Briefing: Automated Insect Monitoring via AI and Electrical Field Sensors
Source: Odgaard, F.B., Kjærbo, P.V., Poorjam, A.H.
et al. Automated insect detection and biomass monitoring via AI and electrical field sensor technology.
Sci Rep 15, 29858 (2025).
https://doi.org/10.1038/s41598-025-15613-5Date: Received - 11 April 2025 | Accepted - 08 August 2025 | Published - 14 August 2025
Executive Summary
This document outlines a novel, automated insect monitoring system that uses electrical field sensors and artificial intelligence to provide a non-invasive, continuous alternative to traditional methods. The system addresses the critical need for improved insect monitoring in the face of global declines, aiming to overcome the labor-intensive, lethal, and temporally limited nature of conventional techniques like Malaise traps.
The core technology detects atmospheric electrical field modulations caused by flying insects. A differential sensor design suppresses environmental noise, while a cloud-based AI pipeline processes the signals. This pipeline employs a Convolutional Neural Network (CNN) for insect detection, a probabilistic algorithm for Wing-Beat Frequency (WBF) analysis, and a lookup-based algorithm for biomass estimation.
A field validation study conducted in a Danish nature reserve compared the system against standard Townes Malaise traps. The results demonstrated a moderate to strong positive correlation between sensor and trap data for insect counts (Spearman’s ρ up to 0.725). However, the correlation for biomass was weaker and not consistently significant. A major discrepancy in magnitude was observed, with sensors recording approximately three times more insect counts and 26 times more biomass than the traps. This is attributed to fundamental methodological differences (passive sensing vs. single capture) and significant uncertainty within the system's current biomass estimation algorithm.
Notably, the sensor system exhibited higher measurement consistency between its own units (sensor-sensor correlation for biomass ρ = 0.867) than paired Malaise traps (Malaise-Malaise correlation for biomass ρ = 0.641), although this difference was not statistically significant (P = 0.057). The study concludes that while the technology shows significant promise for scalable, non-lethal insect monitoring, the biomass algorithm requires substantial refinement and calibration before it can be used for absolute estimation.
1. The Challenge in Conventional Insect Monitoring
Insects, comprising over half of all described species, are vital for ecosystem stability through functions like pollination, nutrient cycling, and pest control. Alarming reports of declines in insect abundance, biomass, and species richness underscore the urgent need for effective monitoring to support conservation and safeguard ecosystem services.
However, conventional monitoring techniques present significant challenges:
• Labor-Intensive: Methods such as pan, pit, light, and Malaise traps require substantial manual effort for insect collection, sorting, counting, and weighing.
• Invasive and Lethal: These trap-based approaches remove insects from the local population, posing a potential threat to fragile species and raising ethical concerns. The validation study for this new system highlighted this impact, with 55,443 insects killed in just two Malaise traps during the sampling period.
• Limited Granularity: Traditional methods typically provide data at coarse temporal intervals (e.g., daily or weekly), limiting insights into finer-scale activity patterns.
Automation and non-invasive technologies are critical for overcoming these limitations, enabling continuous data collection across large areas without disrupting local ecosystems.
2. A Novel Automated Monitoring System
The presented system offers a comprehensive, automated solution for non-invasive insect monitoring, from data acquisition in the field to data analysis in the cloud.
2.1. Operating Principle and Sensor Design
The system's core innovation is its ability to passively detect flying insects by exploiting natural electrical effects.
• Detection Mechanism: As insects fly, they acquire a positive electrical charge through air friction (triboelectric effect) and disrupt the ambient atmospheric electric field. These combined effects create unique electrical signatures that the sensor detects.
• Differential Probe Design: To function in noisy outdoor environments, the sensor employs two identical electrostatic probes spaced 28 cm apart. This differential measurement approach effectively mitigates distant, common-mode noise sources like atmospheric disturbances and radio signals.
• Detection Volume: The design creates a detection volume sensitive to nearby insects. However, it also creates a "blind plane" of zero sensitivity on the symmetry plane directly between the two probes. The sensor's sensitivity is size-dependent, meaning larger insects are detectable at greater distances than smaller insects.
2.2. System Architecture and Data Pipeline
The system is composed of three integrated components:
1. Field Sensor Units: The core sensor, housed in a weatherproof unit, uses an ESP32 microcontroller to acquire signals, perform real-time preprocessing, and transmit data via cellular communication. The sensors are solar-powered for continuous daylight operation.
2. Cloud Processing Infrastructure: Data is sent to a cloud-based pipeline that performs a series of processing steps:
◦ Removes power line interference (50/60 Hz) using a specialized comb filter.
◦ Detects the presence of flying insects using an AI model.
◦ Calculates the Wing-Beat Frequency (WBF) of detected insects.
◦ Estimates the body mass of the insects.
3. User Interface: Processed data on insect activity (counts) and biomass is aggregated and made available through a user interface for analysis and export.
2.3. AI-Powered Data Processing
The analytical power of the system resides in its sophisticated data processing algorithms.
