A low-cost detection system for concealed pests based on multi-modal data fusion

Abstract

The application provides a low-cost concealed pest detection system based on multi-modal data fusion. The system includes a data acquisition module, a data analysis module and a user interaction module. The data acquisition module includes a handheld detection terminal and a field sound monitoring pile. The handheld detection terminal integrates an optical camera, a near-infrared spectrum sensor and a sound wave sensor, and is used for synchronously collecting external images, internal component information and pest activity sound of crops. The field sound monitoring pile integrates a sound wave sensor and is used for long-term monitoring of pests in the soil. The system realizes non-destructive early accurate identification of concealed pests such as Sitophilus zeamais, Sitophilus oryzae, Plodia interpunctella and Rhyzopertha dominica by fusing and analyzing multi-dimensional data such as images, spectra and sound. The application adopts low-cost sensors and a simplified hardware architecture, cross-verification and joint judgment of multi-modal data are performed through a confidence fusion algorithm, an early concealed detection mode is set to improve the identification ability of pests without obvious external characteristics, and the technical threshold is significantly reduced, so that small and medium-sized farmers can afford and conveniently use the system.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a low-cost detection system for concealed pests based on multimodal data fusion. Background Technology

[0002] Effective monitoring and control of agricultural pests are crucial for ensuring grain yield and quality. Among the many types of agricultural pests, the control of elusive pests is particularly difficult. Elusive pests, such as corn weevils, rice weevils, wheat moths, and grain borers, selectively hide inside crop stems, leaf strata, or deep in the soil roots during key stages of their life cycle. The damage caused by these pests is significantly concealed, making early detection difficult with conventional field survey methods. By the time crops show obvious symptoms, the damage has often already caused irreversible losses, seriously threatening agricultural production efficiency.

[0003] In existing technologies, farmers mainly rely on visual observation or experience to detect pests. This method is inaccurate and has a significant time lag, only able to identify damaged crops that have already shown external symptoms, making it difficult to effectively monitor pests in their early stages. Although some high-precision detection equipment exists, such as laboratory-grade spectrometers or professional acoustic detection devices, which can provide relatively accurate results, these devices are expensive and complex to operate, making them unsuitable for small and medium-sized farmers. Furthermore, some technical solutions rely on only a single type of sensor, such as optical images or sound signals, which have limited ability to identify damage hidden inside crop tissue, resulting in a high false positive rate. Summary of the Invention

[0004] The purpose of this invention is to propose a low-cost detection system for concealed pests based on multimodal data fusion, in order to solve the shortcomings of existing technologies such as difficulty in detecting concealed pests, low accuracy, high equipment cost, and high false judgment rate of single sensor solutions.

[0005] To achieve the above objectives, this invention proposes a low-cost detection system for concealed pests based on multimodal data fusion, comprising a data acquisition module, a data analysis module, and a user interaction module; The data acquisition module is used to collect multimodal data on crop growth status, including a handheld detection terminal and at least one field sound monitoring stake; the housing of the handheld detection terminal integrates an optical camera, a near-infrared spectral sensor and a first acoustic sensor; the field sound monitoring stake integrates a second acoustic sensor. The data analysis module is connected to the data acquisition module to receive and process multimodal data from the data acquisition module and output pest identification results. The user interaction module communicates with the data analysis module to present pest identification results to the user.

[0006] Furthermore, the front end of the handheld detection terminal housing is provided with an acoustic coupling structure, which is used to fit against the surface of crop stems or leaves to achieve physical contact acoustic coupling; the receiving end of the first acoustic sensor is located inside the acoustic coupling structure.

[0007] Furthermore, the sound coupling structure is made of soft silicone material.

[0008] Furthermore, the optical camera, near-infrared spectral sensor, and first acoustic sensor are connected to the main controller inside the handheld detection terminal via a universal interface module; the universal interface module adopts a standardized communication protocol and physical interface.

[0009] Furthermore, the field sound monitoring stake includes a solar power supply unit and a wireless communication unit; the solar power supply unit includes a solar panel and a small-capacity battery; the wireless communication unit uses low-power wide-area network technology to periodically send the collected sound wave data to the data analysis module.

[0010] Furthermore, the data analysis module performs the following steps for multimodal data fusion and identification: it analyzes the compositional changes of the data collected by the near-infrared spectral sensor, performs smoothing using the Savitzky-Golay filtering algorithm, suppresses baseline drift using standard normal variable transformation, predicts the cellulose and water content of crop tissues using a partial least squares regression model, identifies internal component anomalies, and quantifies them into a first numerical confidence level.

