en.Wedoany.com Reported - Water conservancy projects are critical infrastructure for ensuring national flood control safety, water supply security, food security, and ecological security. Their safe and stable operation directly affects the safety of people's lives and property, sustainable socio-economic development, and the health and balance of ecosystems. Termites are the "invisible killers" of embankments in water conservancy projects, characterized by strong concealment, rapid reproduction, wide-ranging damage, and severe disaster consequences. They build long-term nests inside embankments, excavating intricate tunnel systems that directly compromise the integrity and stability of the earth structure. This increases soil porosity and reduces shear strength. Under high water levels during flood seasons, this can easily trigger major hazards such as piping, seepage, landslides, and even dam breaches, causing immense losses. During the catastrophic 1998 Yangtze River flood, approximately 80% of embankment hazards were triggered by termite damage, fully highlighting the importance of termite control in the safety management of water conservancy projects.

China currently has a total embankment length exceeding 300,000 km, a large portion of which was built between the 1950s and 1970s. These embankments suffer from low construction standards, long operational lifespans, and limited maintenance conditions, making termite infestation a particularly prominent problem. According to statistics, the termite infestation rate for embankments in southern river basins like the Yangtze, Pearl, and Huaihe Rivers exceeds 60%, and for old embankments, it is over 80%. With global warming, the distribution range of termites is expanding northward, and the severity of infestation is continuously increasing, presenting unprecedented challenges for termite control in water conservancy projects.

For a long time, termite monitoring in China's water conservancy projects has primarily relied on traditional methods such as manual inspection, bait trapping, and manual excavation verification. These methods are not only labor-intensive, inefficient, and limited in coverage but also heavily dependent on the experience of monitoring personnel. They suffer from significant issues like delayed response, high misjudgment rates, and numerous monitoring blind spots, making it difficult to achieve early detection and pre-warning of termite hazards. With the rapid development of the Internet of Things (IoT), Artificial Intelligence (AI), and sensor technology, automated and intelligent termite monitoring technology has gradually become a research hotspot. Monitoring devices based on single modalities like vision, sound, and temperature are being gradually applied in engineering practice, but they still face significant technical bottlenecks: the visual modality is easily affected by underground darkness, humidity, sediment occlusion, and foreign object interference, resulting in insufficient recall rates for small termite targets (3-5mm) and prominent missed detection issues; the sound modality is susceptible to environmental noise interference such as rainwater scouring, soil vibration, insect and rodent activity, and water flow sounds, leading to poor environmental adaptability and false alarm rates generally exceeding 8%; the temperature modality can only identify significant thermal anomalies generated by large nests, failing to capture the weak temperature rise signals from the initial activities of scattered individual termites, posing a high risk of missed alarms. Furthermore, single modalities lack an effective cross-validation mechanism. In the complex and variable environment of water conservancy projects, it is difficult to simultaneously ensure high recognition rates and low false/missed alarm rates, failing to meet practical engineering monitoring needs.

Currently, domestic and international research has not yet formed a mature multi-modal collaborative sensing and deep fusion termite monitoring technology system. Foreign-developed termite monitoring devices are mostly based on a single modality, suffering from low recognition accuracy, short battery life, high cost, and inadaptability to China's complex water conservancy environments. Domestic research in China started relatively late. Most existing products use single-sensor technology. Some attempts at multi-modal fusion also suffer from insufficient fusion depth, poor algorithm robustness, and low engineering maturity, failing to achieve efficient collaboration and precise fusion of multi-modal data, thus unable to meet the actual needs of termite monitoring in water conservancy projects.

To this end, this paper focuses on the core technology of "vision-sound-temperature" three-modal fusion, conducting research, development, and application validation of an underground passive termite monitoring device. It focuses on breaking through key technologies such as multi-modal data acquisition, preprocessing, feature fusion, decision verification, termite species and caste identification, and nest location prediction. Through extensive laboratory testing and nationwide multi-scenario field applications in water conservancy projects, the device's performance and practicality are comprehensively validated, providing implementable and scalable core technical support for intelligent termite monitoring and early warning in water conservancy projects, promoting the transformation of termite control in China's water conservancy projects from "manual inspection, post-event treatment" to "intelligent monitoring, pre-event warning".

Construction of Multi-Modal Sensing System and Data Acquisition

1. Multi-Modal Sensor Configuration and Signal Characteristics

The multi-modal sensor monitoring device adopts an integrated passive underground structure design, requiring no external power supply or wiring. It can be directly buried in areas with high termite activity. Using image, sound, and temperature sensors as the core sensing units, it constructs a multi-dimensional collaborative sensing system for termite activity, achieving synchronous capture of multi-source signals at the same time and space, providing high-quality data support for subsequent multi-modal fusion recognition.

