Directional pest control for consumer pest identification

Overview

An on-device image identification system for consumer pest diagnosis and treatment support. It identifies eight common pest species and separates them from the three conditions they are most often confused with: disease, nutrient disorder, and a healthy leaf. The distinguishing component is a directional error-control layer that goes beyond normal classifier output. It detects predictions that would imply the opposite intervention to what the visual evidence supports, and abstains when the evidence is ambiguous. Instead of guessing, the system returns a cost-calibrated unclear response plus what additional observation would settle the case. Every prediction and abstention is logged for audit and future improvement.

The benefits are practical: when the economic or safety cost of choosing the wrong treatment is high, a wrong direction classification is more damaging than a generic error. By designing specifically for this failure mode, the system reduces the chance that a consumer applies a product that is directly opposite to what is actually needed.

Technical specifications

Key technical and performance details:

• Classifier architecture: ConvNeXt-Base at 384×38 px, trained on 259,000 labeled images.

• Current pest taxonomy: spider mite, aphid, thrips, whitefly, fungus gnat, caterpillar, leaf miner, and mealybug.

• Measured accuracy: 94.3% overall accuracy and 83.5% on pest classes, on a held-out split with no image leakage.

• Calibration: calibration error of 0.113, with temperature scaling fitted on a 644-case holdout the model had not seen.

• Directional error control: identifies classifier mistakes that point a user to the opposite control approach, then withholds high-cost guesses when evidence is contradictory.

• Deployment: runs on-device without a cloud round trip. Planned packaging is ONNX output with a documented API and target-hardware benchmarking.

Technology readiness level

The core model and abstention mechanism are implemented and measured, not just planned. According to the developers, approximately 300 tracking training runs have been logged since 2025, a confusion matrix is retained, the calibration curve has been fitted, and the abstention layer is delivered. However, the existing training corpus is primarily from cultivated plant-−images, so a new household-pest taxonomy is the current validation gap. The proposed validation plan has four phases: (1) map new pest labels onto the existing schema and train a baseline model, (2) measure directional errors on a clean holdout and report a confidence interval, (3) apply targeted constraints and cost-aware abstention, then remeasure the same error rate, and (4) package as an ONNX model with an on-device benchmark. A go/no-go decision after phase 2 allows partners to stop if the actual directional error rate turns out to be low.