SoftTech AI Registry
John Deere · See & Spray Gen 2

Model-Selected Target Spraying

John Deere See & Spray Gen 2 uses camera imagery and trained machine-learning models to distinguish weeds from crops during targeted spraying. A model-derived weed identification determines which observed targets enter the selective-spray path, and downstream spray controls physically apply herbicide at the corresponding target/location without individual operator approval for each qualifying target.

Recorded characteristics

Function
Smallest controlled object: one model-identified weed target / corresponding spray location (not an exact one-inference/one-nozzle-pulse/one-plant relationship; not a whole field or spraying job). Causal chain: camera observation → trained machine-learning model differentiates crop from weed → model-derived weed identification determines the selective-spray branch → downstream spray/nozzle control executes the qualifying target action → herbicide is physically applied at the corresponding target/location. Deere states: "Using camera vision technology and machine learning, See & Spray™ differentiates in-season crops from weeds — and only sprays the weeds." Model-decision boundary: the learned component owns the classification-sensitive judgment of whether the observed target belongs to the weed/target class. The downstream execution rule may effectively be "qualifying weed target → spray"; that does not displace model ownership, because the downstream machinery does not independently establish which observed target is a weed. The model is not recorded as directly actuating nozzles. Physical consequence: herbicide is physically applied at the corresponding model-selected target/location; no biological, efficacy, crop-health or environmental outcome is recorded. Analytical classification: Category 2 — the learned weed/crop judgment is necessary to determine which individual observed target enters the selective-spray branch; downstream machinery executes that model-dependent determination. Duplicate boundary: #168 Meta = model-derived judgment over a platform-native digital object → digital enforcement; #165/#166 = model-derived control/setpoint → controlled equipment/process state; #169 Waymo = model-derived trajectory → machine's own movement; #170 = model-derived judgment about an independently existing external physical target → selective physical intervention at that target/location.
Data access
Imagery from boom/sprayer-mounted cameras observing the crop row and field surface during spraying. With See & Scout (standard on Gen 2 from 2027), camera data may be recorded and streamed to the customer's John Deere Operations Center account. Exact production input feature set not publicly established.
Actions
Can take actions
External actions
Yes
Human confirmation
Not required
Permission basis
Not established
Administrative control
Constraint boundary (operator/system constraints, not credited to the model): operator activation, configuration and operation of the spraying job; spraying mode (targeted in-crop, targeted fallow); crop/application configuration (supported crops; broadleaf weeds only in wheat and barley); chemical/product and tank configuration (single or optional dual-product; operator-chosen mixes); operating restrictions (targeted-mode speeds "up to 16 mph depending on crop and configuration"); nozzle system options (ExactApply, Individual Nozzle Control Pro). Human confirmation not_required applies only to the individual qualifying target action: the operator configures/enables and operates the spraying job, but individual model-identified qualifying targets do not require separate human approval before targeted spraying executes. No claim is made that the overall system operates without human involvement. Outside action yes: the model-derived target judgment results in physical action against an independently existing external target/location, outside the model/software decision boundary. Permission basis not_established: the operator's legal/agricultural authority and equipment ownership are not Registry software permission mechanisms. Default state not_established: operating-mode selection by the operator does not establish a customer-deployment default under the Registry methodology. Auditability (evidence-supported only): Operations Center integration generates weed pressure maps, as-applied data and savings insights (January 2026 announcement).
Default state
Not established
Availability
Current commercial production: See & Spray Gen 2 single-product hardware comes standard on all 2027 and newer 400R and 600R Series sprayers (MY27 408R, 410R, 412R, 612R, 616R). Targeted in-crop application in corn, soybeans, cotton, sorghum, canola, sugar beets, peanuts and edible beans; wheat and barley broadleaf weeds only; targeted fallow. Product options may not be available in all regions.
Licensing
"In-crop application passes require the purchase of a renewable software license. Targeted fallow applications require no additional cost." Hardware standard with an optional factory deduct.
External model or provider
John Deere's own trained machine-learning models (no third-party model provider established).
Limitations and uncertainty
Automated monitoring deliberately withheld because Deere's published website terms restrict automated crawling/data extraction. Deere pages require JavaScript and were read once in a browser; no repeated fetch was made. Evidence control: the Gen 2 product page carries the classification-sensitive evidence; the January 2026 announcement is used for platform context, operating restrictions and records. Select, Premium, Ultimate and broader See & Spray material is not used to assert any Gen 2 component or behaviour. Reversibility: future spraying can be stopped or reconfigured, but herbicide already physically discharged at a target/location cannot simply be undone; an executed application is not straightforwardly reversible. Exclusions: (1) autonomous vehicle movement; (2) route/path authority; (3) biological weed-kill outcome; (4) guaranteed herbicide efficacy; (5) environmental-outcome authority; (6) autonomous chemical selection; (7) model-owned application-rate decision; (8) Variable Rate functionality (biomass detection with operator-set rate/threshold controls — a different proposition); (9) broadcast spraying, including the dual-tank broadcast pass; (10) crop-health decision; (11) online learning/retraining during field operation; (12) exact plant-species identification; (13) exact one-inference/one-nozzle-pulse/one-biological-plant correspondence; (14) transfer of Gen 2 architecture to Premium; (15) to Ultimate; (16) to Select; (17) direct model control of valves/actuators; (18) exact target-to-nozzle implementation; (19) See & Scout field insights as an action. Not publicly established: (1) exact production ML architecture; (2) exact machine-level model output representation; (3) whether the output is a class label, segmentation, coordinates, confidence score or compound representation; (4) confidence representation; (5) exact classification threshold; (6) sensitivity-setting effect; (7) complete post-model decision logic; (8) downstream veto logic; (9) exact target-to-nozzle mapping; (10) target-to-nozzle timing; (11) nozzle-command protocol; (12) internal control interface/bus; (13) valve/actuator implementation; (14) exact action latency; (15) failure handling; (16) retry behaviour; (17) behaviour when a required nozzle is unavailable; (18) exact production model/version; (19) training-data composition; (20) training-data volume; (21) retraining cadence; (22) deployment/update cadence; (23) production online learning; (24) local field-driven model updating; (25) model rollback/version reversion; (26) complete chemical/application constraint set; (27) complete environmental restriction logic; (28) exact operating-speed restrictions for all relevant modes; (29) complete crop-specific constraint set; (30) maximum autonomous action scope during a job; (31) per-target inference record; (32) per-target confidence retention; (33) model version associated with individual actions; (34) per-target explanation; (35) retention period for target/action records; (36) whether mapping preserves one-to-one target history; (37) exact model parity with Premium; (38) exact model parity with Ultimate; (39) exact model parity with Select; (40) exact execution/nozzle parity across product variants.

Evidence