[0011] Furthermore, when the cellulose content of the tested area is more than 15% lower than the healthy baseline value or the moisture content is more than 20% lower than the healthy baseline value, the system determines that there is an abnormality in the internal composition of that area.

[0012] Furthermore, the data analysis module performs spectral analysis of the audio data collected by the first and second acoustic sensors to identify the characteristics of pest sounds. It uses a 50 Hz power frequency notch filter to eliminate power grid interference and a bandpass filter from 100 Hz to 8 kHz to retain the main frequency bands of pest activity sounds. It extracts the Mel frequency cepstral coefficients, spectral centroids, and sub-band energy ratio sound signature feature vectors to generate a second numerical confidence level.

[0013] Furthermore, the data analysis module performs external feature analysis on the optical image data, uses a deep learning-based image segmentation model to segment crop leaf regions, uses a convolutional neural network to extract multi-scale visual features, identifies insect holes, lesions, leaf miners, or honeydew insect frass residues, and generates a third numerical confidence score.

[0014] Furthermore, the data analysis module employs a confidence fusion algorithm based on an adaptive weighting mechanism to perform fusion calculations on the first, second, and third numerical confidence scores. When the second numerical confidence score is greater than the second threshold, the first numerical confidence score is greater than the first threshold, and the third numerical confidence score is less than the third threshold, the system activates the early concealment detection mode and increases the adjustment amplitude coefficient of the acoustic modality.

[0015] Compared with the prior art, the advantages of the present invention are: 1. This invention combines an optical camera, a near-infrared spectral sensor, and an acoustic sensor to integrate image, spectral, and sound modal data, enabling non-destructive detection and early identification of pests such as corn weevils, rice weevils, wheat moths, and grain borers that hide inside crops or in the soil. This significantly improves the accuracy and timeliness of identification and provides a reliable basis for precision prevention and control.

[0016] 2. This invention provides a flexible monitoring mode that combines portability and long-term monitoring. The system integrates a handheld detection terminal and a fixed monitoring stake in the field, which supports both rapid point detection by manual inspection and long-term unattended area monitoring. It meets the pest monitoring needs of different scenarios and stages, and significantly improves the coverage and continuity of monitoring.

[0017] 3. This invention significantly reduces technical costs and the barrier to entry by employing low-cost near-infrared sensors and ordinary optical cameras. Through modular design and integration of sensors, along with universal interface modules, hardware costs are greatly reduced. The system is easy to operate, requiring no professional training, making it affordable and convenient for small and medium-sized farmers. It effectively solves the problems of high prices and complex operation associated with traditional high-precision equipment.

[0018] 4. By effectively avoiding the limitations of single-sensor detection, the multimodal data fusion system of this invention can comprehensively utilize three types of information—external morphology, internal composition, and activity characteristics—for cross-validation and joint judgment. This overcomes the shortcomings of relying solely on images, which easily overlooks internal hazards, and relying solely on sound, which is easily affected by environmental interference, thus greatly improving the reliability of detection.

[0019] 5. By enabling early warning, the system can identify and warn pests in advance through abnormal internal components and characteristic sounds before they cause obvious external symptoms or serious internal damage. This significantly advances the control window and avoids or significantly reduces irreversible losses such as crop growth stagnation and yield decline caused by pests.

[0020] 6. This invention has the ability to monitor both underground and above-ground pests. The system includes a portable handheld terminal and a fixed field monitoring stake, which can continuously or at fixed points monitor the activities of pests inside the above-ground stems and leaves and pests in the underground root system, respectively. This realizes three-dimensional pest monitoring of the entire growth period of crops and all spatial dimensions, and expands the applicable scenarios of the system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the underlying execution architecture of the multimodal fusion decision algorithm according to an embodiment of the present invention.

[0022] Figure 2 This is a system-level data flow and functional module topology diagram of the multimodal data fusion and decision-making process according to an embodiment of the present invention.

[0023] Figure 3 This is a flowchart of the confidence-based fusion verification and judgment logic in an embodiment of the present invention.

