Login
Home FAQ
Frequently Asked Questions

Find Answers to Your Questions

Everything you need to know about Meridian Sentinel. Can't find what you're looking for?

GENERAL & OVERVIEW

What is Meridian Sentinel ?

Meridian Sentinel is a satellite-based verification infrastructure that helps African governments and institutions prevent fraud in agricultural programs. Using freely available satellite imagery and machine learning, we can verify 100% of program beneficiaries cheaper than manual verification ($50-100 per farm).

Key capabilities:

  • Detect ghost farms (zero vegetation)
  • Verify actual farm area (prevent area inflation)
  • Identify crop types (prevent crop misreporting)
  • Predict yields (estimate production)
  • Monitor crop health in real-time (early warning)

Who is Meridian Sentinel for ?

Our primary customers are :

African Governments (Ministries of Agriculture)

  • Input subsidy programs
  • Cash transfer programs
  • Crop insurance schemes
  • Agricultural extension services

Development Banks & Donors

  • World Bank agricultural projects
  • African Development Bank programs
  • USAID, DFID, EU development projects

Insurance Companies

  • Index-based crop insurance
  • Parametric insurance products
  • Claims verification

Input Suppliers & Agribusinesses

  • Distributor compliance monitoring
  • Contract farming verification
  • Supply chain traceability

What problem does Meridian Sentinel solve ?

African governments lose billions annually to agricultural programfraud. The problems :

Current Situation:

  • Manual verification costs $50-100 per farm
  • Only 5-10% of beneficiaries can be verified (budget constraints)
  • Verification takes 6-18 weeks
  • Field agents can be bribed

Fraud Types :

  • Ghost farms : 10-20% of registered farms don't exist
  • Area inflation : Farmers claim 5 hectares, actually farm1 hectare
  • Crop misreporting : Claim high-value crops (rice), plant (maize)
  • Multiple registrations : Same farmer registered in multiple districts

Result : Billions wasted, legitimate farmers can't access programs, poor policy decisions based on bad data
Our Solution : 100% verification at 80% lower cost using satellites

How is Meridian Sentinel different from other satellite agriculturecompanies ?

We're specifically built for fraud detection in African smallholder agriculture

Commercial providers typically charge between $5 and $20 per farm, focusing primarily on yield monitoring for commercial farms operating in larger fields ranging from 2 to 10 hectares. Their solutions are generally proprietary black-box models, deployed exclusively on the cloud, and are mainly used in global markets such as the USA and Europe. In contrast, Meridian Sentinel offers a much more affordable solution and is designed to detect fraud rather than just monitor yield. It specifically targets smallholder farmers managing tiny plots of land (around 0.1 to 2 hectares), with a focus on African regions. Unlike commercial providers, Meridian Sentinel uses open-source, transparent models and provides flexible deployment options, including cloud, on-premise, and hybrid setups.

Key differentiators :

  • 10-50x cheaper than alternatives
  • Purpose-built for fraud detection (not generic monitoring)
  • Trained on 19,000+ African fields (Nigeria, Ghana, Kenya)
  • Works on smallholder mixed cropping systems
  • Open-source ML models (transparent, auditable)
  • Government-friendly (data sovereignty, on-premise option)

HOW IT WORKS (TECHNICAL)

How does satellite verification work ?

5-step process :

STEP 1: Government/Executing Organisation (EO) uploads beneficiarylist

  • Required data :
  • Farmer name
  • National ID number
  • GPS coordinates (latitude, longitude) OR address
  • Declared crop type
  • Declared farm area (hectares)

STEP 2: We extract satellite imagery
Data sources :

  • Sentinel-2: 10m resolution, every 5 days (FREE fromESA)
  • Landsat 8/9: 30m resolution, every 16 days (FREE fromNASA)
  • SentinelSentinel-1 SAR: 10m, cloud-penetrating radar (FREE)

STEP 3: Machine learning analysis
10 ML models analyze each farm:

Field Boundary Detection

  • Detects actual field boundaries
  • Calculates actual area (hectares)
  • Compares to declared area
  • FLAGS if difference >20%

Crop Classification

  • Analyzes NDVI time series (vegetation patterns)
  • Identifies crop type (maize, rice, cassava, etc.)
  • Compares to declared crop
  • AnFLAGS if mismatch