Visual Modality Sensing: Utilizes a 5-megapixel high-resolution CMOS image sensor and a 120° wide-angle low-distortion lens. Exposure parameters and focusing strategies are optimized for small termite targets, enabling clear identification of individual termites as small as 3mm. To adapt to the dark, humid underground environment, a low-light imaging scheme (minimum illumination 0.01 lux) is adopted, coupled with adaptive noise reduction and contrast enhancement algorithms. This effectively suppresses noise interference in the underground environment, ensuring stable output of core features such as termite morphology, body surface texture, movement trajectory, and population density without external light sources, providing clear and effective image data for visual recognition.

Sound Modality Sensing: Employs a -30dB high-sensitivity MEMS acoustic sensor to capture weak acoustic signals generated by termite activities such as feeding (gnawing wood, soil), crawling, and trophallaxis.

Temperature Modality Sensing: Uses a high-precision NTC temperature sensor with a measurement range of -40 to 85°C, accuracy of ±0.1°C, and response time ≤100ms. It can precisely capture local micro-temperature rise signals of 0.3 to 1.5°C generated by termite colony metabolic activities. The sensor adopts a multi-point layout, with three temperature measurement points evenly distributed inside the device. Through continuous sampling and sliding window filtering algorithms, it monitors temperature change trends in real-time, effectively distinguishing termite metabolic heat signals from interference items like natural soil thermal disturbances, solar radiation, and water heat conduction, providing reliable temperature data for temperature recognition.

2. Temperature-Humidity Fusion Sensing and Environmental Compensation

The on-site environment of water conservancy projects is complex and variable. Environmental factors like high humidity and extreme temperature changes can cause temperature sensor drift, affecting the reliability of temperature modality recognition. To solve this problem, this study designs temperature-humidity fusion sensing and environmental compensation logic, enabling the humidity sensor and temperature sensor to work collaboratively, achieving precise correction of temperature signals and enhanced anti-interference capability. The specific implementation logic is as follows: First, synchronously collect ambient temperature and relative humidity data. Based on extensive experimental data on termite activity, construct a temperature-humidity-termite activity correlation model, clarifying the activity intensity and metabolic heat characteristics of termites under different temperature and humidity conditions, providing a theoretical basis for subsequent fusion recognition. Second, when the ambient relative humidity > 85%, activate the temperature drift compensation algorithm. Based on the correlation between humidity value and temperature drift amount, correct the temperature measurement data in real-time, rectifying sensor measurement errors in high-humidity environments to ensure temperature data accuracy. Third, define the high termite activity zone (temperature 25-30°C, humidity 60%-80%). Within this zone, termite activity is frequent, and metabolic heat signals are obvious. Appropriately increase the thermal confidence weight output by the temperature modality to enhance recognition sensitivity. Fourth, in extreme temperature and humidity ranges (temperature < 15°C or > 35°C, humidity < 40% or > 90%), termite activity intensity significantly decreases, and metabolic heat signals are weak. Appropriately raise the single-modality confidence threshold to avoid misjudgment caused by environmental factors, ensuring the robustness of fusion recognition.

Experimental verification shows that the temperature-humidity fusion mechanism improves the stability of the temperature modality by over 12% in complex environments such as the southern plum rain season, coastal high humidity, and northern low temperatures. The temperature measurement error is controlled within ±0.2°C, providing more reliable temperature input for multi-modal fusion recognition.

3. Multi-Modal Data Synchronization and Preprocessing

Temporal synchronization and spatial registration of multi-modal data are prerequisites for efficient fusion. This device adopts a hardware timestamp synchronization scheme, configuring a unified clock module for the image, sound, and temperature sensors. This ensures that the acquisition time synchronization error of the three signal types is ≤10ms, achieving strict alignment of multi-source signals at the same time and in the same area. Simultaneously, a spatial registration design is adopted, placing the temperature and sound sensors on the same side as the visual acquisition area, ensuring that the three sensor types monitor the same spatial range, avoiding feature mismatch caused by spatial misalignment.

Due to the complexity of the underground environment, the collected multi-modal data contains significant noise interference (e.g., sediment occlusion in images, environmental noise in sounds, instantaneous jumps in temperature). A preprocessing pipeline is needed to improve data quality, laying the foundation for subsequent feature extraction and fusion recognition. The specific preprocessing steps are as follows.