[0024] Figure 4 This is a flowchart illustrating the judgment process of the adaptive weight adjustment mechanism in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0026] This invention provides a low-cost detection system for concealed pests based on multimodal data fusion. The system mainly consists of three functional modules: a data acquisition module, a data analysis module, and a user interaction module. The data acquisition module is responsible for acquiring multimodal raw data on crop growth status from the field. The data analysis module is responsible for fusing and intelligently analyzing the received multimodal data. The user interaction module is responsible for presenting the analysis results to the user in an intuitive format. These three modules work collaboratively via wired or wireless communication to achieve early, accurate, and low-cost detection of concealed pests.

[0027] The data acquisition module is the foundation of the system for obtaining field information, and its design balances portability with the practical needs of long-term monitoring. This module specifically includes a handheld detection terminal and one or more field sound monitoring posts. The handheld detection terminal is a portable device designed for easy field inspections by farmers. Its casing features an ergonomic grip design, and the front of the casing integrates three core sensor components: an optical camera, a near-infrared spectral sensor, and a first acoustic sensor. The optical camera is used to acquire high-definition optical images of crop leaves and stems to obtain information on externally visible damage. The near-infrared spectral sensor operates in the wavelength range of 900 nm–1700 nm and is used to non-destructively acquire near-infrared reflectance spectral data from crop tissues. This spectral information has a close correlation with the content of key physiological components such as water and cellulose within the leaves. The first acoustic sensor is used to collect the weak sound signals generated by pests feeding, crawling, and other activities within the crop tissues.

[0028] To improve the quality of sound acquisition and effectively suppress environmental noise interference, a sound coupling structure is specially designed at the front of the handheld detection terminal's housing. This structure is typically made of soft silicone material through a molding process. During use, the operator tightly adheres this sound coupling structure to the back of the stem or leaf of the crop being tested, allowing the receiving end of the first sound wave sensor to achieve good acoustic coupling with the crop tissue through physical contact. This effectively transmits the sound wave signals generated by pest activity while attenuating the interference of environmental wind noise on the acquisition process. The specific shape of the sound coupling structure is a bowl-shaped structure with a converging front end. The diameter of its open end is slightly larger than the typical size of the stem or leaf being tested, ensuring that it completely covers the area being tested, forming a closed acoustic channel. The interior of the housing is filled with sound-insulating material to further isolate the adverse effects of housing vibration on the accuracy of sound wave acquisition.

[0029] The optical camera, near-infrared spectral sensor, and primary acoustic sensor are all electrically connected to the main controller inside the handheld detection terminal via a universal interface module. This universal interface module employs standardized communication protocols and physical interface designs, including both USB and SPI bus interfaces, enabling plug-and-play integration of the three sensors. This modular design not only effectively reduces hardware costs but also significantly improves system maintainability and scalability. When a sensor fails, the user can quickly replace it without replacing the entire device.

[0030] Field sound monitoring posts are fixed devices deployed in fields for long-term, unattended monitoring. Each post integrates a second acoustic sensor within its casing, with performance specifications identical to the first acoustic sensor on a handheld detection terminal. This sensor primarily monitors the sound signals of pest activity in the soil layer or near the root zone of crops. The field sound monitoring posts further integrate two auxiliary functional units: a solar power supply unit and a wireless communication unit. The solar power supply unit consists of a solar panel and a small-capacity battery pack composed of lithium iron phosphate batteries. The solar panel has a nominal power of five watts, and the battery pack has a capacity of three thousand mAh, capable of maintaining continuous power supply for at least seventy-two hours even under extreme conditions of continuous rainy weather. The wireless communication unit uses LoRa spread spectrum communication technology, enabling the periodic transmission of collected acoustic data to a data analysis module, achieving remote data feedback. The casing of the field sound monitoring post is manufactured using polyoxymethylene engineering plastic through injection molding, exhibiting excellent corrosion resistance and aging resistance, allowing for long-term stable operation in field environments without frequent maintenance.

[0031] The data analysis module is the core of the entire system's intelligent processing. It can be deployed on the local embedded processor of the handheld detection terminal or on a cloud server for higher computing power. This module is responsible for receiving and fusing multimodal data from the data acquisition module, analyzing the data step-by-step according to a preset processing flow, and ultimately outputting the pest identification results. (Refer to...) Figure 2 As shown, the data processing and fusion process includes the following four main stages.

