plant canopy at a fixed time interval during 8am to 10am every day.

[0023] Preferably, the environmental monitoring station collects air temperature and relative humidity; wherein, when the temperature is 35-40℃ and the humidity is 75%-95%, it is used for the generation of feature vector.

[0024] Preferably, in the training stage of the dual-channel convolutional neural network model, a transfer learning mechanism and a dynamic feature library updating strategy are introduced, which includes the following steps:

[0025] 1) Pre-train the model using the source region labeled guava disease and pest multispectral data set, the source region is the historical data of the planting base in South China; migrate the convolutional layer weights of the pre-trained model to the target region model, and fine-tune on the small sample data set composed of the potential period leaf multispectral feature vectors collected in the target region and the disease-free leaf spectra confirmed based on PCR detection, the sample amount of the small sample data set is not less than 200 groups;

[0026] 2) The model fine-tuned in step 1) is deployed in the target area, and a seasonal healthy leaf spectrum template library is established in the edge computing terminal, including a rainy season template group and a dry season template group; the rainy season template group corresponds to the healthy leaf spectrum curve when the air relative humidity is greater than or equal to 85%, and the dry season template group corresponds to the healthy leaf spectrum curve when the air relative humidity is less than or equal to 75%; the multi-spectral imaging device automatically matches and calibrates the template group according to the real-time collected environmental humidity data;

[0027] 3) At least 10 healthy leaves determined by the model are selected by the edge computing terminal every day, and the average spectrum thereof is calculated as a reference spectrum; the seasonal template group selected in step 2) is used to calculate the spectral angle distance SAD between the current healthy leaf reference spectrum and the matching template every day, and the calculation formula is:

[0028] wherein is the standard reflectance of the matching seasonal template group at the wave band i, R i is the reflectance of the reference spectrum at the wave band i, and n is the number of wave bands; when the SAD value is less than 0.15 radian, the current reference spectrum curve is added to the corresponding seasonal template library;

[0029] 4) The seasonal template group updated in step 3) is applied to multi-spectral image correction, and the corrected feature data is input into the fine-tuned model in step 1); a domain adaptation layer is added before the full connection layer in the model fine-tuning training stage, which uses the maximum mean difference algorithm to calculate the feature distribution difference between the source domain and the target domain, and adds the weighted loss function after weighting with a weight coefficient β, β ∈ [0.3, 0.5]; only the template matching mechanism is enabled in the deployment reasoning stage.

[0030] Preferably, an edge computing resource collaborative management mechanism is integrated in the dynamic feature library update strategy, including:

[0031] a) A lightweight pre-screening module is deployed in the edge computing terminal to monitor the absolute value of the reflectance difference of healthy leaves at the pest-sensitive wave band 426 nm and the disease-sensitive wave band 840 nm in real time; when the reflectance difference at the 426 nm or 840 nm wave band exceeds the healthy reference range ± 15%, the full-waveband spectral angle distance SAD calculation process is triggered; otherwise, the SAD calculation is skipped and the current seasonal template group is directly used;

[0032] b) The SAD calculation and template update operation in step 3) are limited to be performed during the period from 00:00 to 05:00 every day; during this period, the convolution layer operation of the disease identification branch is suspended, only the 426 nm wave band monitoring function of the pest identification branch is maintained, and 80% of the computing resources are allocated to the SAD process by the task scheduler of the edge computing terminal;

[0033] c) Establishing a spectrum compression engine on the cloud server side, performing PCA dimension reduction on the original spectrum data of the seasonal template group, and compressing the original six-dimensional spectrum data to a three-dimensional feature vector; the compressed feature vector is returned to the edge computing terminal through the 4G / 5G network;

[0034] d) The edge computing terminal establishes a three-version circular queue storage area for each seasonal template group, and only retains the latest three versions of compressed spectrum data; when the number of versions exceeds three due to the addition of new templates, the oldest version of data in the queue is automatically deleted and the storage space is released.

[0035] Preferably, a health benchmark range dynamic updating mechanism is added in step a), which specifically includes the following operations:

[0036] a-1) Establishing a health benchmark range updating module on the edge computing terminal, which extracts the reflectance data of the 426nm and 840nm bands from the current matched seasonal template group, and calculates the health benchmark range of each band:

[0037] Formula for the minimum value of the health benchmark range of the 426nm band:

[0038] Formula for the maximum value of the health benchmark range of the 426nm band:

[0039] Formula for the minimum value of the health benchmark range of the 840nm band:

[0040] Formula for the maximum value of the health benchmark range of the 840nm band:

[0041] represents the reflectance value of the i-th sample in the current seasonal template group at the 426nm band;

[0042] represents the reflectance value of the j-th sample in the current seasonal template group at the 840nm band;

[0043] min is the minimum value function; max is the maximum value function;

[0044] a-2) The health benchmark range used in step a) is updated every 24 hours, and the update time is set to the period from 00:00 to 05:00 every day, and is executed after the SAD calculation and template updating operation in step b) is completed;

[0045] a-3) When the seasonal template group is updated in step 3), the health benchmark range is immediately triggered to recalculate and overwrite the old value.

[0046] Preferably, the training data construction includes the following steps:

[0047] b-1) Collecting the multispectral image of the guava plant canopy in step S2, at the same time, marking the leaf of the same plant for puncture sampling, the puncture position is located at the geometric center point of the 200 pixel x 200 pixel rectangular area on both sides of the main vein of the leaf ± 5 cm;

[0048] b-2) Establish a unified time axis for each sample, where the multispectral image collection time is recorded as T0, and the puncture leaf sample is detected in the PCR laboratory at time T1; Limit the time delay ΔT = T1-T0≤36 hours, and the overtime sample is automatically rejected;

[0049] b-3) The disease sensitive index DSI and the insect pest sensitive index NDRE generated in step S3 are bound with the PCR detection results according to the following rules:

[0050] When ΔT≤24 hours: directly use the original PCR label, and the anthracnose Ct value ≤32 is marked as 1, and the anthracnose Ct value >32 is marked as 0; Fruit fly egg density ≥5 eggs / cm 2 marked as 1, and the fruit fly egg density <5 eggs / cm 2 marked as 0;

[0051] When 24 hours < ΔT≤36 hours: apply an attenuation coefficient k = 1-0.02×(ΔT-24) to the PCR label, which is based on the negative correlation between the timeliness of PCR detection and the proliferation rate of the pathogen;

[0052] For samples with ΔT>24 hours, if the PCR detection is negative, it is directly marked as 0; only for positive samples, the attenuation coefficient k is applied;

[0053] b-4) Store the bound data set as: {[DSI, NDRE, temperature, humidity, ΔT}, [disease label x k, pest label x k]}, complete the training data construction.

[0054] Preferably, in step b-3), the generation of the attenuation coefficient k further introduces a PCR correction factor λ, which specifically includes:

[0055] c-1) During the period from puncture sampling to the completion of PCR detection, the temperature data of the sample storage environment is continuously recorded by the Internet of Things environment monitoring station, and the average temperature T avg during this period is calculated;

[0056] c-2) According to ΔT and T avg , calculate the PCR correction factor λ:

[0057] λ = e -0.03×ΔT×[1+0.05×(Tavg-25)] , λ ∈ [0.2, 1.0];

[0058] c-3) When T avgWhen the temperature is 28℃, the attenuation coefficient k is updated as: k = max(0, 1-0.02*(DeltaT-24)*lambda); otherwise, the linear attenuation coefficient is used;

[0059] d) a two-dimensional look-up table of DeltaT and T avg is established, and when DeltaT > 24 hours, the discrete modified value of lambda is obtained through interpolation mapping.