• Insect Detection (CNN): A Convolutional Neural Network (CNN) is used to classify 1-second signal segments. Each segment is converted into a spectrogram (a visual representation of frequency over time), which serves as the input to the CNN. The model was trained on a large, manually annotated dataset and demonstrated high classification performance on a held-out test set:
◦ AUC (Area Under Curve): 0.96
◦ F1-Score: 0.79
◦ Precision: 0.77
◦ Recall: 0.81
• WBF Calculation: For segments classified as containing an insect, the probabilistic YIN (pYIN) algorithm estimates the fundamental frequency, or WBF. A post-processing step filters out unreliable signals (e.g., those with a WBF below 20 Hz or with drastic frequency changes) to reduce false positives. Adjacent 1-second segments with similar WBFs are aggregated to represent a single, continuous insect event.
• Biomass Estimation: Biomass is estimated using a lookup-based algorithm that maps the calculated WBF to body mass via a reference table derived from scientific literature. This method was chosen because a direct statistical test showed only a weak negative correlation between WBF and body mass (Pearson's ρ = -0.32). The algorithm handles ambiguity by:
◦ Defining WBF regions (dense, less dense, sparse) based on the number of potential insect taxa at a given frequency.
◦ Filtering candidates based on their presence in Europe.
◦ Averaging mass for multiple remaining candidates (though observation probabilities are currently placeholders).
◦ Assigning a median mass value if no match is found in dense regions.
3. Field Validation and Comparative Analysis
The system was validated against the widely used Townes Malaise trap to assess its real-world performance.
3.1. Experimental Design
• Location: A nature reserve in the Capital Region of Denmark.
• Setup: Two sites were established, each containing one electrical field sensor and one Malaise trap deployed in close proximity.
• Duration: The sampling period ran from June 17, 2024, to September 12, 2024.
• Analysis: Due to non-normal data distribution, Spearman's correlation tests were used to compare insect counts and biomass between the methods.
3.2. Key Findings: Sensor vs. Malaise Trap
The comparative analysis revealed moderate-to-strong correlations for insect counts but highlighted significant challenges in biomass estimation.
Comparison Metric | Site 1 Correlation (ρ) | Site 1 P-value | Site 2 Correlation (ρ) | Site 2 P-value
Insect Counts | 0.725 | < 0.0001 | 0.569 | 0.0029
Biomass | 0.644 | 0.00051 | 0.332 | 0.098 (NS)
• Magnitude Discrepancies: The sensors recorded significantly higher values than the traps.
◦ Counts: OLS slopes indicate sensors recorded approximately 3 times higher insect counts.
◦ Biomass: Sensor estimates were roughly 26 times higher than trap measurements.
• Attribution of Discrepancies: These differences are attributed to several factors:
1. Methodology: The sensor is a passive, non-invasive device that may detect the same insect multiple times as it passes through the detection area. The Malaise trap, by contrast, captures an insect only once.
2. Biomass Algorithm Uncertainty: The large discrepancy in biomass is compounded by ambiguity in the WBF-to-biomass mapping, especially in lower frequency ranges where many species have similar WBFs.
3. Trap Biases: Malaise traps are known to underestimate certain taxa, such as heavy-bodied beetles (Coleoptera), which may lead to an underestimation of total biomass in the trap samples.
3.3. Key Findings: Inter-Method Consistency
The study also compared the consistency of measurements between the two sensors and the two Malaise traps.
Comparison Type | Spearman's ρ | Statistical Test (vs. Malaise-Malaise)
Sensor-Sensor Counts | 0.758 | Z-score = 1.029 (P = 0.304)
Malaise-Malaise Counts | 0.597 |
Sensor-Sensor Biomass | 0.867 | Z-score = 1.907 (P = 0.057)
Malaise-Malaise Biomass | 0.641 |
The sensor system demonstrated higher correlation between its units for both counts and biomass, suggesting greater measurement consistency. While this difference was not statistically significant at the α = 0.05 level, the result for biomass (P = 0.057) approached significance and warrants further investigation.
4. Limitations, Future Directions, and Conclusion
4.1. Identified Limitations
• Biomass Algorithm: The primary limitation is the biomass estimation algorithm. Its reliance on a literature-based lookup table with inherent ambiguities and uncertainties (e.g., whether source weights were dry or wet) necessitates further calibration before it can provide reliable absolute biomass estimates.
• Geographic Bias: The CNN model was trained and evaluated primarily on data from Northern Europe, which may limit its generalization to other ecological contexts with different insect fauna.
• Noise Susceptibility: Despite the differential design, power line noise from nearby sources can still present a challenge that requires specialized digital filtering.
4.2. Proposed Future Enhancements
• Refine Biomass Estimation: Future work should focus on improving the biomass algorithm. This could involve developing a predictive model that utilizes the negative correlation between WBF and body mass, potentially in combination with the existing lookup-based method.
• Expand Training Data: The geographic scope of the CNN's training data must be expanded to ensure robust performance across diverse habitats and regions.
• Explore Advanced AI: An alternative approach using object detection frameworks on spectrograms could be explored. Such models could simultaneously detect and localize insect signals, potentially improving resilience to noise and overlapping events.
4.3. Conclusion
This study successfully demonstrates a sensor-based approach for automated, continuous, and non-lethal insect monitoring. The system provides high-resolution temporal data and shows moderate-to-strong correlation with traditional methods for insect activity counts. By mitigating the labor, ethical concerns, and destructive nature of lethal sampling, this technology represents a promising advancement for ecological research. While the biomass estimation component requires significant refinement, the overall system offers a scalable and powerful new tool for understanding and conserving insect populations.