[0032] The first stage is the data processing for the near-infrared spectral composition variation analysis channel. After acquiring the raw spectral data from the near-infrared spectral sensor, the module first performs preprocessing operations, specifically including two steps. The first step uses the Savitzky-Golay filtering algorithm to perform multi-point smoothing on the raw spectral data to eliminate random noise. In this embodiment, the selected window width is eleven sampling points, and the polynomial order is second. Smoothing effectively suppresses the interference of high-frequency random noise on spectral characteristics. The second step performs standard normal transformation to suppress spectral line shifts caused by differences in sample surface scattering and baseline drift.

[0033] After preprocessing, the system uses a pre-trained partial least squares regression model to perform quantitative analysis on the processed spectral data and predict the cellulose and moisture content of the current detection site based on the spectral characteristics.

[0034] The component anomaly identification process is achieved through two methods: one is to establish a component benchmark database of healthy crop tissues at different growth stages, and compare the component content of the currently detected part with the health benchmark values ​​of the same variety and growth stage item by item; the other is to conduct multiple measurements of the same detection area at different time points, and identify abnormal trends in component content through time series analysis. When the cellulose content of the detected part is more than 15% lower than the health benchmark value, or the moisture content is more than 20% lower than the health benchmark value, the system determines that there is an internal component anomaly caused by tissue damage due to pest feeding in that area, and quantifies the degree of anomaly as a value between zero and one as the component anomaly confidence level C2 output, where zero represents complete health and one represents severe anomaly.

[0035] The second stage is the audio data processing of the acoustic wave feature recognition channel. After the module synchronously receives audio signals from the first and second acoustic wave sensors, it first performs acoustic signal preprocessing, specifically including the following two filtering steps. The first step is to eliminate power grid interference and its harmonic components using a 50Hz power frequency notch filter. The second step is to retain the main frequency band of pest activity sounds using a 100Hz-8000Hz bandpass filter. After the above preprocessing, environmental background noise in the acoustic signal is effectively suppressed, and the target sound signal generated by pest activity is highlighted.

[0036] The preprocessed continuous audio signal is divided into frames according to preset frame length and overlap rate parameters. In this embodiment, the frame length is set to 25ms. For each frame signal, the system extracts a set of voiceprint feature vectors that can characterize the essential features of the sound. The feature vectors contain the following three types of feature parameters: Mel frequency cepstral coefficients, which are extracted in 13 dimensions and spectral decomposition is performed using 26 Mel filter banks; the spectral centroid reflects the frequency centroid distribution of the sound signal; and the sub-band energy ratio, which divides the frequency band from 100Hz to 8000Hz into 8 equal-width sub-bands and calculates the ratio of the energy of each sub-band to the total energy as a feature parameter.

[0037] In the offline phase, the system pre-established a background acoustic signature template library for healthy crop tissues and a acoustic signature feature template library for target pests. The background acoustic signature template library contains acoustic feature spectra of 20 major crop varieties at different growth stages, with each crop's template consisting of the statistical mean and covariance matrix of at least 120 samples. The pest acoustic signature feature template library contains acoustic signature feature spectra of feeding and crawling sounds of 12 major elusive pests, including corn weevils, rice weevils, wheat moths, and grain borers. Each pest's template is also based on statistical modeling of no less than 100 positive samples. During real-time analysis, the system performs Gaussian mixture model likelihood matching calculations on the extracted real-time acoustic signature feature vectors with both the background and pest templates. By comparing the matching distance differences between the real-time features and the two types of templates, and combining this with a preset likelihood difference threshold, the system determines whether there are any suspicious pest activity acoustic signature events, and generates an acoustic signature matching confidence score C1 as the output of the second stage.

[0038] The third stage is the optical image processing of the external image detection channel. After receiving the raw image data acquired by the optical camera, the module first performs image preprocessing, including the following three steps: image enhancement using an adaptive histogram equalization algorithm to improve contrast; white balance correction using a gray-world white balance algorithm to eliminate the influence of light source color shift; and noise suppression using a bilateral filtering algorithm to smooth image details while maintaining edge sharpness.

[0039] After preprocessing, the system uses a pre-trained deep learning-based image segmentation model to process the images. This segmentation model adopts an improved version of the U-Net architecture, using ResNet-50 as the encoder backbone and adding an attention mechanism module to the decoder. The segmentation model is trained on a dataset containing 20,000 labeled images. The segmentation model accurately segments crop leaf regions from complex field backgrounds, forming regions of interest containing only the target crop tissue.