[0060] The present application at least includes the following beneficial effects:

[0061] By fusing multispectral imaging and environmental parameters, a multi-dimensional feature vector is constructed to input a double-channel convolutional neural network, thereby realizing early and synchronous warning of diseases and pests. The independent processing mechanism of the disease identification branch and the pest identification branch is used to respectively associate key factors such as DSI index and temperature, NDRE index and humidity, thereby improving the sensitivity to latent diseases and pests. By limiting the device height range (1.8-2.2m) and the waveband parameters (such as 426nm±10nm pest-sensitive waveband), it is ensured that the multispectral image covers the complete canopy and reduces light interference. The automatic rotation of the gimbal enhances the spatial sampling capability, covers the back area of the leaf, and improves the representativeness and consistency of the reflectance data, thereby providing a reliable basis for feature extraction.

[0062] By constraining the temperature and humidity suitable range (temperature 35-40℃, humidity 75%-95%), the high environmental conditions of diseases and pests are focused, and invalid data is eliminated. This enhances the discriminability of the feature vector, makes the model more accurately capture the cooperative change rule of the environment and the spectrum (such as the sensitive reflectance characteristics of anthracnose under high temperature and high humidity), and reduces the false positive rate.

[0063] The transfer learning mechanism is combined with dynamic template updating, the healthy leaf spectrum library is established seasonally (rainy season / dry season group), and the template is automatically adjusted through SAD calculation. The domain adaptation layer reduces the difference between the source domain and the target domain in the fine-tuning stage, thereby improving the cross-region adaptability of the model. After deployment, the template matching mechanism suppresses the baseline drift, thereby maintaining the long-term discriminant stability.

[0064] The light-weight pre-screening module filters invalid calculations through the reflectance difference threshold, the resource scheduler allocates computing power to the SAD process, and the PCA dimensionality reduction is combined to compress the spectral data. The circular queue storage mechanism optimizes the space utilization, reduces the load of the edge terminal, and ensures the efficient parallelism of real-time monitoring and template updating.

[0065] The healthy benchmark range dynamic updating module calculates the extreme range (such as the reflectance of the 426nm waveband) based on the seasonal template group in real time, and automatically refreshes every 24 hours. The real-time trigger updating mechanism responds to the template changes, so that the pre-screening module accurately tracks the healthy leaf spectrum

[0066] Puncture sampling and spatiotemporal alignment of multispectral images (ΔT≤36 hours), combined with decay coefficient correction of sample labels beyond time, compensate for pathogen proliferation effects. Rules distinguish between negative / positive sample processing, enhance label reliability, and improve the model's ability to capture the timing of disease and pest development.

[0067] PCR correction factor λ introduces non-linear effects of temperature and humidity, optimizing the decay coefficient. Two-dimensional lookup table maps ΔT and T avg Interpolation makes label correction more consistent with the actual proliferation of microorganisms, improving the authenticity of training data.

[0068] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following detailed description, and it is intended to cover any and all adaptations of the present application. DETAILED DESCRIPTION

[0069] The present application will be further described in detail below, so that those skilled in the art can implement it according to the description.

[0070] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0071] It should be noted that the experimental methods described in the following embodiments are all conventional methods, and the reagents and materials, unless otherwise specified, can be obtained commercially.

[0072] A fusion multi-spectral imaging early warning method for guava diseases and pests, the method comprising the following steps:

[0073] S1: Deploying multi-spectral imaging equipment and Internet of Things environment monitoring stations in the guava planting area, the environment monitoring stations transmit data in real time to the edge computing terminal through the LoRa wireless transmission protocol;

[0074] S2: Intervals of multi-spectral imaging equipment are used to collect multi-spectral images of guava plant canopy, and the original image is corrected by diffuse reflection white board to eliminate environmental light interference, and the correction formula is: I original is the original image, I dark is the dark current calibration image, I white is the total reflection calibration image, I corrected corrected image; and calculate the reflectance of each waveband

[0075] S3: The edge computing terminal is used to preprocess the corrected multi-spectral image, the preprocessing includes: extracting a rectangular region of interest of 200 pixels x 200 pixels on both sides of the leaf vein; calculating the reflectance of each waveband, and fusing to generate disease sensitive index DSI and pest sensitive index NDRE, RE is red edge band reflectance, NIR is near-infrared band reflectance, and RED is red light band reflectance;

[0076] S4: The DSI, NDRE, 426 nm band reflectance, 840 nm reflectance, temperature, and humidity data are combined into a multi-dimensional feature vector to construct training data, and the training data is input into a dual-channel convolutional neural network model deployed on a cloud server. The dual-channel convolutional neural network includes an independent disease identification branch and a pest identification branch, wherein the disease branch takes the DSI, 840 nm reflectance, and temperature as input; the pest branch takes the NDRE, 426 nm reflectance, and humidity as input, and the outputs of the two branches are fused through a fully connected layer;

[0077] S5: The early warning threshold is determined based on the model validation set: when the model output latent period disease probability is higher than 65% or the pest infestation probability is higher than 60%, an early warning instruction is automatically triggered, which includes the type of disease and pest, the coordinate position of occurrence, and the type of biological control agent matched; the early warning instruction is pushed to the farmer terminal through the 4G / 5G mobile communication network to perform precise pesticide application operation.

[0078] Existing passion fruit disease and pest monitoring mainly relies on manual patrol or single-dimensional data collection. Traditional methods are difficult to identify diseases and pests in the latent period, and usually need to take control measures after visible lesions or yellowing appear on the leaves, at which time the plants have been substantially damaged. Some technologies use visible light imaging to identify lesions, but the passion fruit canopy is dense, the lesions on the back of the leaves are easily blocked, and visible light cannot capture early physiological changes. Some other methods rely on laboratory PCR detection, but the sampling to result feedback takes as long as 24-36 hours, and the puncture sampling damages the integrity of the plants, making it difficult to achieve large-scale continuous monitoring. Environmental parameters and spectral data are analyzed in isolation, and the quantitative relationship between "red edge band shift under high temperature and high humidity and anthracnose outbreak" cannot be established, resulting in early warning lagging behind the actual development process of diseases and pests.

[0079] To solve the above problems, the technical solution deploys a multispectral imaging device in the passion fruit planting area. The device is fixed on a height of 1.8-2.2 meters from the canopy through an adjustable support and is equipped with an automatic rotating pan-tilt to cover the front and back of the leaves. The device includes six channels of 426 nm±10 nm pest-sensitive band, 840 nm±26 nm disease-sensitive band, etc., and collects canopy images from 8:00 to 10:00 every day. An Internet of Things environmental monitoring station is also deployed to transmit temperature and humidity data to an edge computing terminal through the LoRa protocol.

[0080] The multi-spectral raw image is corrected by a diffuse white board to eliminate environmental light interference, and the calculation formula is based on the dark current calibration image and the total reflection calibration image to generate reflectivity data. The edge computing terminal extracts a 200x200 pixel area on both sides of the main vein of the leaf, calculates the reflectivity of each waveband, and generates a disease sensitive index DSI (red edge and near-infrared waveband ratio) and a pest sensitive index NDRE (normalized red edge difference index).

[0081] DSI, NDRE, 426nm reflectivity, temperature, and humidity are combined into a multi-dimensional feature vector and input into a dual-channel convolutional neural network of a cloud server. The disease recognition branch takes DSI, 840nm reflectivity, and temperature as input; the pest recognition branch takes NDRE, 426nm reflectivity, and humidity as input, and the outputs of the two branches are fused through a fully connected layer. The model verification set sets the early warning threshold to be a disease probability of 65% or a pest probability of 60%, and after triggering the early warning, automatically sends the type of plant diseases and insect pests, coordinates, and matched biological pesticide information to the farmer terminal, and can also link to the plant protection unmanned aerial vehicle for precise pesticide application.

[0082] Technical principle: Multi-spectral imaging captures blue light absorption anomalies caused by real fly eggs through the 426nm waveband, and monitors leaf structure damage caused by anthracnose through the 840nm near-infrared waveband. Environmental parameters are limited to a high incidence of plant diseases and insect pests interval of 35-40℃ temperature and 75-95% humidity, and invalid data at low temperature and low humidity are excluded. The dual-channel network design separates the learning of disease and pest characteristics: the disease branch is associated with the DSI index and high temperature environment (suitable conditions for anthracnose reproduction), and the pest branch is associated with the NDRE index and high humidity environment (real fly oviposition induction conditions), avoiding feature confusion through branch independence. Edge computing terminal preprocessing reduces cloud load, and LoRa protocol adapts to the needs of low-power wide-area transmission in farmland.