① Image Data Preprocessing: First, apply a Gaussian filter to remove Gaussian noise from the image. Then, use histogram equalization to enhance image contrast, improving the distinction between termite targets and the background. For small termite targets in the underground environment, apply a small target enhancement algorithm to amplify and enhance features in tiny regions of the image, avoiding missed detection of small targets. Finally, use morphological operations (dilation, erosion) to remove fine impurities from the image while preserving the complete morphological features of the termite targets.

② Sound Data Preprocessing: Apply a bandpass filter to remove noise signals outside the 100-500Hz frequency band. Then, use wavelet denoising to remove high-frequency interference and impulse noise from the signal. Frame the denoised sound signal with a frame length of 256ms and a frame shift of 128ms. Extract core acoustic features such as short-term energy, zero-crossing rate, spectral entropy, and Mel-frequency cepstral coefficients (MFCCs) from each frame to form a sound feature vector. Finally, normalize the feature vector to map it to the same dimension, facilitating subsequent fusion calculations.

③ Temperature Data Preprocessing: Apply a moving average filter (window size 5) to remove instantaneous jump anomalies from the temperature data and smooth the temperature change curve. Calculate the temperature difference between adjacent sampling points to obtain the temperature rise rate feature. Interpolate the multi-point temperature data to construct a two-dimensional temperature field, extracting features like temperature gradient and thermal region range. Finally, normalize the temperature features to unify their dimensions with image and sound features, ensuring compatibility of the fusion algorithm.

Multi-Modal Data Fusion Algorithm and AI Model Optimization

1. "Feature-Level + Decision-Level" Two-Stage Fusion Architecture

To overcome the limitations of single-modality recognition and achieve efficient collaboration and precise fusion of multi-source data, this study designs a two-stage fusion architecture combining feature-level fusion and decision-level fusion. This enhances recognition accuracy and robustness from both the feature expression and decision-making levels, forming a complete recognition pipeline of "data acquisition - preprocessing - feature extraction - feature fusion - decision making".

(1) Feature-Level Fusion Algorithm

Feature-level fusion deeply integrates features extracted from the three modalities (image, sound, temperature) to form a more representative joint feature vector, compensating for the shortcomings of single-modality features and enhancing the feature expression capability for small termite targets. The specific implementation process is as follows.

① Extract core features from the three modalities respectively. Image features are extracted using the backbone network (CSPNet) of the optimized YOLOv10-M model, yielding a 256-dimensional deep convolutional feature vector containing core information like termite morphology and texture. Sound features are extracted via MFCCs, yielding a 128-dimensional acoustic feature vector containing information like the frequency and energy of termite activity sounds. Temperature features are extracted via temperature gradient and temperature rise rate, yielding a 64-dimensional temperature feature vector containing information like the variation pattern of termite metabolic heat.

② Combine feature concatenation and attention mechanism fusion to achieve deep fusion of the three feature types. Concatenate the image, sound, and temperature feature vectors to form a 448-dimensional initial joint feature vector. Introduce a channel attention mechanism (SE-Net) to assign weights to different channels of the joint feature vector, focusing on strengthening feature channels related to termite recognition (e.g., morphological feature channels in images, characteristic frequency band channels in sounds, temperature rise feature channels in temperature) while suppressing interference from irrelevant feature channels, enhancing the representational capacity of the joint feature.

③ Apply Principal Component Analysis (PCA) to reduce the dimensionality of the joint feature vector from 448 dimensions to 128 dimensions. This removes redundant features, reduces computational load, while retaining core recognition features, providing efficient feature input for subsequent decision-level fusion and AI model recognition.

(2) Decision-Level Fusion Algorithm

Decision-level fusion builds upon feature-level fusion by performing cross-validation and comprehensive decision-making on the recognition results of the three modalities, further improving recognition reliability and reducing false and missed alarm rates. This study designs a three-modal voting mechanism based on confidence thresholds. The specific decision logic is as follows.

① Single-Modality Recognition Confidence Calculation. Perform single-modality recognition for image, sound, and temperature respectively, calculating their respective recognition confidences. The image modality outputs image recognition confidence (range 0-1) via the optimized YOLOv10-M model. The sound modality outputs sound feature matching degree (range 0-1) via a Support Vector Machine (SVM) model. The temperature modality outputs thermal confidence (range 0-1) via a temperature anomaly judgment logic.

② Confidence Threshold Setting. Based on extensive laboratory test data and field validation data, determine the recognition confidence thresholds for the three modalities through statistical analysis: image recognition confidence ≥ 0.90, sound feature matching degree ≥ 0.85, thermal confidence ≥ 0.80 (corresponding to temperature anomaly lasting ≥ 30s and temperature rise ≥ 0.3°C). The threshold setting fully considers recognition robustness under different environments, avoiding both missed alarms due to excessively high thresholds and false alarms due to excessively low thresholds.

