On-Farm AI: Weed ID, Biomass Estimation, Disease Detection

The Vision Stack Beyond Pest Scouting

Pest identification is the most visible application of on-farm computer vision, but not the broadest. The same core architecture, a convolutional neural network (CNN) trained on labelled agricultural image datasets, applies to four distinct field problems: pest detection, weed identification, pasture biomass estimation, and crop disease detection. The first of these is covered in the companion spoke on vision-based pest scouting. This spoke covers the remaining three, each with its own commercial deployment path, accuracy profile, and operator decision layer.

The underlying architecture is consistent. A CNN accepts a multi-band image (RGB, near-infrared, or red-edge) as input and outputs a classification or a regression value: "this plant is a broadleaf weed, confidence 0.91", "this paddock contains 2,340 kilograms of dry matter per hectare", "this canopy shows early Septoria infection probability 0.78". The model runs on a GPU or dedicated AI accelerator. When inference runs on-device at the point of data collection rather than in a vendor cloud, the crop health signal stays on the farm. When it runs via a proprietary SaaS platform, the signal is available to whoever holds the data contract.

The cost structure changed decisively in the 2020-2025 period. Model training costs dropped as pre-trained architectures (ResNet, EfficientNet, YOLOv8) became freely available under open licences, allowing agricultural researchers and commercial startups to fine-tune on domain-specific datasets without funding large compute runs from scratch. Inference hardware followed: a Nvidia Jetson Nano at 99-149 USD and a Google Coral USB Accelerator at 24.99 USD (Nvidia and Google 2024 product pricing) now run production-grade vision models in weatherproof enclosures in a field. The marginal cost of an additional inference pass on a deployed system is effectively zero. What that inference covers, and where the output data goes, is the operator's decision.

Weed Identification: See and Spray, LaserWeeder, and the Open Alternative

Commercial weed identification at production scale arrived in two distinct product forms. Blue River Technology's See and Spray system, acquired by John Deere in 2017 for 305 million USD, mounts a camera bar above the crop row and classifies each plant it passes as crop or weed at camera speed. An EfficientDet-derivative CNN running on the onboard processor makes the classification call in milliseconds and activates an individual nozzle only where a weed is detected. In commercial wheat, corn, and soybean trials, John Deere published herbicide volume reductions of 77 to 90 percent compared with blanket application (John Deere See and Spray Ultimate commercial data, 2023). The See and Spray Ultimate system became available to US farmers from 2023 as an attachment for compatible sprayer platforms.

Carbon Robotics LaserWeeder takes a different path: elimination rather than reduction. The system uses a 100-watt fibre laser guided by a custom vision stack, a YOLOv5-derivative architecture trained on over 40 weed species common to US vegetable and row-crop production, to ablate individual weed plants at the growth point. At rated field speed the system destroys more than 200,000 weed plants per hour (Carbon Robotics 2023 field data), eliminating herbicide use entirely in treated rows. Commercial deployments in lettuce, brassica, carrot, and strawberry operations across the western United States began in 2022. The limiting factor is throughput at current field speeds: laser ablation is processing-intensive, and row spacing and crop density affect the achievable pass rate.

The open alternative to both proprietary platforms is a weed detection model fine-tuned on the WeedAI open dataset (Australian Grains Research and Development Corporation, 2021 onwards, weedai.weeds.ai), which contains labelled images across 47 weed species in multiple crop contexts. Ultralytics YOLOv8 models fine-tuned on WeedAI and similar datasets run on Jetson Nano hardware at 25-30 frames per second, sufficient for implement-mounted spot-spray trigger logic. The model weights are operator-owned. The inference data does not leave the machine. The training pipeline is reproducible on any Linux workstation with a mid-range GPU. The data sovereignty question for weed ID is the same question that applies across every Farm Intelligence category: what crop-health signal leaves the farm, and who retains it. The choice between Data Sovereignty at the farm gate and data aggregation at the platform level is made in the hardware and software purchase, not separately.

