Agriculture is the slowest big industry to adopt AI, and also the one with the most to gain. Global food production has to increase by roughly 50 percent by 2050 to meet projected demand, against a backdrop of climate volatility, shrinking arable land, and tightening water supply. Precision agriculture, using satellites, sensors, and machine learning to make per-field decisions, is the most promising path to closing that gap.
A recent survey found that 81 percent of large farms are willing to adopt AI tools for crop management when the tools are affordable and integrate with existing workflows. The bottleneck is not appetite. It is availability of tools that do not require a dedicated data science team to operate.
SciRouter's agriculture API wraps the best open source models (Microsoft's FarmVibes, China's Sinong, CropGPT and others) into a single endpoint. Upload a field boundary or a photo of a diseased leaf, and the API returns actionable answers.
The open source agricultural AI stack
Agricultural AI was mostly proprietary until around 2023, when several major open source efforts launched in parallel:
- FarmVibes.AI from Microsoft Research. An open source precision agriculture platform with pretrained workflows for yield estimation, weed detection, and crop type mapping. It unifies satellite, weather, drone, and soil sensor data.
- Sinong from a Chinese research consortium. One of the largest public collections of crop disease datasets and pretrained models, released in 2024.
- CropGPT and variants. Language-model wrappers that let farmers query agronomic knowledge bases in natural language.
- Prithvi-EO (NASA and IBM). A foundation model for satellite imagery that fine-tunes well for crop-type mapping and anomaly detection.
Each of these is powerful but none of them are easy to self-host. The API wraps them with the same schema so you can swap between them or run them in parallel.
Crop disease detection
Disease detection is the most photogenic precision agriculture task and also one of the most useful. Early identification of wheat rust, potato late blight, tomato bacterial spot, or rice blast can prevent yield losses of 20 to 50 percent when caught in time. The classical approach involves an agronomist walking the field, which scales poorly. A smartphone camera plus a trained classifier scales dramatically better.
import httpx
API_KEY = "sk-sci-..."
BASE = "https://scirouter.ai/v1"
# Upload a photo of a suspicious leaf
with open("leaf.jpg", "rb") as f:
response = httpx.post(
f"{BASE}/agriculture/detect-disease",
headers={"Authorization": f"Bearer {API_KEY}"},
files={"image": ("leaf.jpg", f, "image/jpeg")},
data={
"crop": "tomato",
"model": "sinong-v2",
"return_severity": "true",
},
timeout=60,
)
result = response.json()
print(f"Top disease: {result['top_prediction']['label']}")
print(f"Confidence: {result['top_prediction']['confidence']:.1%}")
print(f"Severity (0-3): {result['severity']}")
print(f"Recommended action: {result['advisory']}")The call returns a disease label, severity score, and a short advisory generated from a curated agronomy knowledge base. For researchers, the raw model logits are also returned so you can evaluate uncertainty and ensemble across models.
Yield prediction from satellite time series
Yield prediction is a different kind of problem. Instead of a single photo, it uses a time series of satellite imagery over a growing season combined with weather data, soil properties, and historical yields. Modern models treat this as a multimodal time-series regression and can predict end-of-season yield to within 5 to 10 percent of actual for major grain crops.
The API takes a field boundary (GeoJSON), a crop type, and a growing season. It fetches Sentinel-2 and weather data automatically, runs a pretrained FarmVibes workflow, and returns an expected yield with an uncertainty interval.
- Satellite inputs. Sentinel-2 optical for vegetation indices, Sentinel-1 SAR for biomass under clouds, MODIS for higher-frequency coarse-resolution data.
- Weather inputs. ERA5 reanalysis and short-range forecasts for temperature, precipitation, radiation, and humidity.
- Soil inputs. SoilGrids for texture, organic carbon, and nutrient proxies.
Weed detection and variable-rate spraying
Another high-impact application is weed detection from drone imagery to drive variable-rate herbicide application. Instead of spraying the entire field uniformly, targeted spraying reduces chemical use by 50 to 90 percent depending on weed pressure. FarmVibes includes pretrained weed classifiers, and the API exposes them for both orthomosaic imagery and real-time drone feeds.
Soil moisture and irrigation
For irrigated agriculture, knowing the soil moisture state of each part of a field enables precise water application. SMAP satellite data provides global soil moisture at a coarse resolution, and downscaling workflows (using Sentinel-1 SAR and weather data) can bring that to field scale. The API exposes downscaled soil moisture as one of its standard outputs.
Insurance and risk assessment
Parametric crop insurance pays out based on objective indices (temperature days, rainfall thresholds, NDVI deficits) rather than traditional loss adjustment. This is much faster to settle and less prone to fraud, but it needs reliable remote sensing to work. The API provides the underlying measurements (NDVI, EVI, rainfall aggregates) that feed parametric insurance contracts.
CropGPT and natural language access
One of the most interesting recent developments is CropGPT-style natural language wrappers. A farmer can ask “why does my north field look worse than my south field this week?” and get a grounded answer that pulls from recent satellite imagery, weather history, and agronomic knowledge bases. The API exposes a chat endpoint that combines retrieval over the same data used by the other endpoints.
This is the kind of interface that makes precision agriculture actually usable for people who are not GIS analysts. Farmers, extension workers, and advisors can get answers without opening a mapping application.
Who is using the agriculture API
- Agricultural extension services in developing markets where per-field AI advice would otherwise be cost-prohibitive.
- Insurance underwriters building parametric products.
- Cooperatives aggregating many small farms and providing agronomy as a service.
- University researchers studying climate impacts on crop production.
- NGOs running early-warning systems for food security in at-risk regions.
Getting started
The quickest path is Agriculture Lab, which provides a web interface for drawing field boundaries, uploading leaf photos, and running the various workflows. It is designed for agronomists and researchers who do not want to write code but need to move faster than a desktop GIS allows.
For production integrations (farm management systems, insurance platforms, advisory apps), the REST API and Python SDK are the right surface. The API is stable, versioned, and covered by the same SLA as the rest of SciRouter.