Yield Prediction

  • Predicts yield (tons/hectare)
  • Uses satellite data + weather + soil
  • Compares to historical averages
  • FLAGS if suspiciously low (abandoned field)

Supporting models :

  • Planting date detection
  • Harvest date detection
  • Crop health monitoring
  • Drought/pest impact

STEP 4: Fraud flagging

Risk scoring (0-100):

  • HIGH RISK (80-100): Area inflation >50%, wrong crop, zero vegetation
  • MEDIUM RISK (50-79): Area mismatch 20-50%, low yield, late planting
  • LOW RISK (0-49): All verified, consistent with expectations
  • Ghost farm: Zero vegetation detected.
  • Area inflation: Declared 5 ha, actual 1 ha.
  • Crop misreporting: Declared rice, actual maize.
  • Abandonment: Planted but not maintained

STEP 5: Government/EO action

Dashboard shows :

  • Verified farms (90-95%): Approve automatically
  • Flagged farms (5-10%): Investigate with field visits

Government/EO investigates ONLY high-risk cases (5-10%of total)

  • 90-95% auto-verified (no field visit needed)
  • Massive cost savings

What if there's cloud cover? Can satellites see through clouds ?

Cloud cover is manageable with our multi-temporal approach :

The Challenge :

  • Tropical Africa has 50-80% cloud cover during rainy season
  • Optical satellites (Sentinel-2, Landsat) cannot see through clouds

Our Solutions :
CLOUD MASKING (Automatic)
Machine learning identifies clouds automatically :

  • Cloud detection algorithm (99% accuracy)
  • Removes cloud pixels
  • Fills gaps with temporal interpolation
  • Creates seamless composite

SYNTHETIC APERTURE RADAR (SAR) (Backup)
Sentinel-1 SAR penetrates clouds !

  • Uses radar waves (not optical light)
  • Works in any weather
  • 10m resolution
  • Good for: Crop detection, field boundaries, irrigation
  • Not as good for: Crop types (needs optical data)
  • Strategy: Use SAR when optical data insufficient

What if farmers don't have GPS coordinates ?

Multiple solutions, no problem:
OPTION 1: Address Geocoding (Most common)

Farmers provide :

  • Village/district name
  • Local landmarks
  • Directions ("500m north of village chief's house")

We use :

  • Google Maps API
  • OpenStreetMap
  • Bing Maps API
  • Local GIS databases
  • Convert addresses to GPS coordinates
  • Accuracy: 90-95% (within 100-500 meters)
  • Good enough for satellite verification (10m resolution)

OPTION 2: Mobile App (For new registrations)
Government field agents use mobile app :

  • Phone GPS captures coordinates automatically
  • Photo of farm (for verification)
  • Farmer details entered in app
  • Syncs to cloud database

OPTION 3: Map Interface (Government/EO office)
Government/EO staff use web-based map tool :

  • Zoom to village/district
  • Click on farm location
  • System records GPS coordinates
  • Farmer verifies location (SMS with map link)
  • Accuracy: 50-100 meters (depends on staff knowledge)
  • Fast: 2-3 minutes per farmer

OPTION 4: Farmer Self-Registration (Future)
Farmers use USSD/SMS :

  • Dial shortcode (e.g., *123*FARM#)
  • Phone GPS captures location
  • Farmer confirms via SMS
  • System validates and registers

Pilot Experience :

  • 30% of beneficiary lists had bad/missing GPS coordinates
  • Solution: Used address geocoding + manual correction
  • Result: 95% of addresses successfully geocoded
  • Remaining 5%: Government field agents collected GPS

Recommendation: Don't let missing GPS stop you!

  • Start with what you have (addresses)
  • We'll geocode automatically
  • Field agents fill gaps (5-10% of farms)
  • Build GPS database over time

How does the system handle mixed cropping? African farmers often plant multiple crops in the same field.