③ Voting Decision Rule. Adopt a "three conditions all met" voting rule. Only when the recognition confidences of all three modalities (image, sound, temperature) reach their corresponding thresholds is it determined as "termite activity present", triggering an alarm. If the confidence of any modality does not reach its threshold, it is determined as "no termite activity", considered an interference signal. Simultaneously, introduce a continuous detection trigger mechanism. The same device needs to detect termite activity three consecutive times (with a 30s interval between each detection) before formally initiating an alarm, further reducing false alarms caused by occasional interference.

Experimental verification shows that this decision-level fusion mechanism reduces the device's false alarm rate from over 8% for traditional single modalities to 0.92%, and the missed alarm rate to 0.3%, significantly improving recognition reliability.

2. Key Technologies for Temperature Modality Fusion

As a core component of the three-modal fusion, the recognition accuracy of the temperature modality directly affects the overall fusion effect. To address the issues of weak termite metabolic heat signals in underground environments and susceptibility to environmental interference, this study focuses on developing key technologies for temperature modality fusion, including micro-temperature rise discrimination logic, temperature field localization and nest prediction, and temperature anti-interference filtering, enhancing the recognition accuracy and anti-interference capability of the temperature modality.

(1) Micro-Temperature Rise Discrimination Logic

The micro-temperature rise signals (0.3-1.5°C) generated by termite colony activity are extremely weak and easily confused with natural soil thermal disturbances. Therefore, a multi-point collaborative micro-temperature rise discrimination logic is designed. The specific process is: ① Single-point discrimination: If the temperature difference of a single temperature measurement point over three consecutive samples (sampling interval 10s) is ≥ 0.3°C and lasts for ≥ 30s, mark this point as a "thermal anomaly point". ② Multi-point collaborative discrimination: If three adjacent temperature measurement points are simultaneously marked as "thermal anomaly points" and the thermal anomaly area range is ≥ 5cm × 5cm, determine it as "termite colony aggregation thermal anomaly", excluding misjudgment caused by a single point failure or local soil thermal disturbance. ③ Temperature-humidity correction: Correct the thermal confidence based on the simultaneously collected humidity data and the temperature-humidity-termite activity correlation model. If within the high termite activity temperature-humidity zone, appropriately increase the thermal confidence; otherwise, decrease it, further improving discrimination accuracy.

(2) Temperature Field Localization and Nest Prediction

Based on multi-point temperature data, construct a two-dimensional temperature field model to locate the termite aggregation center and predict the nest position, providing precise location guidance for on-site treatment. The specific implementation methods are as follows: ① Temperature field construction: Interpolate the temperature data from the three measurement points using the Kriging interpolation algorithm to construct a two-dimensional temperature field within a 1m range around the device, visually presenting the temperature distribution pattern. ② Aggregation center localization: Identify the high-temperature region in the temperature field and calculate its geometric center. This center is the termite aggregation center, with a localization error ≤ 0.5m. ③ Nest prediction: Based on termite nesting habits, temperature field characteristics, and historical data, establish a nest position prediction model. Using parameters like the temperature gradient of the high-temperature region, duration of thermal anomaly, and range size, predict the depth and extent of the nest (with nest depth prediction error ≤ 0.3m), providing precise guidance for on-site excavation treatment and reducing ineffective excavation.

(3) Temperature Anti-Interference Filtering

Temperature interference in underground environments mainly includes solar radiation heat, soil heat conduction, and water evaporation endothermic cooling. These interference signals differ significantly from termite metabolic heat signals. Through feature analysis and algorithmic filtering, interference signals can be effectively distinguished from valid signals. Specific filtering strategies are as follows: ① Solar radiation heat filtering: Temperature changes caused by solar radiation are characterized by large area, gradual change, and synchronicity, with no obvious local high-temperature center in the temperature field. By monitoring the spatial distribution and rate of temperature change, this type of interference can be effectively filtered. ② Soil heat conduction filtering: Temperature changes caused by soil heat conduction are characterized by slowness, global scope, and no abrupt changes, with a small temperature gradient. By calculating the rate and gradient of temperature change, soil heat conduction can be distinguished from termite metabolic heat. ③ Water evaporation endothermic filtering: Temperature changes caused by water evaporation are characterized by instantaneous, local, and cooling effects, opposite to the warming characteristics of termite metabolic heat. By monitoring the temperature change trend (warming/cooling), this type of interference can be effectively filtered. Experimental verification shows that the temperature anti-interference filtering accuracy reaches 99.2%, effectively avoiding misjudgment caused by environmental temperature interference.

