insectnet-research-0.2.0-perch2

InsectNet Research 0.2.0

A deterministically built research classifier for broad insect presence, cicadas, and Orthoptera, fitted on frozen 1,536-dimensional Google Perch 2 embeddings. Exact row-level training manifests are retained privately rather than redistributed.

Status: frozen public research artifact; these exact joblib bytes are not deployed. The locked independent dog challenge produced broad insect activations in 11/26 windows. Do not use this model for unattended ecological counts or alerts.

Artifact SHA-256: 27bf603a6dec2df2789b3bf9241f5e035ccdea5909c4ecf252623ff9304afe32
Dataset hash: d59cde46a933b85c1cd46f944c6df5c9c6a7587efe7c354b5cbe22b2d0698240

Current operational relationship

Since July 2026, a later JSON/NPZ successor from the InsectNet research family has operated as a conservative review-only head in the private field site's shared Perch 2 listener. That successor is insectnet-dev2-field-probe, bundle ID ff8d28f5e8d0416eb63e7b958b6b950b69dd89a0baca0089f399b20cbc3c529f, with field review threshold 0.9995. Its exact checked bundle is now published as InsectNet Research 0.3.0.

The successor is a distinct artifact; it is not this joblib under another name. Its threshold crossings remain unconfirmed model assertions. Deployment does not transfer field-validation claims to this frozen public candidate. See the living deployment contract.

Outputs

The fitted heads are independent probabilities, not a softmax. The serialized inference contract suppresses subtype outputs unless insect_present also passes.

The source evidence supports Cicadidae versus broad Orthoptera. It does not support inventing cricket/katydid versus grasshopper labels.

Feature contract

• Backbone: Google Perch 2 CPU release
• Audio input to Perch: five seconds, 32 kHz, mono, float32
• Input to this artifact: one 1,536-dimensional Perch embedding
• Verified Perch model-tree SHA-256: 3fb2d54b3e34534f1130052b25737e54bbb5ebfd340ec040d4510772b64c81ff
• Perch weights are not included

The official Perch code repository is Apache-2.0. The available Kaggle metadata did not expose a separate model-weight license during this build, so this card does not extend the code license to the weights by assumption.

Training data

InsectSet459 contains 2,096 unique permissive Train/Validation observations from 334 contributor groups: 1,009 cicada and 1,087 Orthoptera windows. Group IDs, not windows, were assigned to train, validation, and test.

No source audio is redistributed here.

Internal grouped evaluation

Test macro AP: 0.963. Test macro F1: 0.910.

These results show that the grouped source partitions are learnable. They do not establish stationary-microphone field reliability.

Locked independent dog challenge

The final artifact was scored once on 26 permissively licensed Domestic Dog sounds from 26 observers. These samples were not used for fitting or threshold selection.

Outdoor taxon recordings may contain incidental insects, but a 42.3% broad activation rate is too high to treat this model as field-ready.

Private stationary-microphone activation audit

A private reviewed frog archive supplied 1,308 five-second windows from 436 recordings. It was reviewed for frog presence, not insect absence, so these are activation rates—not false-positive rates:

Summer frog recordings can contain real insects. Listening review is required before any of these activations can be called errors. No private paths or audio are included here.

Use

import joblib
from insectnet_perch2.candidate import active_labels, predict_candidate

package = joblib.load("insectnet-research-0.2.0-perch2.joblib")
embedding = ...  # shape (N, 1536), from the verified Perch 2 feature contract
scores = predict_candidate(package, embedding)
labels = [active_labels(package, row) for row in scores]

Do not feed waveforms directly into this artifact.

What is validated

• Exact retained-private data, window, embedding, and artifact hashes are recorded.
• The private build audit found no contributor/observer/source groups crossing partitions.
• Only CC0, CC BY, and the explicitly noncommercial ESC-50 research lane were used.
• Broad/subtype hierarchy is serialized and tested.
• The model loads and scores finite Perch embeddings.
• Internal grouped metrics and independent challenge failures are both published.

What is not validated

• local field precision or recall
• unattended field detections
• species-level insect identity
• cricket/katydid versus grasshopper separation
• robustness to frogs, ventilation, rain, distant birds, or recorder changes
• calibrated ecological abundance estimates

Known failure mode and next research requirement

Frozen linear transfer on this corpus still carries source/domain shortcuts. Adding chicken, cat, and poultry hard negatives improved known-source rejection but did not generalize to the untouched dog challenge. The next serious model should use reviewed same-recorder hard negatives, more diverse iNaturalist non-insect taxa, and likely supervised backbone fine-tuning or a richer temporal head rather than another threshold-only adjustment.

Sources

• InsectSet459: academic-datasets/InsectSet459, pinned revision documented in provenance
• ESC-50: Piczak, DOI 10.1145/2733373.2806390
• Ross-308: Díaz de Cerio et al., DOI 10.34810/DATA3437
• Perch: Ghani et al., Global birdsong embeddings enable superior transfer learning for bioacoustic classification (2023)

Source summaries, retained-private manifest hashes, metrics, and challenge reports are included beside the artifact. The public bundle supports artifact/summary consistency checks but cannot independently reconstruct every row or split assignment without the retained private manifests and source audio.

Git source

This historical Perch 2 artifact belongs to the InsectNet repository; it is not part of the legacy BirdNET-logit InsectNet repository.