Source-linked AI summary
Effectiveness of IoT and Deep Learning for Detection and Severity Assessment of Postelectrotermes militaris in Tea Plantations
D. K. C. Senevirathna, A. A. E. Nanayakkara, H. M. C. K. Kulathunga, J. K. D. P. Nadula, R. M. Mapatuna, Malithi Nawarathne, Jaliya L. Wijayaraja, P. D. Senanayake, Samitha Vidhanaarachchi, Kalpani Manathunga
TL;DR
ULWT can damage tea plants internally before visible symptoms appear, creating a need for scalable, noninvasive monitoring. This study develops an IoT acoustic-sensing framework with CNN classification, severity estimation, and geospatial tracking; on a held-out test set, it achieved 81.5% accuracy and 83.0% recall, while remaining a screening and decision-support tool.
Problem
ULWT infestations can cause internal damage before visible symptoms, while existing practices lack systematic early-stage monitoring, predictive capability, and spatial tracking.
Method
The framework collects trunk acoustics with an IoT device, analyzes Fourier-derived spectrograms using a CNN, and combines infestation probability, amplitude, and nearby infested plants for severity estimation.
Results
81.5% accuracy, 80.6% precision, 83.0% recall, 81.8% F1-score, and 0.819 AUC were achieved on the held-out test set under field conditions.
Takeaways & Limitations
Severity and geospatial outputs can support higher-risk-area identification, prioritization of confirmatory inspections, and targeted field interventions.
Takeaways & Limitations
The study used relatively few independent field recordings and a single train-validation-test partition, so segmented clips do not represent independent biological samples.
Abstract
from arXiv · showhide
Tea plantations are vulnerable to Postelectrotermes militaris, commonly known as the Upcountry Live Wood Termite (ULWT), which can cause substantial damage when infestations remain undetected. This study proposes an IoT-enabled acoustic monitoring framework integrated with deep learning for early detection and severity assessment of ULWT infestations in tea plantations. Research Method: Audio signals were captured non-invasively from tea trunks using a high-sensitivity microphone connected to a Raspberry Pi-based IoT device, with geographic coordinates recorded for spatial tracking. After trimming, resampling, and segmentation, 2,000 ten-second samples were obtained, comprising 1,000 healthy and 1,000 infested samples, and divided into 1,600 training, 200 validation, and 200 test samples. The dataset used in this study is publicly available on Kaggle (Senevirathna et al. 2026). Fourier-derived spectrograms trained a CNN for infestation classification and probability estimation. A weighted severity model combined CNN probability, mean acoustic amplitude, and nearby infested plants within 5 m, with geospatial mapping used to visualize infestation distribution. Findings and Values: Field trials in a ULWT-affected tea plantation in Pundaluoya demonstrated feasibility under realistic environmental noise. On the held-out test set, the CNN achieved 81.5% accuracy, 80.6% precision, 83.0% recall, 81.8% F1-score, and 0.819 ROC-AUC. Beyond binary infestation detection, the framework introduced quantitative severity assessment using infestation probability, acoustic amplitude, and nearby infested plants. The resulting severity and geospatial outputs can support plantation managers in identifying high-risk areas, prioritizing field inspections, and implementing more timely and targeted control measures.
1. Introduction
ULWT can cause serious internal damage while tea bushes appear healthy, making visual inspection and current management practices insufficient for systematic early detection and spatial tracking.
- ULWT colonizes living woody tissues, forming heartwood and internal-stem galleries that progressively weaken tea plants.
- Visually healthy canopies may conceal substantial internal deterioration, limiting visual inspection, surface treatments, and natural predation.
- Removing and burning infested bushes, followed by Mana grass cultivation before replanting, increases costs and can create a three- to four-year income gap.
- Current management relies on labour-intensive, expertise-dependent inspections and lacks systematic early monitoring, predictive capability, and spatial tracking.
- Acoustic sensing is motivated as a non-destructive approach for detecting concealed termite activity where internal damage develops before visible symptoms.
- The proposed framework combines trunk-level acoustic acquisition, deep-learning classification, severity estimation, user notification, and geospatial visualization.