3. YOLOv10-M Optimization Based on Multi-Modal Constraints

YOLOv10-M, as a lightweight object detection model, offers advantages like fast detection speed, high accuracy, and few parameters, making it suitable for deployment on edge devices. However, the model still has shortcomings in small termite target recognition and anti-interference in complex environments. Therefore, targeted optimization of the YOLOv10-M model is carried out based on multi-modal fusion requirements to enhance the model's recognition capability and robustness for small termite targets. Specific optimization measures are as follows.

① Add a small object detection head. For small termite targets of 3-5mm, add a new small object detection head (output feature map size 1024×1024) to the YOLOv10-M model to strengthen the extraction and recognition capability for small target features, improving the small target recall rate. Simultaneously, adjust the anchor box sizes of the detection head. Based on the actual size of termites (3-5mm), design three sets of anchor boxes (4×4, 5×5, 6×6) to match the morphological features of small termite targets, reducing missed detections caused by anchor box mismatch.

② Introduce a temperature feature gating mechanism. Integrate the temperature feature vector into the model's feature fusion layer via a gating unit, achieving synergy between temperature and image features. When the temperature feature detects a thermal anomaly, the gating unit opens, strengthening the recognition weight of termite targets in the image features. When the temperature feature shows no anomaly, the gating unit closes, reducing the recognition weight of image features, decreasing misjudgment of non-termite targets, and enhancing the model's anti-interference capability.

③ Integrate sound energy features. Integrate the short-term energy feature from the sound features into the model's classification head, working synergistically with image features for classification judgment. When the sound short-term energy is within the characteristic frequency band of termite activity (100-500Hz) and the energy value reaches the threshold, increase the model's classification confidence for termite targets; otherwise, decrease the classification confidence, further suppressing misjudgment caused by environmental noise.

④ Model lightweight compression. Apply knowledge distillation technology, using YOLOv10-L as the teacher model and YOLOv10-M as the student model. Transfer the knowledge from the teacher model to the student model. While ensuring recognition accuracy, compress the model parameters by 60%, increase inference speed by 40%, and achieve single-frame inference time ≤ 20ms, meeting the real-time monitoring requirements of edge devices while reducing device power consumption and extending battery life.

Laboratory tests show that the optimized YOLOv10-M model achieves a recognition rate of 99.95% for small termite targets, an improvement of 7.65% over the original model. The missed alarm rate is reduced to 0.3%, and the false alarm rate to 0.92%, meeting the real-time and accuracy requirements for termite monitoring in water conservancy projects.

4. Termite Species and Caste Identification Technology

Different termite species and castes pose varying degrees of hazard to embankments. Accurately identifying the species and caste can provide a scientific basis for targeted control, improving control effectiveness. Based on multi-modal feature differences, this study constructs a termite species and caste identification model to identify five major harmful species (Odontotermes formosanus, Macrotermes barneyi, Coptotermes formosanus, Reticulitermes spp., and Macrotermes annandalei) and three castes (workers, soldiers, alates) found in China's water conservancy projects.

(1) Species Identification

Significant differences exist in the morphology, sound, and metabolic heat characteristics of different species: Odontotermes formosanus workers are 3-4mm long, dark brown, with feeding sound frequencies concentrated at 200-300Hz and metabolic heat temperature rise of 0.5-0.8°C. Macrotermes barneyi workers are 4-5mm long, light yellow, with feeding sound frequencies concentrated at 150-250Hz and metabolic heat temperature rise of 0.8-1.2°C. Coptotermes formosanus workers are 3-5mm long, milky white, with feeding sound frequencies concentrated at 250-350Hz and metabolic heat temperature rise of 0.6-0.9°C. Reticulitermes spp. workers are 2-3mm long, grayish-white, with feeding sound frequencies concentrated at 100-200Hz and metabolic heat temperature rise of 0.3-0.5°C. Macrotermes annandalei workers are 5-6mm long, yellowish-brown, with feeding sound frequencies concentrated at 300-400Hz and metabolic heat temperature rise of 1.0-1.5°C.

Based on these feature differences, a multi-modal species identification model is constructed: fuse image morphological features, sound spectrum features, and temperature rise features, input them into an SVM classifier, and train with a large number of samples to achieve accurate identification of the five species. Tests show that the species identification accuracy is > 98%, with the identification accuracy for Odontotermes formosanus and Macrotermes barneyi reaching over 99%.

(2) Caste Identification

Different castes (workers, soldiers, alates) of the same species exhibit significant differences in morphology, behavior, and metabolic characteristics: Workers are smaller, primarily responsible for feeding and nest building, active frequently, with weaker metabolic heat. Soldiers are larger, with developed heads, aggressive, less active, with moderate metabolic heat. Alates are the largest, have wings, are active during the reproductive season, and have stronger metabolic heat.

