Lydos Air delivers eight surgical crop-intelligence engines composed into a single KSL-signed decision pipeline. From EPPO-coded entomovision classification through multispectral Pest Stress Index, wingbeat acoustic species matching, predator-modulated trophic forecasting, and ETL/EIL integrated pest management decisions to spot-spare nozzle valve planning, biocontrol drone-drop scheduling, and privacy-preserving federated learning — every action in the agriculture workflow is chain-hashed, auditable, and operated entirely within the farm's data perimeter.
Pest, beneficial, and neutral arthropod classification from crop-canopy imagery. Every detection is EPPO-coded (e.g. HELIAR for Helicoverpa armigera, APISME for Apis mellifera), role-tagged, and written to a chain-hashed audit ledger. Raw imagery never enters the platform database — only sha256 reference, bounding box, and confidence score.
NDVI + NDRE + canopy thermal readings are composed into a transparent Pest Stress Index (0.5·NDRE component + 0.3·thermal deviation + 0.2·NDVI deficit). RFC 7946 GeoJSON pollinator corridors are evaluated point-in-polygon with Jordan curve ray-casting. Each grid cell receives HEALTHY / WATCH / STRESSED / CRITICAL classification with chain-hashed audit.
Dominant wingbeat frequency extracted from a 40 kHz MEMS microphone recording is matched against a KSL-signed species library. Seed values: Apis mellifera 230 Hz, Bombus terrestris 150 Hz, Helicoverpa armigera 60 Hz, Tuta absoluta 80 Hz. Bandwidth-aware linear-decay confidence mathematics; critical for night operations when visual detection fails.
Predator-prey (Coccinella → Aphis) and parasitoid-host (Trichogramma → Helicoverpa egg) directed-graph with predator-modulated first-order decay forecast. 24-hour observation window, 1–168 h horizon. A suppression factor ≥ 0.40 triggers a natural-control veto recommendation passed to the IPM decision engine.
A pure-deterministic EPPO/EFSA ETL/EIL decision matrix merges upstream signals — entomovision counts, multispectral PSI, pollinator corridor flags, wingbeat role summary, and trophic veto — into five human-readable codes: NO_SPRAY_ZONE, MONITOR_ONLY, BIOCONTROL_RECOMMENDED, SPRAY_APPROVED_CAUTION, SPRAY_APPROVED_URGENT. Every decision carries reason codes, is chain-hashed, and KSL-signed.
Per-cell IPM decisions are deterministically mapped to per-nozzle OPEN / OPEN_REDUCED / CLOSED valve actions across a 50 cm grid. Unknown or unmapped decisions default to CLOSED (fail-closed, preventing accidental spray). Transparent latency model: 5 ms base + 0.2 ms per nozzle, with a warning flag when the budget exceeds 80 ms.
Per-pass release plans for biological control agents — parasitoid wasp cards, ladybird vials, lacewing larvae, predatory mite sachets, aphid parasitoid capsules. Regulatory clearance references are stored as hashed values only. Target pest EPPO codes must be a subset of the registered agent authorisation list; default state is NOT_DISPATCHED (fail-closed) because biological agents are living organisms.
Privacy-preserving cross-farm federation: each insight shares only EPPO code, broad geographic region prefix (TR-W, EU-S, US-MW, AS-E, …), sha256 embedding fingerprint, confidence in [0,1], and anonymous participant hash. Raw imagery, GPS coordinates, IP addresses, and email addresses never enter the database. Federated rounds produce a deterministic aggregate signature; participants run local averaging on their own hardware.
The European Plant Protection Organisation's open Bayer-code taxonomy is the single source of truth for all insect identifiers across the entomovision, wingbeat, trophic, and IPM engines. EPPO codes ensure species identifiers are unambiguous across regulatory jurisdictions.
Economic Threshold Level and Economic Injury Level values used in the IPM Decision engine follow EFSA scientific opinion on pest risk. Threshold inputs are registered per crop-pest EPPO pair by the operator; the platform enforces EIL ≥ ETL structurally.
Precision-agriculture spray drone missions operate under FAA Part 107 Commercial UAS rules in US airspace. The platform's preflight validation, geofence enforcement, Remote ID metadata capture, and multi-operator approval chain are designed to align with Part 107 operational requirements.
European agricultural drone deployments targeting SAIL II–IV operations. The platform's ConOps documentation support, multi-operator approval, ground risk buffer enforcement, and hash-chained audit ledgers align with the SORA evidence requirements for Specific category risk assessment.
Every chain-hash record across all eight crop-intelligence engines uses SHA-256 per NIST FIPS 180-4. The federated insight aggregate signature is the SHA-256 of sorted chain hashes joined with a pipe delimiter — deterministic and permutation-invariant, verified by participants locally.
Raw farm imagery, GPS coordinates, operator email, and user identifiers never leave the originating site. Federated learning shares only hashed embeddings and broad region prefixes. This structural constraint is enforced at the schema level — not a configuration flag — and is verified by test invariants on every CI run.
A general object detector produces a label and bounding box. The entomovision engine maps that detection to an EPPO Bayer code — a globally unique species identifier maintained by the European Plant Protection Organisation — and assigns a functional role (pest, beneficial, or neutral). This structured output feeds directly into the IPM decision matrix: the pest count advances toward ETL/EIL thresholds, while a beneficial count can trigger a natural-suppression veto. A plain bounding box cannot drive that decision logic.
Any grid cell whose centroid falls inside a registered RFC 7946 GeoJSON corridor polygon is classified as NO_SPRAY_ZONE before the ETL/EIL matrix is even evaluated. The nozzle valve plan maps NO_SPRAY_ZONE to CLOSED for every nozzle over that cell. This precedence rule is hard-coded in the deterministic decision matrix and cannot be overridden by pest pressure alone.
No. Each farm's raw imagery, GPS logs, and operator identifiers stay on-site. Only the federated insight record — EPPO code, broad region prefix, sha256-hashed embedding fingerprint, confidence score, and anonymous participant hash — is submitted to the federated round. The aggregate round produces a deterministic signature that participants use to verify locally computed federated averaging on their own hardware. Lydos Air does not perform cross-farm model training.