[0083] The technical scheme has the following beneficial effects: 1) Traditional visible light imaging can only identify dominant lesions, while the multispectral 426 nm band can detect potential period real fly eggs, and the red edge band (730 nm) captures the chlorophyll fluorescence changes caused by anthracnose, which is 24-48 hours earlier than the visible light scheme; 2) The prior art processes meteorological data and images separately, and cannot quantify the complex rules such as “when humidity > 85%, 840 nm reflectivity decreases by 3% to indicate anthracnose outbreak”, and the present scheme establishes an environment-spectrum collaborative criterion through feature vector fusion; 3) The automatic rotating holder solves the problem of crown layer shielding, and improves the coverage rate of leaf back by more than 40% compared with the fixed angle device; the time period limitation (8:00-10:00) avoids the interference of strong light at noon, and ensures the stability of reflectivity data. 4) The present scheme avoids damaging the plant by sampling through non-invasive online monitoring, and compresses the warning feedback time from more than 36 hours to within 10 minutes. 5) The technical scheme is aimed at the power limitation of edge devices, and places image correction and exponential calculation in the terminal, and only transmits 6-dimensional feature vectors (including DSI, NDRE, 426 nm reflectivity, 840 nm reflectivity, temperature and humidity) to the cloud, which reduces the data amount by 98% compared with the original image transmission. 6) In order to overcome the network fluctuation in farmland, the LoRa protocol is used to maintain transmission in a low signal-to-noise ratio environment, and the receiving sensitivity reaches-148 dBm, which is better than-90 dBm of the traditional WiFi.

[0084] In another technical scheme, the multispectral imaging device is fixed on an adjustable support at a height of 1.8 m to 2.2 m from the crop canopy, and an automatic rotating holder is configured to realize full-coverage of the canopy, the multispectral imaging device includes imaging channels of blue light band 450 nm±16 nm, green light band 560 nm±16 nm, red light band 650 nm±16 nm, red edge band 730 nm±16 nm, near-infrared band 840 nm±26 nm and pest-sensitive band 426 nm±10 nm, and the multispectral imaging device collects multispectral images of the guava plant canopy at a fixed time interval during 8:00 to 10:00 every day.

[0085] An adjustable support is erected in the guava planting area, and a laser range finder is used to accurately control the height of the bottom of the device from the top surface of the canopy to be 2.0±0.2 m, the horizontal rod of the support is provided with an automatic rotating holder driven by a stepping motor, the rotating angle of the holder covers 0-360° horizontally, and the adjusting range of the pitch angle is-30° to +45°. The device selects a six-channel multispectral imager, the pest-sensitive channel is strictly limited to 426 nm±10 nm, and narrow-band interference filters are used to suppress the crosstalk of the 450 nm band; the near-infrared channel is set to 840 nm±26 nm to match the characteristics of the leaf structure damage caused by anthracnose. The device is started to preheat at 7:50 every day, and the collection task is performed at 8:00-10:00 every day, and the holder is triggered to complete the east-west 120° fan-shaped scanning every time the collection is performed, and the time consumption of single scanning is 5 minutes.

[0086] In another technical solution, the environmental monitoring station collects air temperature and relative humidity; wherein, when the temperature is 35-40℃ and the humidity is 75%-95%, it is used for the generation of the feature vector.

[0087] The monitoring station is equipped with a high-precision temperature and humidity sensor, and the sampling interval is set to 5 minutes. Real-time data is transmitted to the edge computing terminal through the LoRa protocol. The terminal has a built-in data filtering module, which sets double determination conditions: temperature threshold lower limit 35℃, upper limit 40℃, humidity threshold lower limit 75%, upper limit 95%. Only when the real-time data falls within the temperature and humidity threshold interval at the same time, the system will mark this period as an "effective monitoring window" and trigger the subsequent multispectral image acquisition instruction. During the effective window period, the environmental monitoring station and the multispectral imaging device work together to ensure that the temperature and humidity data and the leaf reflectivity data are strictly aligned in time and space. The edge computing terminal binds the temperature and humidity data of the effective window period with the DSI and NDRE indices generated in step S3 to form a 6-dimensional feature vector, which is transmitted to the cloud dual-channel convolutional neural network. For non-effective window period data, the system automatically labels it as "low-risk sample" and archives it to an independent database, which is only used for model negative sample training to avoid interfering with the potential period warning criterion.

[0088] The temperature and humidity threshold is derived from the study of the biological characteristics of passion fruit pests and diseases. The conidia of the anthracnose pathogen reach the peak germination rate at 35-40℃, less than 30% at less than 35℃, and mycelial growth is inhibited at more than 40℃. The oviposition behavior of fruit flies is most active in the humidity range of 75%-95%, and the hatching rate of eggs decreases by more than 50% when the humidity is less than 75%. This biological mechanism is verified by PCR detection: the positive detection rate of anthracnose in suspected diseased leaf samples collected at a temperature of 33℃ and a humidity of 70% is only 8.7%, while the positive rate of samples in the same area jumps to 62% when the temperature rises to 36℃ and the humidity reaches 80%. The threshold setting avoids two interference scenarios: one is to avoid mistaking the temporary wilting of leaves caused by high temperature during the day as disease symptoms, which usually has a humidity of less than 60%; the other is to exclude the spectral noise when the humidity is high in the morning but the temperature is not enough, at which time the water film reflection on the leaf surface will interfere with the disease feature extraction at 840nm. The dual-channel neural network design strengthens the coupling analysis of environment and spectrum: the disease branch binds the DSI index with the temperature to capture the correlation rule that every 2.5% decrease in red edge band reflectivity within the 35-40℃ interval corresponds to an increase in the probability of anthracnose; the pest branch links the NDRE index with the humidity to establish a quantitative relationship between abnormal fluctuations in 426nm reflectivity and oviposition intensity of fruit flies within the 80%-95% humidity range.

[0089] In another technical solution, in the training stage of the dual-channel convolutional neural network model, a transfer learning mechanism and a dynamic feature library updating strategy are introduced, which specifically includes the following steps:

[0090] 1) Pre-training the model using a source region labeled guava pest and disease multispectral dataset, the source region being historical data from a plantation base in South China; migrating the convolutional layer weights of the pre-trained model to the target region model, and fine-tuning on a small sample dataset composed of leaf multispectral feature vectors collected in the target region during the podzolization period and corresponding disease-free leaf spectra confirmed by PCR detection, the small sample dataset having a sample size of no less than 200 groups;

[0091] 2) Based on the model fine-tuned in step 1), when deployed in the target region, a seasonal healthy leaf spectrum template library is established on the edge computing terminal, including a rainy season template group and a dry season template group; the rainy season template group corresponds to healthy leaf spectrum curves when the air relative humidity is greater than or equal to 85%, and the dry season template group corresponds to healthy leaf spectrum curves when the air relative humidity is less than or equal to 75%; the multispectral imaging device automatically matches and calibrates the template group according to the real-time collected environmental humidity data;

[0092] 3) The edge computing terminal selects at least 10 healthy leaves determined by the model every day, calculates the average spectrum as the reference spectrum, and uses the seasonal template group selected in step 2) to calculate the spectral angle distance SAD between the current healthy leaf reference spectrum and the matching template every day, the calculation formula being:

[0093] Wherein is the standard reflectance of the matching seasonal template group at band i, R i is the reflectance of the reference spectrum at band i, n is the number of bands; when the SAD value is less than 0.15 radians, the current reference spectrum curve is added to the corresponding seasonal template library;

[0094] 4) Apply the updated seasonal template group in step 3) to multispectral image correction, and input the corrected feature data into the fine-tuned model in step 1); add a domain adaptation layer before the fully connected layer in the model fine-tuning training stage, which uses the maximum mean difference algorithm to calculate the feature distribution difference between the source domain and the target domain, and adds it to the loss function after weighting by a weight coefficient β, β ∈ [0.3, 0.5]; only the template matching mechanism is enabled in the deployed inference stage.