Good question! We specifically designed for this:

  • African smallholders practice intercropping (maize +beans, cassava + maize)
  • European/American models trained on monoculture (single cropper field)
  • Standard crop classification fails on mixed crops

Our Solution :
MULTI-CLASS CLASSIFICATION (Primary approach)
Instead of: "This field is maize OR beans"
We classify: "This field is maize + beans (intercrop)"

Classes in our model :

  • Maize (pure)
  • Cassava (pure)
  • Rice (pure)
  • Sorghum (pure)
  • Millet (pure)
  • Beans (pure)
  • Maize + Beans (intercrop)
  • Maize + Cassava (intercrop)
  • Cassava + Beans (intercrop)
  • Other (mixed)

SPECTRAL UNMIXING (Advanced technique)

Estimate proportion of each crop :

  • 60% maize, 40% beans
  • 70% cassava, 30% maize
  • Useful for yield estimation

Method: Analyze spectral signatures

  • Maize signature: High NIR, medium Red
  • Beans signature: Medium NIR, low Red
  • Combined: Weighted average

Fraud Detection Impact : Mixed cropping actually helps fraud detection :

  • Fraudsters typically claim pure, high-value crops (e.g., rice)
  • Real farmers often intercrop (maize + beans)
  • Mismatch between claimed and detected crop is a fraud indicator

Example :
Farmer claims: "I'm planting pure rice (high subsidy)"
Satellite detects: Maize + cassava intercropping

  • FLAG for investigation
  • Likely fraud (wrong crop claimed)

Bottom line : Mixed cropping is not a limitation—our models are trainedspecifically for this African context.

ACCURACY & RELIABILITY

What happens if the model makes a mistake ?

Multiple safeguards prevent wrong decisions :
SAFETY MECHANISM 1 : Human-in-the-Loop

We NEVER recommend automatic rejection based solely on satellite data.

Workflow :

  • Satellite flags farm as high risk (e.g., "area inflation detected")
  • Government reviews on dashboard
  • Government field agent visits farm (physical verification)
  • Agent confirms or refutes satellite finding
  • Government/EO makes final decision (approve/reject/investigate further)
  • Human always makes final decision
  • Satellite provides evidence, not verdict
  • Prevents false rejections

SAFETY MECHANISM 2 : Risk Scoring (Not Binary)
Instead of: "FRAUD" or "NOT FRAUD" (binary)

We provide: Risk score 0-100 (continuous)

Example outputs:

  • Score 95: Very high risk → Investigate immediately
  • Score 65: Medium risk → Review evidence, may investigate
  • Score 20: Low risk → Approve (likely legitimate)
  • Government/EO sets own threshold (e.g., investigate all >70)
  • Flexible based on resources available
  • Can adjust based on experience

SAFETY MECHANISM 3: Confidence Intervals
For each prediction, we provide confidence :

  • "Field area: 1.2 hectares (confidence: 85%)"
  • "Crop type: Maize (confidence: 92%)"
  • "Yield: 2.5 tons/ha (confidence: 70%)"
  • Low confidence predictions flagged for manual verification
  • Government knows when to trust satellite vs. field visit
  • Transparent about uncertainty

SAFETY MECHANISM 4: Multiple Check
We don't flag based on single criterion :
Example: Area Inflation Check

  • Declared: 5 hectares
  • Satellite detected: 1.2 hectares
  • Difference: -76% (huge discrepancy)

But we also check:

  • Crop health: Is vegetation present? (Yes)
  • Crop type: Does it match declared? (Yes, maize)
  • Historical: Has farmer been in program before? (Yes, 3 years, always legitimate)s
  • Only flag if multiple indicators point to fraud
  • Reduces false positives
  • Context-aware decisions

SAFETY MECHANISM 5 : Appeals Process
Farmers can appeal if rejected :

  • Government/EO notifies farmer of rejection
  • Farmer requests review (provide evidence)
  • Government/EO field agent re-verifies
  • If satellite was wrong, farmer approved + we retrain model
  • If farmer was fraudulent, rejection stands
  • Checks and balances
  • Protects legitimate farmers
  • Continuous improvement (learn from mistakes)

What if we find a systematic error ?
Example: Model consistently underestimates cassava area by 20%
Our response :

  • IMMEDIATE: Adjust model output (+20% correction for cassava)
  • SHORT-TERM: Retrain model with more cassava training data
  • LONG-TERM: Collect more ground truth, improve accuracy
  • COMPENSATION: Review and re-process affected cases (if farmers rejected)
  • Transparent about errors
  • Quick fixes
  • Continuous improvement

Can the system detect new types of fraud we haven't seen before?