Based on these differences, further achieve caste identification on the basis of species identification: Distinguish body size and morphological differences through image features, distinguish activity frequency and acoustic signal differences through sound features, and distinguish metabolic heat differences through temperature features. Construct a multi-modal caste identification sub-model to achieve accurate differentiation of workers, soldiers, and alates, with identification accuracy > 98%. The identification accuracy for soldiers reaches 99.2%, providing precise guidance for targeted control (e.g., deploying specific agents for soldiers).

Experimental Testing and Field Application Validation in Water Conservancy Projects

1. Laboratory Performance Testing

To comprehensively verify the device's core indicators such as multi-modal recognition performance, environmental adaptability, and battery life, systematic tests were conducted in a professional termite laboratory. The test environment simulated the underground concealed environment of water conservancy projects. Termite sample groups and interference groups were set up. The recognition performance of single modalities and three-modal fusion was compared. The device's environmental adaptability and battery life were also tested to ensure it meets practical engineering requirements.

(1) Recognition Performance Test

Samples of workers, soldiers, and alates from the five major harmful species mentioned above were selected, with 100 individuals per species per caste, totaling 1500 termite samples. Interference samples included common underground organisms like ants, cockroaches, and earthworms, as well as environmental interference signals like soil vibration, rainwater scouring, and water flow sounds, totaling 500 interference sample groups. Termite samples and interference samples were placed separately into test boxes simulating the underground environment. The monitoring device was deployed for continuous monitoring over 72 hours. The device's recognition results and alarm situations were recorded. Core indicators such as recognition rate, alarm accuracy rate, false alarm rate, and missed alarm rate were calculated. Simultaneously, the recognition performance of single vision, single sound, and single temperature modalities was tested separately and compared with the three-modal fusion recognition performance.

Test results show that under three-modal fusion recognition, the comprehensive termite recognition rate is 99.96%, with specific rates: Odontotermes formosanus 99.98%, Macrotermes barneyi 99.97%, Coptotermes formosanus 99.96%, Reticulitermes spp. 99.95%, Macrotermes annandalei 99.94%. The alarm accuracy rate is 99.08%, false alarm rate 0.92%, and missed alarm rate 0.3%. For single modalities: vision recognition rate 92.3%, false alarm rate 7.8%, missed alarm rate 7.7%; sound recognition rate 89.5%, false alarm rate 9.1%, missed alarm rate 10.5%; temperature recognition rate 85.7%, false alarm rate 6.3%, missed alarm rate 12.4%.

In summary, the three-modal fusion recognition performance is significantly superior to single modalities, effectively solving the problems of low recognition accuracy and high false/missed alarm rates associated with single modalities. It enables accurate recognition of small termite targets, meeting the precision requirements for termite monitoring in water conservancy projects.

(2) Environmental Adaptability Test

Given the complex and variable environment of water conservancy projects, environmental adaptability tests including high/low temperature tests, waterproof tests, salt spray tests, and high humidity tests were conducted to verify the device's stable operation capability under different environments.

① High/Low Temperature Test. The device was placed in constant temperature chambers at -20°C, -10°C, 0°C, 25°C, and 55°C, running continuously for 72 hours. The device's operating status and recognition performance were recorded every 12 hours. Test results show that the device operates normally within the range of -20°C to 55°C, with a recognition rate consistently above 99.5% and no failures, meeting the usage requirements of different climate regions nationwide.

② Waterproof Test. The device was immersed in a test box with 1.5m water depth for 30 minutes. After removal, the interior was checked for water ingress, and the device's operating status and recognition performance were tested. Test results show no water ingress inside the device, normal operation, recognition rate maintained above 99.9%, achieving IP68 protection rating, meeting usage requirements in environments with floodwater accumulation and heavy rain.

③ Salt Spray Test. The device was placed in a salt spray chamber with 5% sodium chloride solution for 48 hours. After the test, the corrosion condition of the device's shell and metal parts, as well as its operating status and recognition performance, were checked. Test results show no significant corrosion on the shell and metal parts, normal operation, recognition rate maintained above 99.5%, meeting usage requirements in coastal salt spray environments.

④ High Humidity Test. The device was placed in a constant temperature and humidity chamber with 95% relative humidity, running continuously for 72 hours. The device's operating status and recognition performance were tested. Test results show normal operation, recognition rate maintained above 99.8%, the temperature-humidity fusion compensation mechanism effective, with no temperature measurement drift or misjudgment, meeting usage requirements in southern high-humidity environments.