2. Methodology
The methodology collects acoustic signals from tea trunks with an IoT device, converts them into Fourier-based spectrograms for CNN classification, and combines model and spatial indicators to estimate severity.
- The end-to-end system integrates trunk acoustic collection, signal preprocessing, CNN classification, severity estimation, user notification, and location-based visualization.
- Data Collection: Recordings were collected from healthy and infested trunks in realistic plantation environments using a high-sensitivity microphone and Raspberry Pi-based IoT device.
- Severity Estimation: The CNN generated infestation probabilities, while severity combined that probability, acoustic amplitude, and nearby infested plants within a 5 m radius.
- Preprocessing: Preprocessing trimmed recording edges, resampled audio uniformly, segmented recordings into non-overlapping 10-second clips, and preserved low-frequency termite-related characteristics.
- Data Collection: The balanced dataset was partitioned before segmentation so clips from the same original recording remained within their assigned training, validation, or test subset.
- Classification Model Selection and Training: Fourier-based spectrograms represented time-varying frequency content, and a CNN was selected to learn localized hierarchical patterns from these image-like inputs.
- IoT Device: The IoT device used a Raspberry Pi controller, microphone, trunk interface, cloud connectivity, and mobile-application support for field data collection and transmission.
3. Results and Discussion
Field deployment showed that the IoT device could capture ULWT-related acoustic signals under plantation conditions, while the CNN distinguished infested from healthy samples and supported severity and spatial monitoring. Test performance was moderate, with field variability, limited independent recordings, and a training–validation gap defining important boundaries for interpretation.
- IoT device: The trunk-fitting IoT device captured acoustic recordings in ULWT-affected plantation environments, reduced surrounding noise, and enabled cloud-based data transfer.Field deployment demonstrated suitability for acoustic monitoring under realistic plantation conditions.
- Deep learning classification: The CNN achieved 81.5% accuracy on the held-out test set using Fourier-derived spectrogram representations.Training accuracy reached approximately 95%, while validation accuracy stabilised at approximately 81% after epoch 30.
- Deep learning classification: The model achieved an AUC of 0.819, with a 0.20 false-positive rate and 0.83 true-positive rate at the selected threshold.These results indicate discrimination across classification thresholds, but the operating point still produced false positives and false negatives.
- Deep learning classification: The infested class achieved 80.6% precision, 83.0% recall, and 0.818 F1-score, with 83 infested and 80 healthy samples correctly classified.The confusion matrix recorded 20 healthy samples misclassified as infested and 17 infested samples misclassified as healthy.
- Limitations: The study’s evidence is bounded by few independent field recordings, unmeasured environmental parameters, and a training–validation accuracy gap suggesting incomplete transfer to unseen samples.Segmenting recordings increased clip counts but not the numbers of independent plants, colonies, locations, or recording events.
- Application and contribution: The framework combined CNN predictions, acoustic amplitude, severity estimation, and geographic information to support plantation-level monitoring and targeted inspection.Its outputs are intended as screening and decision-support tools rather than replacements for expert field assessment.
4. Conclusions
The study presents an integrated AI–IoT framework for non-destructive ULWT detection, severity estimation, and geospatial monitoring, demonstrating feasibility under field conditions while identifying priorities for broader validation.
- The framework combines trunk-mounted acoustic acquisition, mobile field data collection, CNN classification, severity estimation, and web-based geospatial visualization.
- 81.5% accuracy, 80.6% precision, 83.0% recall, 81.8% F1-score, and 0.819 AUC were achieved on the held-out test set.These results demonstrate feasibility of acoustic-based ULWT detection under field conditions.
- The framework extends beyond binary classification by combining infestation probability, acoustic amplitude, and nearby infested plants to estimate severity and visualize potential spatial distribution.The approach provides an initial severity-estimation framework rather than an independently validated measure of infestation intensity or spread.
- False-negative predictions mean the system should currently support screening and decision-making rather than replace expert field inspection.
- Future work will expand data across plantation and environmental conditions, evaluate alternative deep learning architectures, and further validate severity estimation.These efforts are intended to improve robustness and generalisability for wider plantation-level application.