[0095] The existing guava disease and pest early warning model faces significant obstacles when deployed across regions. When the model trained in the South China region is directly applied to new regions such as Yunnan or Guizhou, the baseline of the spectral reflectance of healthy leaves will systematically shift due to differences in soil composition, changes in light conditions, and the alternation of rainy and dry seasons. For example, the reflectance of leaves at the 426 nm band is generally about 12% higher in high-humidity environments during the rainy season than in the dry season. The traditional static template library cannot adapt to this seasonal spectral drift, resulting in feature discrimination errors. At the same time, fixed model parameters cannot capture the different expressions of disease and pest sensitive bands in different regions. For example, the response intensity of fruit fly eggs in the red edge band (730 nm) in Yunnan planting areas is 30% weaker than in the South China region, causing an increase in pest detection rate. Existing transfer learning solutions only rely on source domain data fine-tuning and do not establish a dynamic regional adaptability mechanism, limiting the model's generalization ability.

[0096] The technical solution constructs a cross-regional adaptive early warning system, including the following steps:

[0097] First, a dual-channel convolutional neural network is pre-trained on the cloud server side using historical labeled data sets from the South China region. The data set contains 2000 groups of multispectral feature vectors of anthracnose and fruit fly pest. After pre-training, the convolutional layer weights are frozen, and only the fully connected layer is kept in a trainable state. When the model is deployed to the target region of Yunnan, collect local paddy soil period leaf samples, and confirm 200 groups of healthy and diseased leaf data through PCR detection to form a small sample data set. Fine-tune the fully connected layer with this data set to learn the specific spectral response patterns of the target region.

[0098] A seasonal spectral template library is established on the edge computing terminal. The rainy season template group includes the spectral curves of healthy leaves when air humidity is ≥85%, and the dry season template group corresponds to data when humidity is ≤75%. The multispectral imaging device automatically matches the template group based on real-time humidity, such as calling the rainy season template calibration image when the humidity is 88%. Select 10 healthy leaves determined by the model daily and calculate their six-band average reflectance as the baseline spectrum. Calculate the difference between the baseline and the current seasonal template using the spectral angle distance (SAD) algorithm: when the SAD value exceeds 0.15 radians, add the new baseline spectrum to the corresponding seasonal template library, allowing the template library to evolve dynamically with the physiological state of the leaves.

[0099] A domain adaptation layer is inserted between the convolutional layer and the fully connected layer during the model fine-tuning stage. This layer uses the maximum mean difference algorithm to quantify the feature distribution difference between the South China source domain and the Yunnan target domain, and adds it to the loss function with a weight coefficient β = 0.4. Remove the domain adaptation layer after deployment and only keep the seasonal template matching mechanism for real-time correction.

[0100] Technical principle: The thickening of the leaf cuticle in the rainy season leads to an increase in reflectivity at the 840 nm near-infrared band, and the weakening of transpiration in the dry season leads to an increase in blue light absorption at 426 nm. The dual-template design decouples the differentiated effects of humidity on the spectrum, reducing the false positive rate by 40% compared to a single template. The spectral angle distance calculation excludes the interference of light intensity differences and focuses on the changes in reflectivity curve shape. The 0.15 rad threshold ensures that the template is only updated when the leaf physiological state deviates significantly from the historical baseline, avoiding temporary environmental noise pollution to the template library. The maximum mean difference algorithm maps the source domain and target domain feature distribution through the reproducing kernel Hilbert space, minimizing the statistical distance between the two. The weight coefficient β = 0.4 balances the contribution weight of fine-tuning data and prior knowledge in the new domain, preventing overfitting in the target domain.

[0101] Compared with the static transfer learning method, the technical scheme has the following beneficial effects: 1) The traditional method uses a fixed healthy spectrum baseline and does not distinguish between rainy and dry seasons. In this scheme, the two template groups are matched with humidity, and in the rainy season test in Yunnan, the specificity of identifying anthracnose is improved from 74% in the traditional scheme to 93%, and the false positive rate is reduced by 19 percentage points. 2) Existing domain adaptation technology only aligns the feature distribution in the training stage, and cannot be continuously optimized after deployment. This scheme combines online template library updating and offline domain adaptation layer, so that the F1 score of the model on new samples in the dry season in Guizhou remains stable at more than 92%, with a fluctuation range less than 35% of the traditional method. 3) Compared with the global template updating strategy, this scheme filters out invalid update requests through the SAD threshold, reducing the daily calculation amount of the edge terminal by 62% and compressing the storage occupancy to 1 / 3 of the traditional scheme.

[0102] Deployment optimization and limitation avoidance: To address the storage pressure caused by the expansion of the template library, a PCA dimensionality reduction engine is designed on the cloud server side: the six-dimensional original spectrum is compressed to a three-dimensional feature vector, retaining 95% of the spectral variation information and reducing the data volume by 50%. The edge terminal uses a circular queue storage mechanism, and each seasonal template group only retains the last three versions, with the oldest data automatically discarded when the number of versions exceeds. To address the convergence delay caused by the domain adaptation layer, a two-stage training strategy is adopted: in the early stage, the domain loss is closed (β = 0) to accelerate parameter convergence, and in the later stage, β = 0.4 is enabled to optimize feature alignment. In the transition season (such as the first week of the dry-to-rainy season), when the humidity fluctuates within 75%-85% for five consecutive days, the system runs the dual-template group in parallel to output the results for voting, avoiding discrimination shock during the season switching period.

[0103] In another technical scheme, an edge computing resource collaborative management mechanism is integrated into the dynamic feature library updating strategy, including:

[0104] a) A lightweight pre-screening module is deployed on the edge computing terminal to monitor the absolute value of the reflectance difference between the 426 nm and 840 nm sensitive bands of healthy leaves in real time. When the reflectance difference of the 426 nm or 840 nm band exceeds the healthy baseline range ± 15%, the full-band spectral angle distance (SAD) calculation process is triggered. Otherwise, skip the SAD calculation and directly use the current seasonal template group;

[0105] b) The SAD calculation and template update operations of step 3) are limited to the 00:00-05:00 period. During this period, the convolution layer operation of the disease recognition branch is suspended, only the 426 nm band monitoring function of the pest recognition branch is maintained, and 80% of the computing resources are allocated to the SAD process through the task scheduler of the edge computing terminal;

[0106] c) A spectral compression engine is established on the cloud server side to perform PCA dimensionality reduction on the original spectral data of the seasonal template group, compressing the original six-dimensional spectral data to a three-dimensional feature vector. The compressed feature vector is returned to the edge computing terminal through the 4G / 5G network;

[0107] d) The edge computing terminal establishes a three-version circular queue storage area for each seasonal template group, retaining only the last three versions of compressed spectral data. When the number of versions exceeds three due to the addition of new templates, the oldest version of data in the queue is automatically deleted and the storage space is released.

[0108] Traditional edge computing terminals face significant resource management defects when performing dynamic feature library updates. Existing technologies use a full-time spectral angle distance calculation strategy, such as performing a complete six-band SAD calculation and template update every two hours, resulting in continuous high load of terminal computing resources. This mechanism is particularly inefficient in farmland monitoring scenarios, as pest identification tasks only require 426 nm band monitoring at night, but traditional systems still force full convolution operations, resulting in more than 80% of invalid power consumption. At the same time, the original spectral data is directly stored locally, and a single six-dimensional spectral vector occupies storage space. With the accumulation of rain and drought dual-template groups, the storage occupancy expands to that of the traditional solution after three months. The limited resources of the edge computing terminal are continuously squeezed by inefficient tasks, ultimately leading to an increase in real-time disease and pest monitoring response delay and even task interruption.