(3) Battery Life Performance Test

The device uses a 38Ah high-capacity lithium thionyl chloride battery and low-power design. The battery life test method simulates actual working scenarios: the device operates according to a strategy of "waking up every 30 minutes at night, working for 20 seconds each time; waking up every 1 hour during the day, working for 20 seconds each time". Battery power consumption is recorded. Test results show that the average power consumption of the device is 0.012mW, and the battery life reaches 4.8 years, meeting the engineering requirement of "one-time burial, four-year maintenance-free", eliminating the need for frequent battery replacement and reducing damage to the embankment structure.

2. Field Application Validation in Water Conservancy Projects

To verify the device's performance and practicality in actual water conservancy project scenarios, field application validation was conducted at 182 water conservancy projects nationwide, covering key river basins such as the Yangtze River, Pearl River, Huaihe River, and Yellow River. The projects included various types like reservoirs, embankments, and sluice gates, spanning diverse typical environments such as southern high temperature and humidity, central hilly areas, northern low temperatures, and coastal salt spray. A total of over 2,300 devices were deployed, monitoring an area exceeding 5 million m². Traditional monitoring points (manual inspection + bait trapping) were set up simultaneously as a control to compare and analyze the application effectiveness of the device.

(1) Application Scenario Classification and Deployment Plan

Based on the type of water conservancy project and termite damage characteristics, application scenarios were classified into four categories, using a standardized deployment plan.

① Reservoir Dam Scenario. Mainly deployed in high termite activity areas such as the downstream slope, berm, sides of the spillway, and abutments of the dam. A quincunx layout pattern is adopted with a device spacing of 15m, focusing on monitoring termite activity inside the dam body to prevent piping and seepage hazards.

② Embankment Project Scenario. Mainly deployed in areas such as the downstream slope, toe, and floodplain of the embankment. A linear layout pattern is adopted with a device spacing of 20m, covering the entire embankment line to prevent seepage and landslide risks caused by termite nesting.

③ Sluice Gate Project Scenario. Mainly deployed in the soil around the sluice gate and on both sides of the diversion/drainage channels. Device spacing is 10m, focusing on monitoring termite activity in the foundation area of the sluice gate to prevent hazards like seepage and foundation settlement.

④ Old Embankment Scenario. Given the high termite infestation rate and weak structure of old embankments, devices are deployed more densely with a spacing of 10m, focusing on monitoring dangerous sections and weak points to achieve early detection and treatment of termite hazards.

(2) Field Application Results

After 12 months of field application validation, all performance indicators of the device met or exceeded design requirements, showing significant advantages over traditional monitoring methods. The specific application results are summarized as follows.

① Recognition and Alarm Performance. A total of 127 termite activities were monitored, successfully providing early warnings for 109 termite hazards, with a 100% hazard treatment accuracy rate. The alarm response time was ≤ 20s, compared to 7-15 days for traditional monitoring methods. The device achieved a recognition rate of 99.95%, alarm accuracy rate of 98.5%, false alarm rate of 1.4%, and missed alarm rate of 0.3%, all superior to industry standard requirements (alarm accuracy rate ≥ 95%, false alarm rate ≤ 5%, missed alarm rate ≤ 5%).

② Operational Efficiency and Cost. The average annual operation and maintenance cost for traditional manual inspection of 1km of embankment is approximately 20,000 RMB. Using this device, the annual operation and maintenance cost per km of embankment can be reduced by 14,000 RMB, a 70% reduction. A single person can install over 30 devices per day, improving efficiency by more than 15 times compared to the traditional bait trapping method (2km installation per person per day). The device operates maintenance-free for 4.8 years, eliminating the need for frequent embankment excavation for battery replacement, reducing damage to the embankment structure and lowering maintenance workload.

③ Hazard Treatment Effect. All 109 termite hazards were treated promptly. Among them, 87 were early-stage scattered termite activities, and 22 were small to medium-sized nests. Follow-up monitoring after treatment showed no recurrence of termites, effectively preventing further development of termite damage, avoiding major hazards like piping and seepage, and ensuring the safe operation of water conservancy projects.

④ Adaptability to Different Scenarios. In southern high-temperature and high-humidity scenarios (e.g., Huangtian Reservoir, Guangdong), the device maintained a recognition rate above 99.8%. In northern low-temperature scenarios (e.g., Yellow River Embankment, Shandong), the device operated normally at -20°C with a recognition rate above 99.5%. In coastal salt spray scenarios (e.g., Sanhe Sluice, Jiangsu), the device operated normally after salt spray erosion with a recognition rate above 99.6%. In old embankment scenarios (e.g., Caixian Reservoir, Hubei), the device successfully provided early warnings for 17 termite hazards with a 100% treatment accuracy rate, effectively solving the termite monitoring challenges in old embankments.

