Sagri uses satellite data to turn information on soil carbon, water, methane, crops, and land use into parcel-level intelligence.
Without relying on field surveys, we support a wide range of decision-making, from carbon and environmental MRV to government-scale farmland administration. Our technology is being deployed and validated across Japan, Ukraine, the Philippines, Vietnam, Tanzania, and beyond.
Satellite-native
Built on multi-source Earth observation, including optical, SAR, and thermal data.
Registry-aligned
Designed with major methodologies in mind, including Verra, JCM, and J-Credit.
Field-validated
Calibrated and validated against ground-truth data from around the world.
Government-deployed
Used by municipalities across Japan for satellite- and AI-based farmland inspection.
Sagri’s technology stack is built on five core models, each capturing a different layer of the agricultural system. The models can be combined depending on the use case and are deployed across a range of platforms and workflows, from carbon and environmental MRV to farmland administration.
Accurate farmland boundaries are essential for agricultural analytics and land management. Yet in many regions, parcel-level farmland data remains incomplete or unavailable, while creating and updating it manually requires significant expertise, time, and cost.
AI Polygon automatically delineates farmland parcels from satellite imagery within a specified area and outputs standardized polygon data that can be integrated into GIS platforms and farmland management systems.
How It Works
Multi-resolution
satellite imagery
Satellite imagery with spatial resolutions ranging from 50 cm to 10 m is combined to analyze everything from fragmented smallholder parcels to large-scale commercial farms.
Trained on farmland
data from around the world
The model is trained on multi-year farmland data from across Asia, Europe, North and South America, Africa, and Oceania.
Region-aware
deployment
Local reference data can be used to calibrate the model to regional farming practices, parcel structures, terrain, and landscape characteristics.
Production-ready
outputs
Outputs are delivered in standard GIS formats such as GeoJSON for direct integration into existing systems and workflows.
Every model release is evaluated against independently labeled ground-truth datasets.
Performance is evaluated using industry-standard geospatial segmentation metrics, including Panoptic Quality (PQ).
Performance is assessed separately across large-scale agricultural regions and fragmented smallholder landscapes, ensuring that accuracy is validated under real-world operating conditions across diverse agricultural environments worldwide.
Measuring soil across large areas at the resolution required by carbon markets requires extensive sampling and significant cost.
Sagri’s Digital Soil Mapping combines satellite, terrain, climate, and other environmental data with soil samples to predict soil organic carbon concentration, bulk density, and soil organic carbon stock across a project area.
How It Works
Satellite, terrain, and climate data
We use Sagri’s proprietary environmental covariate stack, built from open satellite, terrain, and climate data.
Public datasets and local samples
Public soil datasets such as LUCAS and KSSL are combined with partner-held data and locally collected calibration samples.
Estimate soil properties
and uncertainty
The model outputs SOC concentration, bulk density, SOC stock, and a 90% prediction interval for every pixel.
Pixel-level reliability
scoring
Each pixel is assigned a reliability score indicating whether it is suitable for credit calculation, planning, or screening.
Project-specific validation
Each project is evaluated for held-out accuracy, prediction-interval coverage, and uncertainty mapping, with results delivered in a project-specific validation report. Because soil prediction accuracy depends on local data density, performance is reported for each individual project.
Forward-in-time validation
The most recent year of samples is withheld from the training data and predicted using only historical data to evaluate forward-in-time performance.
Targeted sampling design
The project area is stratified based on soil and terrain conditions and prediction uncertainty, concentrating sampling where variability and uncertainty are highest to maximize the information gained from each field campaign.
Methane from paddy rice accounts for approximately 10% of agricultural GHG emissions. Alternate Wetting and Drying (AWD) and mid-season drainage can reduce emissions, but continuously monitoring water levels across large areas and accounting for potential increases in nitrous oxide (N₂O) remain major challenges.
DRACO is a process-based model that assimilates satellite data to simulate hourly changes in paddy water levels, vegetation, methane (CH₄), and nitrous oxide (N₂O).
How It Works
Process-based
modeling
Coupled water-balance, vegetation, and biogeochemistry submodels simulate changes in water level, crop growth, and GHG emissions within the paddy.
SAR and optical
satellite data
Sentinel-1 and ALOS-2 SAR provide through-cloud inundation information, while Sentinel-2 provides vegetation indices.
Satellite data
assimilation
Sagri’s proprietary data-assimilation pipeline incorporates satellite observations into the model, keeping simulations anchored to real-world paddy conditions.
Hourly water level,
vegetation, CH₄, and N₂O
DRACO simulates water level, vegetation, CH₄, and N₂O on an hourly basis and calculates the net warming impact, including potential N₂O increases triggered by AWD.
Calibrated across rice cultivars
The model is calibrated across multiple rice cultivars spanning Japan, India, and Southeast Asia.
Field validation of water level, vegetation, and GHGs
In Niigata, Japan, water-level predictions are validated against in-situ water-level loggers. In the Philippines and Cambodia, vegetation and CH₄/N₂O are calibrated and validated against measured plant height and gas-flux data.
Independent validation of inundation detection
The Sentinel-1 inundation classifier is benchmarked separately against site-specific flooded-field labels before being assimilated into DRACO. The ALOS-2 inundation-classification technology has been peer-reviewed and accepted for publication in an academic journal, while papers validating DRACO’s water-level and GHG performance are being prepared for submission.