Yes, through anomaly detection:

SUPERVISED LEARNING (Current):

We train models on known fraud types:

  • Ghost farms (zero vegetation)
  • Area inflation (boundary mismatch)
  • Crop misreporting (wrong crop)
  • Works well for known fraud patterns
  • 89-94% detection rate

Limitation :

Only detects fraud types we've seen before
New fraud schemes may slip through

UNSUPERVISED LEARNING (Advanced):

We also use anomaly detection:

Method:

Build "normal farm" profile:

  • • Typical NDVI patterns
  • • Expected area ranges
  • • Common crop combinations
  • • Historical farming behavior

Flag anything significantly different:

  • • NDVI pattern never seen before
  • • Impossibly high yield claims
  • • Suspicious location (middle of forest)
  • • Inconsistent temporal patterns
  • • Catches unknown fraud types
  • • Also catches data errors (wrong GPS coordinates)
  • • May have false positives (legitimate unusual cases)

Performance:

  • • Catches 60-70% of novel fraud
  • • 20% false positive rate (needs human review)
  • • Good for exploratory analysis

SPECIFIC NEW FRAUD TYPES WE CAN DETECT:

RENTAL FRAUD

Scheme: Fraudster rents real farm for 1 month, registers for program, returns farm after verification

Detection:

  • • Historical analysis: Was field green before?
  • • Ownership records: Does farmer match land registry?
  • • Neighbor comparison: Are surrounding farms registered to same person?
  • • Can flag suspicious cases for investigations

PHOTO FRAUD:

Scheme: Fraudster submits photos of someone else's farm

Detection:

  • • GPS mismatch: Photo GPS ≠ Declared farm GPS
  • • Temporal mismatch: Photo date outside growing season
  • • Duplicate photos: Same photo submitted for multiple farms
  • • Automatic photo forensics

IDENTITY FRAUD

Scheme: One person registers under multiple identities

Detection:

  • • GPS clustering: Multiple "different" farmers at same location
  • • Duplicate farms: Same farm boundary registered twice
  • • Cross-check with national ID database
  • • Requires integration with government systems

PROXY FARMING

Scheme: Urban elite register as "farmers" using actual farmers as proxies

Detection:

  • • Location analysis: Farmer address in city, farm in rural area
  • • Mobile money patterns: Subsidy paid to farmer, immediately transferred to someone else
  • • Cross-check: Is farmer on government payroll? (civil servants can't bebeneficiaries)
  • • Requires government data integration

TEMPORAL FRAUD

Scheme: Plant crop AFTER receiving subsidy, harvest before inspection

Detection:

  • • Planting date analysis: When did field turn green?
  • • Program timeline: Was planting before or after subsidy?
  • • CGrowing season length: Does it match crop type?
  • • Satellite temporal analysis catches this

LEARNING FROM NEW FRAUD:

When government discovers new fraud type:

  • DOCUMENTATION: Government/EO shares case details
  • ANALYSIS: We analyze satellite signatures of fraud
  • RETRAINING: Add new fraud examples to model
  • DEPLOYMENT: Updated model catches similar cases
  • RETROACTIVE: Re-scan past data for similar patterns
  • System gets smarter over time
  • Crowd-sourced fraud detection across all government partners
  • Network effect: Each government benefits from others' discoveries

FRAUD EVOLUTION EXAMPLE:

YEAR 1: Detect ghost farms (zero vegetation)

• Fraudsters adapt: Start planting real crops

YEAR 2: Detect area inflation (boundary mismatch)

• Fraudsters adapt: Claim correct area but wrong crop

YEAR 3: Detect crop misreporting (spectral analysis)

• Fraudsters adapt: Rent real farms temporarily

YEAR 4: Detect rental fraud (historical analysis)

• Fraudsters give up (too hard, not worth it)

Result: Cat-and-mouse game, but we're always learning Eventually: Fraud becomes more expensive than legitimate farming →Program succeeds.

Bottom Line:
System detects both known and unknown fraud throughcombination of supervised learning (known patterns) and anomaly detection (unusual cases). Continuous learning from all partners improves detection over time

GENERAL & OVERVIEW

What is Meridian Sentinel?

Meridian Sentinel is a satellite-based verification infrastructure...

Still Have Questions?

Our team is here to help. Get in touch with us today and we'll provide the answers you need.

Contact Us
Quick response time
Expert support team
24/7 availability