Biomass Estimation: Drone and Satellite CNN Models for Pasture kg DM/ha

Rotational grazing depends on knowing how much forage is in a paddock before sending cattle in. Traditional measurement uses a rising plate meter and a calibration formula: an operator walks transects, takes 30-50 readings, and converts the compressed plate height to a dry matter estimate in kilograms per hectare. On a 500-hectare block with 20 paddocks, this is three to four hours of walking per assessment round. The assessment should ideally run every five to seven days during active growth, which is operationally difficult at working farm scale.

Satellite-based CNN biomass estimation compresses that to a five-minute processing task. Sentinel-2 bands 4 (red, 665nm), 8 (NIR, 842nm), and 8A (red-edge, 865nm) provide the reflectance inputs most correlated with green biomass density. A CNN trained on ground-truth clipping samples outputs a kilogram-per-hectare surface at 10-metre resolution every five days, cloud cover permitting. A CSIRO study validating this approach against 640 physical clipping samples across Australian pasture types (Liu et al. 2022, Remote Sensing in Ecology and Conservation) demonstrated an r-squared of 0.82 at paddock scale. The commercial product PastureMap (Australia, 2016 onwards) uses Sentinel-2 imagery as its primary biomass input, with the satellite data layer free at source and the platform handling model inference and display.

Drone surveys achieve higher accuracy where the revisit frequency and ground resolution justify the flight cost. A DJI Phantom 4 Multispectral or MicaSense Altum-PT sensor flying at 60-80 metres above ground generates a 5-8 centimetre per pixel multispectral orthomosaic. Structure-from-motion processing adds a canopy height model. A ResNet-18 backbone with attention mechanisms trained on canopy height plus NDVI and red-edge inputs achieves r-squared values of 0.88-0.93 against clipping reference data at paddock scale, as demonstrated across multiple University of New England (Australia) trials published between 2021 and 2024 in Precision Agriculture and Remote Sensing. The best-performing open architecture for this task, ResNet-18 with spatial attention heads as described by Moeckel et al. (Remote Sensing 2022), is available under an open licence.

The operator decision these models support is rotation timing: entry at approximately 2,500-3,000 kg DM/ha in temperate pastures, exit at 1,500 kg DM/ha to preserve root reserves and recovery rate. A biomass map generated from Sentinel-2 every five days makes that decision based on the entire paddock surface rather than a 30-point plate-meter walk. The accuracy loss relative to drone survey (r-squared 0.82 vs 0.90) is acceptable for rotation management; it becomes material only for research trial measurement where absolute dry matter values are needed for published results.

Disease Detection: RGB Limits and the Multispectral Advantage

Crop disease detection via computer vision has a fundamental timing problem. By the time lesions are visible in RGB imagery, the pathogen has already colonised the tissue and the infection cycle is established. An RGB-based CNN classifying canopy images against the PlantVillage dataset (Mohanty et al. 2016, Frontiers in Plant Science), which achieved 99.35 percent accuracy across 26 disease classes in controlled conditions, is detecting disease after the fact rather than before the economic threshold. In field conditions with variable lighting, canopy architecture, and growth stages, accuracy for the same models falls to 70-85 percent. The information is still useful for mapping the spatial distribution of established infection, but it is not the intervention-enabling early warning the operator needs.

The multispectral advantage is specific to the red-edge band at 705-745 nanometres. Chlorophyll begins degrading at the cellular level in the three to seven days before a fungal lesion becomes visible, and this degradation registers as a reduction in red-edge reflectance before it registers as a change in visible colour (Mahlein et al. 2019, Plant Disease). A CNN trained on multispectral time-series data, using red-edge reflectance change as a key input feature, can classify early-stage Septoria tritici infection in wheat, wheat stripe rust, and powdery mildew with field-condition accuracy of 80-88 percent at the three-to-seven-day pre-symptom window, compared with 70-85 percent post-symptom for RGB-only classifiers.