[0040] Within the segmented regions of interest of the leaf blade, the system further uses a convolutional neural network to extract multi-scale visual features. This embodiment employs EfficientNet-B4 as the feature extraction network.

[0041] The extracted visual features of the damage are matched with a pre-trained image feature library of pest damage, which contains image samples of 1800 typical damage symptoms caused by 12 major elusive pests. The system calculates the image feature confidence score C3 based on the matching similarity as the output of the third stage. At the same time, the pixel-level location of the damage symptoms is marked on the image and its severity is assessed, which is divided into three levels: mild, moderate and severe.

[0042] The fourth stage is the multimodal fusion decision-making and final recognition result generation process. (Refer to...) Figure 3 As shown, the module fuses the component anomaly confidence score C2, voiceprint matching confidence score C1, and image feature confidence score C3 obtained from the three independent analysis channels to make a final determination of pest presence. The fusion algorithm adopts a confidence fusion strategy based on an adaptive weighting mechanism, and its comprehensive judgment function is defined as follows:

[0043] in, The final score is the combined score after merging. This represents the Sigmoid activation function. For the first The adaptive weight adjustment function corresponding to each confidence level is defined as:

[0044] in, For baseline weights, To adjust the amplitude coefficient, This is the adaptive threshold corresponding to the confidence level. For smoothing coefficients; To preset static weight coefficients, satisfying ; Reference Figure 4 The adaptive weight adjustment mechanism shown is the core innovation of this invention. When the system detects... and and When the system determines that the current detection scenario matches the typical characteristics of early-stage concealed pests, it automatically activates the early-stage concealed detection mode branch. At this point, the adjustment function... In The dynamic increase enhances the contribution of acoustic modes to decision-making, enabling the identification of early-stage, concealed pests without obvious external features; it significantly improves the system's detection sensitivity for early-stage, concealed pests.

[0045] When the confidence combination does not meet the above triggering conditions, the system uses a conventional fusion mode branch for weight adjustment, and each modality is fused according to the preset baseline weights and adjustment parameters. The static weight coefficients of the three modalities are as follows: The initial values ​​were determined through optimization based on a large amount of field trial data, among which the image modality weights... Set to 0.35, spectral mode weight Set to 0.35, acoustic modal weights The weighting coefficient is set to 0.3. The above weighting coefficient can be reconfigured on-site according to the type of target pest and the detection scenario.

[0046] The underlying execution architecture flow of the multimodal fusion decision algorithm of this invention is as follows: Figure 1As shown, the process begins in the parameter initialization module, where the system automatically loads the preset baseline weight coefficients and adaptive threshold parameters upon startup.

[0047] After parameter initialization, the confidence calculation module is entered, and three parallel output results are calculated simultaneously: acoustic modal matching confidence C1, near-infrared component anomaly confidence C2, and optical image feature confidence C3.

[0048] The system's built-in trigger condition judgment module evaluates the three confidence levels in real time, and enters the dynamic weight adjustment module once the preset trigger logic conditions are met.

[0049] In the fusion calculation iteration module, when the iteration count is greater than 2, the dynamic weight allocation function is activated. Key calculation nodes with counts of 3, 4, 8, etc., sequentially perform weighted summation, nonlinear activation, and comprehensive score accumulation operations. When the final calculated comprehensive score S exceeds the preset fusion judgment threshold, the data analysis module outputs pest identification results including pest species, probability of existence, and damage level. These results are then graphically presented to the user through the result visualization module.

[0050] The user interaction module presents the pest identification results generated by the data analysis module to end users in an intuitive and user-friendly format. This module has two optional implementations. The first is a small LCD screen integrated into the handheld detection terminal. The second is a dedicated mobile application installed on the user's smartphone or tablet. This application communicates with the handheld terminal or data analysis server via a wireless network to receive and display richer pest reports, historical trend graphs, and control recommendations, providing a better user experience.