[0109] The core scheme of the application: the technical scheme constructs a resource coordination management mechanism of edge cloud coordination. A light-weight pre-screening module is deployed on the edge computing terminal, which continuously monitors the reflectivity difference of healthy leaves in the 426 nm pest-sensitive waveband and the 840 nm disease-sensitive waveband. When the reflectivity of any waveband deviates from the current healthy benchmark range by more than 15%, the full-waveband SAD calculation flag bit is triggered. The system strictly limits the resource-intensive operation window to the low-load period of 00:00 to 05:00 every day, at which time the convolution operation of the disease identification branch is suspended and only the pest waveband monitoring function is maintained. The edge task scheduler allocates 80% of the computing resources to the SAD process, and sends a spectrum compression request to the cloud.

[0110] The cloud PCA dimension reduction engine receives the original data of the seasonal template group, compresses the six-dimensional spectrum to a three-dimensional feature vector through principal component analysis, and retains 95% of the spectral variation information. The compressed feature vector is returned to the edge terminal through the 4G / 5G network. The edge terminal creates a three-version circular queue storage area for each seasonal template group, and automatically covers the oldest version when a new template is added. The health benchmark range of the pre-screening module is updated at 05:00 every day, and dynamic calibration is achieved by synchronizing the reflectivity extreme value with the latest template group.

[0111] Principle: The light-weight pre-screening is based on the principle of detecting crop physiological characteristics mutation. The abnormal fluctuation of 426 nm waveband reflectivity usually precedes the change of other wavebands in the early stage of pest infestation, and the 840 nm waveband is sensitive to the damage of leaf structure caused by anthracnose. The 15% threshold value is derived from the field statistical data of the upper limit of seasonal reflectivity fluctuation of healthy passion fruit leaves, which can filter the normal fluctuation caused by light change.

[0112] The early morning period limitation takes advantage of the natural idle period of the farmland monitoring system, when the temperature and humidity are usually lower than the disease and pest activity threshold, and the pest identification branch only needs to maintain the minimum frequency scanning. The PCA dimension reduction engine uses eigenvalue decomposition of the covariance matrix to combine the strongly correlated wavebands of the original spectrum into linearly independent principal components. The circular queue storage is based on the principle of "recent data priority", which avoids the accumulation of spectral drift errors by eliminating early templates.

[0113] Compared with the global template updating scheme, the technical scheme breaks through the three-layer limitation: 1) The traditional method requires SAD calculation of the whole wave band every two hours, and triggers 12 times of resource peak value per day. After filtering by the pre-screening module, the actual triggering rate is reduced to once per day, and combined with the concentrated calculation strategy in the early morning, the CPU daily load peak value is reduced. 2) The existing technology directly transmits the original spectral data, and needs to transmit floating point numbers for single update. The scheme adopts the cloud PCA compression and feedback strategy, and the data amount is reduced to 50% of the traditional scheme, and the network transmission time is reduced on average in the test in the mountainous area of Yunnan. 3) Compared with the static template library scheme, the memory occupancy of the edge terminal is stably controlled in a fixed range by the circular queue storage. In the deployment in the Guizhou planting area in the rainy season lasting for 90 days, the storage space fluctuation range is controlled within the range of the traditional scheme, and the system restart caused by storage overflow is avoided.

[0114] In view of the pre-screening module mis-triggering problem, a sliding window filter is added in the reflectivity difference calculation layer: only when three consecutive sampling periods exceed the threshold value, the abnormality is confirmed, so as to avoid the interference of transient environmental noise. In order to cope with the feature distortion risk caused by PCA dimension reduction, the cloud engine is built-in reconstruction error monitoring, and when the reconstruction error of a batch of data exceeds 5%, it is automatically switched to four-dimensional compression mode and an alarm is given.

[0115] In the alternating period of dry and rainy seasons, a double-template group parallel verification mechanism is set: when the humidity is in the transition interval of 75%-85% for five consecutive days, the system outputs the correction results of the rainy season and the dry season at the same time, and selects the optimal solution by the voting mechanism to avoid seasonal switching misjudgment. In view of the network interruption scene, the edge terminal caches the latest compressed template group, and after the communication is restored, the cached data is uploaded first and then the incremental update is requested.

[0116] In another technical scheme, a health benchmark range dynamic updating mechanism is added in step a), which specifically includes the following operations:

[0117] a-1) Establish a health benchmark range updating module in the edge computing terminal, which extracts the reflectivity data of the 426nm wave band and the 840nm wave band from the currently matched seasonal template group, and calculates the health benchmark range of each wave band:

[0118] The minimum value formula of the health benchmark range of the 426nm wave band is:

[0119] The maximum value formula of the health benchmark range of the 426nm wave band is:

[0120] The minimum value formula of the health benchmark range of the 840nm wave band is:

[0121] The maximum value formula of the health benchmark range of the 840nm wave band is:

[0122] represents the reflectance value of the i-th sample in the current season template group at the 426 nm band;

[0123] represents the reflectance value of the j-th sample in the current season template group at the 840 nm band;

[0124] min is the minimum value function; max is the maximum value function;

[0125] a-2) The health reference range used in step a) is updated every 24 hours, and the update time is set to the period of 00:00 to 05:00 every day, and is executed after the SAD calculation and template update operation of step b) is completed;

[0126] a-3) When the season template group is updated in step 3), the health reference range recalculation is triggered immediately and the old value is overwritten.

[0127] The existing passion fruit disease and pest monitoring system has obvious defects in the maintenance of the health spectrum reference. The traditional scheme often presets fixed 426 nm and 840 nm band reflectance threshold ranges in the system initialization stage, for example, the 426 nm band is fixed at 0.18-0.22 reflectance units, and the 840 nm band is fixed at 0.45-0.55 reflectance units. This static reference cannot adapt to the natural spectrum drift of leaves with the growth cycle: when the dry season turns into the rainy season, the cutin layer of passion fruit new leaves thickens, causing the 840 nm band reflectance to generally rise by more than 12%, while the 426 nm band decreases by about 8% due to the accumulation of chlorophyll. The fixed threshold will misjudge such physiological changes as "abnormal reflectance" and trigger the full-band SAD calculation process. In the actual measurement in Yunnan planting area, the traditional system caused an average of 37 times of mis-triggering per day due to the fixed reference, which seriously occupied the edge computing resources.

[0128] To solve the above problems, the technical scheme constructs a closed-loop health reference range updating system in the edge computing terminal. The system automatically performs the following operation chain every 24 hours:

[0129] At 01:00 every day, the task scheduler wakes up the health benchmark range updating module after completing the SAD calculation and template updating. The module extracts all sample data from the currently activated seasonal template group: if it is currently in the rainy season mode, it reads all 426 nm and 840 nm band reflectance values in the rainy season template group; if it is in the dry season mode, it reads the data in the dry season template group. For the 426 nm band, the system traverses all the reflectance values of the samples in the template group, sets the minimum value as the lower limit of the daily health benchmark range, and sets the maximum value as the upper limit. The 840 nm band uses the same logic to independently calculate its minimum and maximum values as the health boundary specific to this band. The calculation results immediately overwrite the old thresholds of the previous day, for example, the 426 nm band range may be adjusted from 0.17-0.20 to 0.15-0.22 after the rainy season template group is updated.

[0130] When the dynamic feature library performs a template addition operation (i.e., a new benchmark spectrum is added after the SAD value exceeds 0.15 radian), the system interrupts the current task and immediately triggers the health benchmark range recalculation. The recalculation process is only for the complete data set after the new template, ensuring that the thresholds used by the pre-screening module always represent the latest health status. In the Guangxi planting area, the highest trigger for real-time updating is 3 times a day during the dry and rainy season switching period, which enables the pre-screening module to accurately track the natural evolution of leaf spectra.

