GubatTanod: Repurposed-phone acoustic chainsaw detector for Indigenous ancestral forests

Problem Statement

Illegal logging and charcoal (uling) poaching in critical Philippine watersheds—such as the Sierra Madre, Palawan, Mount Apo, and Upper Marikina—destroy the natural forest sponge that protects lowland towns from catastrophic flash floods. Indigenous Peoples’ (Dumagat, Palaw’an, Manobo, Subanen) forest guards and DENR Bantay Gubat volunteers risk their lives patrolling thousands of hectares of steep jungle on foot. Satellite deforestation alerts such as GLAD and Global Forest Watch only detect canopy loss days or weeks after ancient dipterocarp trees have already been felled, sawn into flitches, and hauled out by truck. Meanwhile, proprietary acoustic forest sensors cost $500+ per unit in foreign subscription fees, making them unaffordable for local communities.

Proposed Solution

GubatTanod upcycles e-waste—old Android phones with cracked screens that Filipinos have sitting in drawers—into a low-cost, solar-trickle-charged acoustic sentinel network for community forest guards:

1. Upcycled Canopy Sentinel Node: An old Android phone is stripped of bloatware, sealed inside a ₱400 recycled PVC weatherproof housing with a 5W solar panel and an external omnidirectional microphone, and strapped high in a ridge-line tree along known logging trails.
2. Continuous On-Device Spectrogram CNN: Every 5 seconds, the phone wakes its microphone, computes a 64-bin Log-Mel Spectrogram (focusing on the 150 Hz to 2,500 Hz harmonic signature of two-stroke chainsaw engines, gunshots, and heavy diesel logging trucks), and runs a quantized 180 KB Convolutional Neural Network on-device. It accurately rejects tropical rain, thunder, cicadas, hornbills, and wind rustle.
3. Triangulated Trail Alert for Rangers: When two or more nodes detect a sustained chainsaw signature (over 6 seconds across 3 consecutive windows), they transmit a tiny 80-byte packet over a ₱300 LoRa radio dongle or 2G SMS to the barangay/ancestral-domain base station. The base station triangulates the time-of-arrival and signal-energy ratio to plot the logger’s trail sector and texts an immediate alert to Bantay Gubat rangers while the tree is still being cut.

Target Market

• Primary users: Indigenous Peoples’ ancestral domain forest guards (CADT holders under IPRA, RA 8371), DENR Bantay Gubat deputies, and municipal watershed management councils.
• Public-sector and NGO buyers: DENR Forest Management Bureau (LAWIN Forest and Biodiversity Protection System), provincial environment offices (PENRO/CENRO), water districts protecting critical watersheds, and conservation foundations (Masungi Georeserve Foundation, Haribon, Philippine Eagle Foundation).
• Community benefit: Turns local electronic waste into an early-warning shield that protects both ancestral forests and the rangers patrolling them.

How AI Is Used

• Time-frequency Log-Mel Spectrogram CNN: Extracts Mel-Frequency Cepstral Coefficients (MFCCs) and harmonic comb patterns on-device to separate two-stroke chainsaw fundamental frequencies and harmonics from dense tropical bioacoustic background noise (cicadas and heavy monsoon rain).
• Multi-window temporal persistence filter: Eliminates false alarms from passing motorcycles on distant public roads by requiring stationary harmonic persistence characteristic of tree bucking and felling.
• Multi-node acoustic energy localization: Estimates the approximate source zone between canopy nodes using relative acoustic attenuation (dB drop over distance and terrain) so rangers know which ridge trail to intercept.

Community & Ecosystem Integration

• Plugs into DENR LAWIN and Indigenous CADT patrols: Exports verified acoustic incident logs directly into the spatial format used by DENR’s LAWIN patrolling system.
• Ranger safety first: Alerts rangers whether multiple chainsaws or gunshots were detected so community volunteers never walk blind into an armed logging syndicate without PNP/DENR enforcement backup.
• Circular e-waste hardware: Uses donated old Android phones plus a ₱400 solar-PVC kit instead of imported proprietary hardware.

Demo Scenario (3 minutes for judges)

1. Place two old Android phones running the GubatTanod sentinel app at opposite corners of the stage area.
2. Play three loud background audio tracks through a speaker: heavy tropical rain, deafening mountain cicadas, and a passing tricycle. Both phones display live Mel-spectrograms on the big screen and stay green (Forest Normal — 99% Ambient/Bioacoustic).
3. Now play a recording of a two-stroke chainsaw mixed underneath the loud rain and cicada audio (simulating a chainsaw 800 meters away in the canopy).
4. Within 4 seconds, both phones lock onto the 450 Hz–1,800 Hz harmonic bands, flash red (Chainsaw Detected — 96.4% Confidence), and fire a simulated LoRa/SMS packet.
5. The ranger map triangulates the sound between Node 1 and Node 2 onto Trail Sector B (Sierra Madre Ridge) and sends a Tagalog SMS alert with the audio snippet hash for DENR evidentiary logging.

30-Hour Build Plan

Suggested stack: Android Kotlin + TensorFlow Lite Audio / Web Audio API spectrogram fallback, Python FastAPI, Librosa + PyTorch (audio training), PostgreSQL/PostGIS, Next.js + MapLibre, SMS gateway. Public data / APIs: Forest Sound Classification Dataset (FSC22 — IEEE Dataport), Rainforest Connection (RFCx) open bioacoustic benchmark samples, ESC-50 environmental audio dataset, DENR protected area and watershed boundary shapefiles.

Research & Literature References

1. Ioannis Mporas, Ioannis Perikos, Vassilis Kelefouras, and Michael Paraskevas, “Illegal Logging Detection Based on Acoustic Surveillance of Forest,” Applied Sciences, Vol. 10, Issue 20, 7379, 2020. DOI: 10.3390/app10207379
   • How GubatTanod uses this study: Implements the study’s spectral feature extraction windowing and noise-robust classification pipeline to detect chainsaw operation against heavy forest background noise (wind, rain, insects, and birds).
2. Budimir Anđelić, Milutin Radonjić, and Srđan Djukanović, “Sound-based logging detection using deep learning,” 2022 30th Telecommunications Forum (TELFOR), pp. 1-4, 2022. DOI: 10.1109/TELFOR56249.2022.9985055
   • How GubatTanod uses this study: Adopts the compact Mel-spectrogram Convolutional Neural Network architecture shown to achieve high detection accuracy with a tiny memory footprint suitable for battery-constrained edge devices.

Risks & Privacy Notes

• Zero human speech recording: The sentinel node never records or transmits raw voice conversations. It computes frequency-domain energy features in volatile memory and discards the raw audio buffer immediately unless a mechanical chainsaw or gunshot harmonic is triggered.
• Ranger physical safety: When the classifier detects a gunshot signature alongside chainsaw activity, the alert explicitly instructs community volunteers not to approach alone and routes the alert directly to armed DENR enforcement and PNP Mobile Force units.
• Strict 30-hour feasibility constraint: Continuous microphone wake-locks and high tropical canopy humidity rapidly drain and corrode upcycled consumer phones unless paired with custom waterproof solar enclosures.

AI Use Disclosure (draft)

GubatTanod uses a quantized 2D Convolutional Neural Network over Log-Mel Spectrograms to classify chainsaw, logging truck, and gunshot acoustics on-device, and an acoustic attenuation triangulation model to estimate source locations. AI coding assistants helped build the real-time spectrogram visualizer and PostGIS trail map. The audio classifier was validated on public FSC22 forest sound benchmarks.

Rubric Scoring