[0051] The typical workflow of this invention system in actual field applications is as follows: Farmers carry a handheld detection terminal into the crop planting area to be detected, tightly attach the sound coupling structure to the back of the selected crop stem or leaf, and initiate a complete detection process by pressing the detection trigger button on the side of the device. The data acquisition module simultaneously acquires raw data in three modalities: optical image, near-infrared spectrum, and contact sound signal from the detection site. The acquisition process lasts approximately three seconds. After acquisition, the main controller standardizes the format of the three raw data streams through a general interface module, and then selects either local processing or cloud upload processing mode based on the current network connection status. In local processing mode, the ARM Cortex-A72 processor built into the main controller runs a trimmed and optimized neural network model to perform data analysis; in cloud processing mode, the data is uploaded to the cloud server via a 4G mobile network for full model inference calculation. Regardless of the processing mode used, the data analysis module performs parallel processing and fusion analysis of the multimodal data according to the aforementioned four-stage process, and the final pest identification result 20 is displayed to the farmer within five seconds through the user interaction module 8.

[0052] Meanwhile, multiple field sound monitoring stakes deployed in the fields continuously collect sound wave data according to a preset working cycle. Each monitoring stake completes a sampling cycle every fifteen minutes and sends the data to a cloud server for continuous analysis. When a monitoring stake detects suspicious insect activity sound patterns continuously appearing in the soil layer, the system automatically pushes an early warning notification to the mobile application of the associated farmer, reminding the farmer to go to the corresponding area to use a handheld detection terminal for further on-site confirmation and detection.

[0053] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A low-cost detection system for concealed pests based on multimodal data fusion, characterized in that, It includes a data acquisition module, a data analysis module, and a user interaction module; The data acquisition module is used to collect multimodal data on crop growth status, including a handheld detection terminal and at least one field sound monitoring stake; the housing of the handheld detection terminal integrates an optical camera, a near-infrared spectral sensor and a first acoustic sensor; the field sound monitoring stake integrates a second acoustic sensor inside. The data analysis module is communicatively connected to the data acquisition module and is used to receive and fuse multimodal data from the data acquisition module to output pest identification results. The user interaction module is communicatively connected to the data analysis module and is used to present the pest identification results to the user.

2. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The handheld detection terminal has a sound coupling structure at the front end of its housing. The sound coupling structure is used to fit against the surface of crop stems or leaves to achieve physical contact acoustic coupling. The receiving end of the first acoustic sensor is located inside the sound coupling structure.

3. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 2, characterized in that, The sound coupling structure is made of soft silicone material.

4. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The optical camera, the near-infrared spectral sensor, and the first acoustic sensor are connected to the main controller inside the handheld detection terminal via a universal interface module; the universal interface module adopts a standardized communication protocol and physical interface.

5. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The field sound monitoring pile includes a solar power supply unit and a wireless communication unit; the solar power supply unit includes a solar panel and a small-capacity battery; the wireless communication unit uses low-power wide-area network technology to periodically send the collected sound wave data to the data analysis module.

6. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The data analysis module performs the following steps for multimodal data fusion and identification: it analyzes the compositional changes of the data collected by the near-infrared spectral sensor, performs smoothing processing using the Savitzky-Golay filtering algorithm, suppresses baseline drift using standard normal variable transformation, predicts the cellulose and water content of crop tissues using a partial least squares regression model, identifies internal component anomalies, and quantifies them as the first numerical confidence level.

7. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 6, characterized in that, When the cellulose content of the tested area is more than 15% lower than the healthy baseline value or the moisture content is more than 20% lower than the healthy baseline value, the system determines that there is an abnormality in the internal composition of that area.

8. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The data analysis module performs pest characteristic sound spectrum analysis on the audio data collected by the first and second acoustic sensors. It uses a 50 Hz power frequency notch filter to eliminate power grid interference and a 100 Hz to 8 kHz bandpass filter to retain the main frequency bands of pest activity sounds. It extracts the Mel frequency cepstral coefficients, spectral centroids, and sub-band energy ratio sound signature feature vectors to generate a second numerical confidence level.

9. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The data analysis module performs external feature analysis on optical image data, uses a deep learning-based image segmentation model to segment crop leaf areas, uses a convolutional neural network to extract multi-scale visual features, identifies insect holes, lesions, leaf mint marks, or honeydew insect frass residues, and generates a third numerical confidence score.

10. The low-cost detection system for concealed pests based on multimodal data fusion according to claim 1, characterized in that, The data analysis module uses a confidence fusion algorithm based on an adaptive weighting mechanism to calculate the first, second, and third numerical confidence scores. When the second numerical confidence score is greater than the second threshold, the first numerical confidence score is greater than the first threshold, and the third numerical confidence score is less than the third threshold, the system activates the early concealment detection mode and increases the adjustment amplitude coefficient of the acoustic modality.

Images