Conclusion and Outlook

1. Conclusion

Focusing on termite monitoring in water conservancy projects, with the core technology of "vision-sound-temperature" three-modal data fusion, this study conducted research, development, experimental testing, and field application validation of an underground passive termite monitoring device. Key tasks including multi-modal sensing system construction, fusion algorithm design, AI model optimization, termite species and caste identification, and environmental adaptability optimization were completed, achieving the following core results.

① Constructed a high-precision three-modal collaborative sensing system, achieving synchronous acquisition and preprocessing of termite morphology, weak activity sounds, and metabolic micro-temperature rise signals. Combined with a temperature-humidity fusion compensation mechanism, it improved the reliability of multi-modal data in complex environments, solving the problem of capturing weak termite signals in concealed underground environments.

② Designed a "feature-level + decision-level" two-stage fusion architecture. Introduced an attention mechanism and a three-modal voting decision rule, achieving deep fusion and precise decision-making of multi-source features. Combined with the optimized YOLOv10-M model, the device achieved a comprehensive termite recognition rate of 99.95%, alarm accuracy rate of 98.5%, false alarm rate reduced to below 1.5%, and missed alarm rate reduced to 0.3%, outperforming single-modality recognition and traditional monitoring methods, solving the core pain point of high false/missed alarm rates in traditional monitoring.

③ Broke through key technologies for temperature modality fusion, including micro-temperature rise discrimination, temperature field localization and nest prediction, and temperature anti-interference filtering, achieving precise localization of the termite aggregation center and nest position prediction with a localization error ≤ 0.5m. Simultaneously, constructed a termite species and caste identification model, achieving accurate identification of five major harmful species and their three castes with an identification accuracy > 98%, providing a scientific basis for targeted control.

④ The device features strong environmental adaptability and long battery life, capable of stable operation in complex environments such as wide temperature range (-20 to 55°C), water immersion, salt spray, and high humidity, with a battery life of 4.8 years.

⑤ Field application validation at 182 water conservancy projects nationwide showed that the device operates stably in different types and environments of water conservancy project scenarios. It successfully provided early warnings for 109 termite hazards with a 100% treatment accuracy rate. The response speed is superior to traditional monitoring methods. The missed alarm rate decreased by 97%, the false alarm rate decreased by 88%, and operation and maintenance costs were reduced by 70%. It effectively prevented water conservancy project hazards caused by termite damage, achieving the transformation of termite monitoring from "manual inspection, post-event treatment" to "intelligent monitoring, pre-event warning".

2. Outlook

Although this study has achieved significant results, considering the actual needs of termite control in water conservancy projects and technological development trends, further optimization and improvement are still needed in the future.

① Further optimize the multi-modal fusion algorithm by introducing deep learning fusion models (e.g., Transformer fusion models) to enhance the fusion depth and recognition robustness of multi-source features. Further optimize anti-interference strategies for extreme environments (e.g., heavy rainfall, strong vibration, extremely low temperatures) to reduce the false alarm rate to below 1% and the missed alarm rate to below 0.1%.

② Expand recognition functions. Based on the existing species and caste identification, add functions for nest size assessment and hazard level classification. Through multi-modal feature analysis, predict the size, depth, and hazard level of termite nests, providing a more accurate basis for control decisions. Simultaneously, expand the range of multi-species identification to cover all termite species harmful to water conservancy projects in China.

③ Optimize the device hardware design to further reduce power consumption and extend battery life to over 5 years. Simultaneously, optimize the device structure to enhance resistance to soil pressure and external impact, extending the device's service life. Develop miniaturized and lightweight versions suitable for deployment in small water conservancy projects and complex terrain.

④ Promote technology standardization and large-scale application. Based on the research results, take the lead in formulating industry standards for intelligent termite monitoring in water conservancy projects, standardizing processes for device development, deployment, operation, maintenance, and testing. Expand the application scope to cover various types of water conservancy projects, especially old embankments and dangerous sections, to build a national termite monitoring network for water conservancy projects, achieving comprehensive monitoring and intelligent prevention and control of termite hazards.

⑤ Strengthen integration with digital twin water conservancy technology. Integrate the termite activity data, environmental data, and embankment safety data collected by the device into a digital twin water conservancy platform. Construct a digital twin model for termite damage, enabling dynamic simulation, risk prediction, and intelligent treatment of termite activity. This will provide more comprehensive data support for the safe operation and maintenance of water conservancy projects, promoting the transformation of termite control in water conservancy projects towards "intelligent precision control, proactive prevention and control".