[0131] Compared with the prior art, the present scheme has the following beneficial effects: 1) The traditional scheme uses an open-loop setting and needs to be manually calibrated regularly. The present scheme responds immediately through 24-hour forced updating and template change. In the 90-day test in Guizhou planting area, the pre-screening module false trigger rate is reduced from 41 times per day on average in the traditional scheme to less than 5 times. 2) The prior art sets a unified fluctuation tolerance (such as ±10%) for the dual-band, ignoring the differences in environmental response of the 426 nm and 840 nm bands. The present scheme independently calculates the extreme value range for each band, avoiding the triggering of invalid SAD calculation of the 426 nm band due to normal seasonal fluctuations of the 840 nm band. 3) The traditional health benchmark and spectral template library are updated separately, resulting in mismatch between the threshold and the true health status. The present scheme strictly binds the health benchmark data source to the seasonal template group, which improves the pre-screening accuracy to 98% in Yunnan deployment. 4) To prevent threshold distortion caused by abnormal template group data, three protection mechanisms are set: first, when the number of samples in the template group is less than 10, automatically expand to read the previous two versions of historical data until the statistical requirements are met. Second, if the calculation finds that the maximum value of the 426 nm band exceeds the physical upper limit of 0.25, immediately trigger the sensor calibration program and replace it with the last valid value. Finally, if the rainy season template group is not updated for three consecutive days, automatically relax the upper limit of the 840 nm band by 5% to accommodate the lag effect of plant growth, and restore the standard calculation process after the new template is injected.

[0132] In another technical scheme, the training data construction includes the following steps:

[0133] b-1) Collecting the guava plant canopy multispectral image in step S2, at the same time, marking the leaf of the same plant for puncture sampling, the puncture position is located at the geometric center point of the 200 pixel x 200 pixel rectangular area on both sides of the main vein of the leaf ± 5 cm;

[0134] b-2) Establish a unified time axis for each sample, where the multispectral image collection time is denoted as T0, and the puncture leaf sample is detected in the PCR laboratory at time T1; limit the time delay ΔT = T1-T0≤36 hours, and the overtime sample is automatically rejected;

[0135] b-3) The disease sensitive index DSI and the insect pest sensitive index NDRE generated in step S3 are bound with the PCR detection results according to the following rules:

[0136] When ΔT≤24 hours: directly use the original PCR label, and when the anthracnose Ct value is ≤32, it is marked as 1, and when the anthracnose Ct value is >32, it is marked as 0; when the fruit fly egg density is ≥5 eggs / cm 2 marked as 1, and when the fruit fly egg density is <5 eggs / cm 2 marked as 0;

[0137] When 24 hours < ΔT≤36 hours: apply an attenuation coefficient k = 1-0.02×(ΔT-24) to the PCR label, which is based on the negative correlation between the timeliness of PCR detection and the proliferation rate of the pathogen;

[0138] For samples with ΔT>24 hours, if the PCR detection is negative, it is directly marked as 0; only for positive samples, the attenuation coefficient k is applied;

[0139] b-4) Store the bound data set as: {[DSI, NDRE, temperature, humidity, ΔT}, [disease label x k, insect pest label x k]}, complete the training data construction.

[0140] The traditional guava disease and insect pest training data construction relies on the laboratory PCR detection results as the gold standard, but ignores the influence of the time delay (ΔT) from sample collection to detection completion on the authenticity of the label. The existing technology only detects cucumber mosaic virus through a single PCR primer combination, which can identify latent viruses, but does not consider the dynamic proliferation characteristics of pathogenic microorganisms within ΔT. For example, anthracnose bacteria can continue to proliferate after the leaf is detached, and the mycelium proliferation amount can reach 30% of the living body within 24 hours, while the hatching rate of fruit fly eggs increases exponentially with the extension of ΔT under suitable temperature and humidity. When ΔT>24 hours, the PCR detection result reflects the proliferation state after collection, not the true disease and insect pest state at the original multispectral image collection time, resulting in distorted feature correlation learned by the model. Especially in high temperature and high humidity environment, this kind of deviation amplification phenomenon leads to a 37% increase in model false positive rate in Yunnan planting area.

[0141] The training data set construction process of the timeliness correction is as follows:

[0142] At the same time of S2 multispectral image acquisition, the operator labels the leaf for aseptic puncture sampling on the same plant, and the sampling point is strictly limited to the geometric center of the 200 pixel x 200 pixel rectangular area on both sides of the main vein of the leaf within a range of ± 5 cm, to ensure the physiological consistency of the sampling area and the image analysis area. The puncture sample is immediately packaged in a 4°C cold chain box, and the integrated temperature and humidity sensor (accuracy ± 0.5°C) in the box continuously records the environmental data. Each sample is bound with a unique space-time identification code, including the image acquisition time T0 (such as 2025-06-01 10:00:00) and GPS coordinates. The sample reaches the PCR laboratory within 36 hours, and the detection completion time is recorded as T1. The system automatically calculates the time delay ΔT = T1-T0, and the overtime sample is directly rejected.

[0143] PCR detection results and multispectral features are bound according to dynamic rules: when ΔT≤24 hours, the original PCR label is directly used (Ct≤32 for anthracnose is positive, and the real fly egg density≥5 particles / cm 2 is positive); when 24 hours<ΔT≤36 hours, the decay coefficient k = 1-0.02×(ΔT-24) is applied to positive samples, and 0 label is maintained for negative samples. The decay coefficient is designed based on the ex vivo pathogen activity experiment: the anthracnose fungus proliferates 1.8 times at ΔT = 36 hours, and the label value needs to be adjusted by k = 0.76 to match the actual infection degree at T0. Finally, the structured data is generated: {[DSI, NDRE, temperature, humidity, ΔT], [disease label x k, pest label x k]}.

[0144] Compared to existing technologies, this technical solution has the following advantages: 1) Traditional methods only require sampling and detection to be completed within a "reasonable time," without quantifying the upper limit of ΔT. This solution, through a ±5cm constraint on the puncture area and a hard threshold of ΔT ≤ 36 hours, compresses the spatiotemporal deviation of samples from the Yunnan planting area from an average of ±8 hours to ±2 hours, reducing the feature-label mismatch rate by 52%. 2) Existing technologies directly use raw PCR values ​​without correcting for the ex vivo pathogen proliferation effect. This solution introduces a dynamic decay positive label with a k-coefficient, increasing the model's specificity for identifying latent anthrax from 78% to 93% in Guangxi experimental data. 3) Traditional methods retain a 0-label for negative samples exceeding the time limit, but this may lead to false negatives due to pathogen death. This solution maintains a 0-label for negative samples with ΔT > 24 hours, avoiding over-correction and diluting the true negative samples, thus ensuring that the model's specificity is not compromised. 4) To address the label jump issue during the transition period (24 hours < ΔT ≤ 36 hours), a sample temperature and humidity compensation module is deployed at the edge terminal: when the average sample transport temperature Tavg > 28℃ (e.g., during afternoon transport in summer), the nonlinear correction factor λ = e^{-0.03×ΔT×[1+0.05×(T)] is automatically activated. avg -25)]}, update k to max(0, 1-0.02×(ΔT-24)×λ), where λ∈[0.2, 1.0] reflects the microbial thermal acceleration effect. To reduce the computational load on the embedded terminal, pre-generate ΔT-T avg A two-dimensional lookup table is used to obtain the λ value through bilinear interpolation, with a single query taking less than 2ms. For remote areas where ΔT approaches the 36-hour critical value, a phased label pre-writing mechanism is implemented: pre-storage is initiated when ΔT reaches 30 hours, and a data retrieval command is automatically triggered if the timeout period expires. A closed-loop verification system linked to the agricultural drone spraying feedback system is implemented: multispectral data is re-measured 72 hours after spraying in the warning area; if the DSI index decreases by more than 15%, the authenticity of the original label is confirmed, and the attenuation slope parameter of the correction coefficient k is dynamically updated.