Verifying what is actually growing on each farmland parcel at municipality scale traditionally requires field visits, paper declarations, and labor-intensive cross-checking.
Crop Classification analyzes parcel-level satellite time-series data and uses an independent binary classifier for each crop to estimate what is being grown. It calculates a confidence score for each crop and outputs a ranked list of likely crop types for every parcel.
How It Works
Parcel-level
time-series data
Sentinel-2 time-series data is sampled for each farmland parcel polygon.
Crop-specific
season windows
Satellite data is analyzed within crop-specific season windows aligned with growing seasons in Japan.
Independent classifier
for each crop
A dedicated binary classifier independently determines whether each crop is being grown.
Ranked crop predictions
and confidence scores
Confidence scores are calculated for each crop, and the most likely crop candidates are ranked for every parcel.
Regional phenological variation
The model is being extended to account for regional phenological variation across Japan’s diverse climates, from Hokkaido to Kyushu.
Independently versioned by crop
Each crop-specific model can be versioned and deployed independently. For example, the rice model can be retrained without requiring the wheat model to be revalidated.
Performance targets defined per crop
The models are validated on Japanese datasets spanning multiple prefectures and growing seasons. High-volume crops used in subsidy verification, including rice and wheat, achieve production-grade recall. Rather than relying on a single average across all crops, performance targets are defined for each crop according to its use case. Per-crop validation results are shared under NDA.
In Japan, local agricultural committees are required by law to survey the use status of every farmland parcel each year and identify abandoned land. Inspecting every parcel in the field is labor-intensive, which can delay detection and follow-up action.
Land Use Detection classifies the land-use status of each farmland parcel from satellite imagery.
It classifies parcels as actively cultivated, abandoned, converted to non-agricultural use, or reverted to forest, producing a parcel-level status map for annual farmland-use surveys.
How It Works
Parcel-level
time-series data
Sentinel-2 time-series data is sampled for each farmland parcel polygon.
Compare current and
prior-year data
Current-year satellite data is compared with a prior-year baseline to detect changes in farmland use.
Three independent
change-detection models
Three independent detectors identify abandonment, conversion to non-agricultural use, and afforestation.
Parcel-level status
outputs
Analysis results are integrated into the application, where local agricultural committees can review the status of each parcel.
Aligned with survey schedules
The analysis window is adjusted based on delivery timing to align with each municipality’s annual farmland-use survey schedule.
In paid production use across Japanese municipalities
The technology is in paid production use across multiple municipalities in Japan.
Built for Japan’s statutory farmland-survey workflow
The system is built around the farmland-use survey workflow required under Japan’s Agricultural Land Act.
Sagri holds patents covering core technologies for farmland delineation using satellite data and AI, as well as technologies for estimating soil carbon.
AI-Based Automated Generation of
Farmland Parcel Polygons
Technology for Estimating Farmland
GHG Emissions and Supporting
Improved Farming Practices
Sagri’s models are designed and validated with the major methodologies that turn measurement into carbon credits in mind.
Our models are developed and validated through scientific research, field testing, and large-scale training datasets.
Crop Classification Training Data
1M+
Labeled parcel-year records used to train the Crop Classification model
Digital Soil Mapping Training Data
10,000+
Field soil samples used to train the Digital Soil Mapping model
Peer-Reviewed Paper
Published
The ALOS-2 SAR inundation-classification method underlying DRACO has been published in a peer-reviewed academic journal.
DRACO Validation Papers
2 Papers in Submission
Two papers validating DRACO’s water-level and GHG performance are currently in submission. Benchmark research for Digital Soil Mapping is also underway.
Research Partners
JAXA Tohoku University JIRCAS IRRI PhilRice NARO
Government & Multilateral Support
MAFF SBIR JAXA METI SBIR JCM PhilSA WFP TICAD9
No hard minimum, but the economic break-even for a Verra DSM-only carbon project sits around 5,000 ha; rice (VM0051) projects around 3,000 ha.
Smaller engagements are run as in-set tracking or supply chain (Scope 3) projects.
Yes — VT0014 requires calibration samples. The DSM stack dramatically reduces the sample count compared to a sample-only design and tells you where to sample so that each one carries the most information. Quantitative sample reduction varies by project; we benchmark per engagement.
Combination of
(a) Sagri's global covariate stack,
(b) transfer-learned models from analogous landscapes, and
(c) a targeted local sampling campaign of ~100–300 sites stratified by Sagri's sample-design tool.
Field calibration is part of every engagement — we partner with local sampling teams depending on project requirements.
Sentinel-1 and Sentinel-2 (ESA, free); Landsat-8 (USGS, free); Pleiades and SPOT 6/7 (Airbus, paid); ALOS-2 (JAXA, paid). We pay license fees per project on the customer's behalf.
Our DSM sample-level accuracy is in the same range as a leading peer-reviewed US carbon-MRV benchmark, and we additionally validate forward in time — the test that matters for credit-issuance forecasting, which most published competitor work does not report.
DRACO is process-based rather than purely statistical, so it represents the post-drainage N₂O response mechanistically — something ML-only MRV approaches cannot.
We share full per-project validation reports under NDA.
You can view the technology here:
• AgriInsight Map demo
For Governments, Traders, and Producers
From farmland visualization and crop and land-use monitoring to environmental assessment, we support a wide range of needs.
ContactFor Engineers and Researchers
Work with us on R&D using satellite data, soil analysis, and geospatial AI.
Recruit pagePlease feel free to contact us.