Hummingbird Technologies (UK, 2015 onwards) deploys multispectral drone surveys at commercial scale for wheat Septoria, rust, and barley yellow dwarf virus risk mapping across UK and European cereal operations. A survey flight using a fixed-wing drone platform covers 200-400 hectares per day and produces a risk map within 24 hours of data processing. Ceres Imaging (US) operates a comparable service for almond, pistachio, and grape operations in California, mapping water stress and early disease pressure from multispectral aerial imagery. Both are service models rather than operator-owned platforms: the operator pays per flight or per season, the vendor retains the imagery and model outputs.

The on-device alternative is a multispectral sensor such as the MicaSense Altum-PT (approximately 9,500 USD body price) mounted on an operator-owned drone, combined with a locally-run DenseNet-121 or EfficientNet-B4 model fine-tuned on a labelled disease dataset relevant to the operator's crops. The academic literature on agricultural disease detection CNNs is extensive enough that a competent machine-learning practitioner can fine-tune a usable classifier from open pre-trained weights in two to three weeks of work with 2,000-5,000 labelled images from the target crop system. The resulting model runs on a Jetson Nano. The data and the model weights are the operator's.

Edge Inference: Jetson Nano, Coral TPU, Raspberry Pi 5

Three hardware platforms cover the operator-affordable edge inference tier as of 2024-2026 pricing. The Nvidia Jetson Nano Developer Kit, priced at 99-149 USD (Nvidia 2024), packages a 128-core Maxwell GPU delivering 472 GFLOPS of compute alongside a quad-core ARM Cortex-A57 CPU in a 69.6 x 45mm form factor. In practice it runs YOLOv5-nano at 30 frames per second and YOLOv5-medium at 12-18 frames per second for real-time weed and disease classification from a mounted camera. Power draw is 5-10 watts. The board fits into a standard IP65-rated enclosure for field deployment.

The Google Coral USB Accelerator at 24.99 USD (Google 2024) takes a different architectural approach: a dedicated Edge TPU delivering 4 trillion operations per second in a USB stick form factor drawing under 3 watts. The constraint is model compatibility. The Edge TPU runs only quantised TensorFlow Lite models, and architectures must be compiled for the TPU hardware using the Edge TPU compiler. MobileNetV2 and EfficientDet-Lite run at under 3 milliseconds per inference, well beyond real-time for implement-speed applications. The Coral pairs with any Linux host, including a Raspberry Pi 4 or 5, making a complete inference system available for under 75 USD total.

The Raspberry Pi 5 (55-80 GBP for the 8GB variant, Raspberry Pi Foundation 2023) adds the Hailo-8 M.2 accelerator (approximately 45-70 USD, Hailo 2024) via the PCIe interface to reach 13 TOPS for vision workloads. The Hailo-8 runs YOLO architectures natively via the Hailo Model Zoo, including YOLOv8 and RT-DETR variants. At 13 TOPS the system is adequate for drone-mounted real-time classification and for implement-speed ground systems. Total system cost for the Pi 5 plus Hailo-8 plus a housing and power supply is 130-160 USD.

All three platforms run the same open-weight model ecosystem: Ultralytics YOLO family, TensorFlow Lite, and ONNX Runtime. None require cloud connectivity for inference. A model fine-tuned on the operator's own labelled images, deployed on Jetson or Coral hardware, generates crop health data that never leaves the farm network. Blue River See and Spray Ultimate transmits field telemetry to the John Deere Operations Center, where it joins 150 million-plus enrolled acres of equipment data (John Deere Annual Report 2023). Carbon Robotics LaserWeeder retains operational data under a vendor agreement. The open-hardware path eliminates both: inference runs locally, model weights are operator-controlled, and the data stays behind the farm gate. The question of where crop intelligence flows is answered at the point of hardware selection, not in a terms-of-service review. As the Data Sovereignty spoke addresses in detail, the precision agriculture data layer is the rent stack's seventh layer, and the AI model running on a 99-dollar board is the mechanism by which the operator can choose to own it.

A 99-dollar board runs the model. The inference cost is zero. The data stays with the operator who grew the crop.