[0145] In another technical solution, in step b-3), the generation of the attenuation coefficient k further introduces a PCR correction factor λ, specifically including:

[0146] c-1) From the time of puncture sampling until the completion of PCR testing, the temperature data of the sample preservation environment was continuously recorded using an IoT environmental monitoring station, and the average temperature T during this period was calculated. avg ;

[0147] c-2) Based on ΔT and T avg Calculate the PCR correction factor λ:

[0148] λ=e -0.03×ΔT×[1+0.05×(Tavg-25)] , λ∈[0.2,1.0];

[0149] c-3) When T avgWhen the temperature is 28°C, the decay coefficient k is updated as: k = max(0, 1-0.02 x (AT-24) x lambda); otherwise, the linear decay coefficient is used;

[0150] d) Establish a two-dimensional lookup table of AT and T avg , and when AT > 24 hours, the discrete correction value of lambda is obtained by interpolation mapping.

[0151] In the existing construction of passion fruit pest training data, there is a significant deviation between the PCR detection result and the actual collection time of the pest status, mainly due to the fluctuation of the temperature and humidity of the sample storage environment. The traditional method such as the PCR detection scheme of passion fruit cucumber mosaic virus described in the patent CN110117677A can identify latent pathogens, but ignores the dynamic changes of the ex vivo sample during transportation: the mycelium of the anthracnose fungus can proliferate by 30% of the in vivo state within 24 hours at 28°C. The hatching rate of fruit fly eggs increases exponentially in a high temperature and humidity environment. The existing technology only uses a linear decay coefficient to correct the time delay effect (such as k = 1-0.02 x AT), without quantifying the synergistic effect of temperature and humidity on microbial proliferation. For example, when the sample storage temperature reaches 35°C, the accelerated activity of the pathogen causes the PCR detection value to increase by more than 40% compared to the actual collection time, and the linear model cannot reflect such non-linear deviation, resulting in the model learning distorted pest feature associations.

[0152] The technical solution reconstructs the label decay mechanism through a non-linear correction factor lambda, and the specific process is as follows:

[0153] After puncture sampling, the sample is sealed in a cold chain box with a built-in temperature and humidity sensor. The sensor records the sample environment temperature every 5 minutes and transmits it to the edge computing terminal through the LoRa protocol. After the laboratory completes the PCR detection, the system extracts the average temperature Tavg from the collection time T0 to the detection completion time T1, and calculates the time delay AT = T1-T0. For positive samples with AT > 24 hours, the lambda factor generation process is started:

[0154] 1) Temperature-sensitive acceleration factor calculation: based on the microbial thermodynamic model, lambda = e -0.03xATx [ 1+0.05x(Tavg-25) ], where 0.03 is the basic decay rate of the anthracnose fungus and 0.05 is the temperature sensitivity coefficient;

[0155] 2) Non-linear decay correction: when Tavg > 28°C, the linear decay coefficient k is updated as k = max(0, 1-0.02 x (AT-24) x lambda); the original k value is maintained in a low temperature environment (Tavg ≤ 28°C);

[0156] 3) Two-dimensional lookup table application: Pre-generate ΔT-Tavg discrete mapping table, ΔT with 2-hour interval, Tavg with 1℃ interval, obtain λ in real time through bilinear interpolation, single query time <2ms, avoid edge terminal floating point operation overload.

[0157] Technical principle: λ factor design is derived from the thermodynamic characteristics of microbial proliferation: anthracnose in T avg =30℃, the division cycle is shortened to 18 hours (26 hours at room temperature 25℃), and the 1+0.05×(T avg -25) term in the λ formula directly relates temperature and pathogen metabolic rate.e -0.03AT The term quantifies the natural decay of pathogen base within ΔT, and the 0.03 coefficient is fitted through in vitro bacterial culture experiments with an error rate of <5%. The λ correction is applied to positive samples to avoid false negatives caused by natural death of pathogens in low temperature environment. The actual measurement in Guangxi planting area shows that the anthracnose label value is reduced by 28% after λ correction compared with the linear model under the condition of Tavg=32℃ and ΔT=30 hours, which is more consistent with the actual infection state at T0.

[0158] This technical solution breaks through the following three limitations: 1) Traditional methods only rely on fixed time delay threshold (such as ΔT≤36 hours), and do not integrate the influence of temperature on pathogen activity. This scheme introduces an Arrhenius equation variant through the λ factor, which reduces the label correction error from ±35% in the traditional scheme to ±8% in the Yunnan rainy season test; 2) Existing technologies require cloud computing to perform complex exponential operations, and this scheme reduces the CPU load of the edge terminal by 90% through pre-generated lookup table and interpolation mechanism, which is suitable for unstable areas of farmland network; 3) The linear model has a mislabeling rate of 42% when Tavg>30℃, this scheme blocks negative labels through max(0,·) function, and combines with the feedback closed-loop verification of pesticide application (such as DSI index decreases by >15% 72 hours after pesticide application in the early warning area to confirm the effectiveness of λ), which realizes 98% label reliability in Guizhou mountainous area.

[0159] To solve the biological generalization bottleneck of the λ formula, two types of compensation mechanisms are set:

[0160] First, regional strain calibration, that is, collecting local anthracnose strains in the first deployment area, and modifying the 0.03 and 0.05 coefficients through in vitro culture experiments, and the coefficient of South China strain is reduced by 10% to adapt to high humidity environment;

[0161] Second, critical value fault tolerance, that is, when λ<0.2, automatically switch to laboratory reinspection process to avoid excessive decay caused by extreme high temperature (such as Tavg>40℃). To reduce the influence of cold chain box sensor failure, a double node redundancy check is designed, and the historical temperature sequence is interpolated when the deviation is >2℃.

[0162] Although embodiments of the present application have been disclosed in connection with the illustrative embodiments of the present application described above, it should be understood that the application can be embodied in many other specific forms without departing from the spirit or central characteristics of the technical solutions. The present examples are therefore to be considered in all respects as illustrative and not restrictive, and the scope of the application should be governed by the appended claims and their equivalents.

Claims

1. A guava early warning method of plant diseases and insect pests by fusing multispectral imaging, characterized by, The method comprises the following steps: S1: deploying a multi-spectral imaging device and an environmental monitoring station in the passion fruit planting area, and transmitting data to an edge computing terminal in real time through a LoRa wireless transmission protocol; S2: Collect the multispectral images of the guava plant canopy by multispectral imaging equipment at intervals, and eliminate the interference of environmental light by diffuse reflection white board correction of the original image, and the correction formula is: I original is the original image, I dark is the dark current calibration image, I white is the total reflection calibration image, I corrected is the corrected image; and calculating reflectance of each band S3: using the edge computing terminal to preprocess the corrected multi-spectral image, which includes extracting a rectangular region of interest of 200 pixels x 200 pixels on both sides of the main vein of the leaf; The reflectance of each band is calculated, and a disease sensitive index DSI and a pest sensitive index NDRE are generated by fusion, RE is red edge band reflectance, NIR is near-infrared band reflectance, and RED is red light band reflectance. S4: combining DSI, NDRE, 426 nm band reflectivity, 840 nm reflectivity, temperature, and humidity data into a multi-dimensional feature vector to form training data, which is input into a dual-channel convolutional neural network model deployed on a cloud server, the dual-channel convolutional neural network includes independent disease identification branches and pest identification branches, wherein the disease branch takes DSI, 840 nm reflectivity, and temperature as input; the pest branch takes NDRE, 426 nm reflectivity, and humidity as input, and the outputs of the two branches are fused through a fully connected layer; S5: The warning threshold is determined based on the model validation set: when the model output latent period disease probability is higher than 65% or the pest infestation probability is higher than 60%, an early warning instruction is automatically triggered, which includes the type of disease and pest, the coordinate position of occurrence, and the type of biological control agent matched; the early warning instruction is pushed to the farmer terminal through a 4G / 5G mobile communication network to perform precise pesticide application operation.

2. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging according to claim 1, characterized in that, The multi-spectral imaging device is fixed on an adjustable support at a height of 1.8m to 2.2m from the crop canopy, and an automatic rotating pan-tilt is configured to achieve full coverage of the canopy. The multi-spectral imaging device includes imaging channels of blue band 450nm±16nm, green band 560nm±16nm, red band 650nm±16nm, red edge band 730nm±16nm, near-infrared band 840nm±26nm, and pest-sensitive band 426nm±10nm. The multi-spectral imaging device collects multi-spectral images of passion fruit plant canopy at a fixed time interval during 8am to 10am every day.

3. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging as claimed in claim 1, wherein, The environmental monitoring station collects air temperature and relative humidity; wherein when the temperature is 35-40℃ and the humidity is 75%-95%, it is used for generating the feature vector.

4. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging as claimed in claim 1, wherein, During the training phase of the dual-channel convolutional neural network model, a transfer learning mechanism and a dynamic feature library update strategy are introduced, which specifically includes the following steps: 1) Pre-training the model using a passion fruit disease and pest multi-spectral dataset labeled in the source region, the source region being historical data from a planting base in South China; migrating the convolutional layer weights of the pre-trained model to the target region model, and fine-tuning on a small sample dataset composed of latent period leaf multi-spectral feature vectors collected in the target region and corresponding PCR detection confirmed disease-free leaf spectra, the sample size of the small sample dataset is not less than 200 groups; 2) The model fine-tuned based on step 1) is deployed in the target area, and a seasonal healthy leaf spectrum template library is established in the edge computing terminal, including a rainy season template group and a dry season template group; the rainy season template group corresponds to the healthy leaf spectrum curve when the air relative humidity is greater than or equal to 85%, and the dry season template group corresponds to the healthy leaf spectrum curve when the air relative humidity is less than or equal to 75%; the multi-spectral imaging device automatically matches and calibrates the template group according to the real-time collected environmental humidity data; 3) At least 10 healthy leaves determined by the model are selected by the edge computing terminal every day, and the average spectrum thereof is calculated as a reference spectrum; the seasonal template group selected in step 2) is used to calculate the spectral angle distance SAD between the current healthy leaf reference spectrum and the matched template every day, and the calculation formula is: wherein R is the standard reflectance of the matched seasonal template group at waveband i i R is the reflectance of the reference spectrum at waveband i, n is the number of wavebands; when the SAD value is less than 0.15 arcseconds, the current reference spectrum curve is added to the corresponding seasonal template library; 4) The seasonal template group updated in step 3) is applied to multi-spectral image correction, and the corrected feature data is input into the fine-tuned model of step 1); a domain adaptation layer is added before the full connection layer in the model fine-tuning training stage, which calculates the feature distribution difference between the source domain and the target domain by using the maximum mean difference algorithm, and the weighted coefficient β is added to the loss function after weighting, and β ∈ [0.3, 0.5]; only the template matching mechanism is enabled in the deployment and reasoning stage.

5. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging according to claim 4, characterized in that, The edge computing resource collaborative management mechanism is integrated in the dynamic feature library updating strategy, including: a) A lightweight pre-screening module is deployed in the edge computing terminal to monitor the absolute value of the reflectance difference between the healthy leaf at the pest-sensitive waveband 426 nm and the disease-sensitive waveband 840 nm in real time; when the reflectance difference at the 426 nm or 840 nm waveband exceeds the healthy reference range ± 15%, the full-waveband spectral angle distance SAD calculation process is triggered; otherwise, the SAD calculation is skipped and the current seasonal template group is directly used; b) The SAD calculation and template updating operation of step 3) are limited to the period from 00:00 to 05:00 every day; during this period, the convolution layer operation of the disease recognition branch is suspended, only the 426 nm waveband monitoring function of the pest recognition branch is maintained, and 80% of the computing resources are allocated to the SAD process by the task scheduler of the edge computing terminal; c) A spectrum compression engine is established on the cloud server side to perform PCA dimension reduction on the original spectrum data of the seasonal template group, and compress the original six-dimensional spectrum data to three-dimensional feature vectors; the compressed feature vectors are returned to the edge computing terminal through the 4G / 5G network; d) The edge computing terminal establishes a three-version circular queue storage area for each seasonal template group, and only the latest three versions of compressed spectrum data are retained; when the number of versions exceeds three due to the addition of new templates, the oldest version of data in the queue is automatically deleted and the storage space is released.

6. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging as claimed in claim 5, wherein, A dynamic updating mechanism of the healthy reference range is added in step a), which includes the following operations: a-1) A healthy reference range updating module is established in the edge computing terminal, which extracts the reflectance data of the 426 nm waveband and the 840 nm waveband from the currently matched seasonal template group, and calculates the healthy reference range of each waveband: 426 nm band health benchmark range minimum formula: 426 nm band health benchmark range maximum formula; 840 nm band health benchmark range minimum formula: 840 nm band health benchmark range maximum formula: Ri(426) represents the reflectance value of the i-th sample in the current seasonal template group at the 426 nm wavelength band; Rj840 represents the reflectance value of the jth sample in the current season template group at 840 nm wavelength band; min is the minimum function; max is the maximum function; a-2) The health benchmark range used in step a) is updated every 24 hours, with the update time set to the period of 00:00 to 05:00 every day, and the SAD calculation and template update operation of step b) is performed after completion; a-3) When the seasonal template group is updated in step 3), the health benchmark range recalculation is triggered immediately and the old value is overwritten.

7. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging as claimed in claim 1, wherein, The training data construction includes the following steps: b-1) While collecting the guava plant canopy multispectral image in step S2, mark the leaf for piercing sampling on the same plant, and the piercing position is located at the geometric center point of the 200 pixel x 200 pixel rectangular area on both sides of the main vein of the leaf ± 5 cm; b-2) Establish a unified time axis for each sample, where the multispectral image acquisition time is denoted as T0, and the piercing leaf sample is denoted as T1 when the detection is completed in the PCR laboratory; Limit the time delay ΔT = T1-T0≤36 hours, and the overtime sample is automatically rejected; b-3) Bind the disease sensitive index DSI and the insect damage sensitive index NDRE generated in step S3 with the PCR detection results according to the following rules: When ΔT ≤ 24 hours: directly use the original PCR label, anthracnose Ct value ≤ 32 is marked as 1, anthracnose Ct value > 32 is marked as 0; fruit fly egg density ≥ 5 eggs / cm 2 Marked as 1, fruit fly egg density < 5 eggs / cm 2 Marked as 0; When 24 hours < ΔT ≤ 36 hours: Apply an attenuation coefficient k = 1-0.02×(ΔT-24) to the PCR label, which is based on the negative correlation between the timeliness of PCR detection and the growth rate of pathogens; For samples with ΔT>24 hours, if the PCR detection is negative, it is directly marked as 0; Only for positive samples, apply the attenuation coefficient k; b-4) Store the bound data set as: {[DSI, NDRE, temperature, humidity, ΔT}, [disease label x k, insect damage label x k]}, complete the training data construction.

8. The method of early warning of pest and disease of passion fruit using fusion multispectral imaging as claimed in claim 1, wherein, In step b-3), the generation of the attenuation coefficient k further introduces a PCR correction factor λ, which specifically includes: c-1) continuously recording temperature data of the sample storage environment by the Internet of Things environment monitoring station from after the puncture sampling to the completion of the PCR detection, calculating the average temperature T of the period avg ; c-2) according to ΔΤ and T avg Calculate PCR correction factor λ: λ = e -0.03×ΔT×[1+0.05×(Tavg-25)] , λ e [0.2, 1.0]; c-3) when T avg > 28°C, the attenuation coefficient k is updated as: k = max(0, 1 - 0.02 x (AT - 24) x lambda); otherwise the linear attenuation coefficient is maintained. d) a two-dimensional look-up table of ΔT vs. T avg is established, and when ΔT > 24 hours, the discrete correction value of λ is obtained by interpolation mapping.

Images