← Back to Projects
AI Framework · LADM · Land Administration

Intelligent Land Registration for Nepal: A Trustworthy AI Framework for Faster, Accurate, and Inclusive Cadastre & Registry Services

Land RegistrationArtificial Intelligence Cadastral MappingDocument AI LADM · ISO 19152Human-in-the-Loop Fit-for-Purpose
📅 Ongoing · 2025 🏛️ Kathmandu University · Geomatics Eng. 📄 Research Paper

01 Abstract

Land registration is a foundational public service linking people, parcels, and legally recognized rights. In Nepal, legacy paper archives, inconsistent document formats, and mismatches between cadastral maps and ground conditions increase transaction time and weaken land governance. This paper proposes a Nepal-relevant, modular AI framework aligned with fit-for-purpose land administration principles: affordability, participatory verification, incremental upgrading, and clear accountability. The framework integrates computer vision for boundary extraction, document AI for record processing, and anomaly detection for internal quality control - all within a human-in-the-loop design that preserves legal accountability.

02 The Problem

Nepal's land administration faces a classic tension between precision and coverage. Conventional cadastral approaches - high-accuracy surveys, intensive adjudication, engineered databases - require time and capacity difficult to sustain at national scale. Where coverage is incomplete, informal transactions increase, disputes persist, and the registry loses legitimacy.

The operational reality includes high service pressure at land revenue offices, legacy paper archives, inconsistent document formats across regions, and growing mismatches between cadastral maps and ground conditions in peri-urban areas. Post-disaster (earthquake, landslide), the demand for property documentation spikes precisely when the system is under stress.

03 AI Framework - Five Modules

The framework organizes AI into five bounded modules, each solving a specific problem and producing outputs that can be validated independently. This modular approach reduces operational risk and simplifies adoption in government settings.

🛰️
Boundary Extraction
Mask R-CNN and fully convolutional networks applied to UAV/satellite imagery to generate candidate boundary lines. Outputs treated as candidate evidence, not legal boundaries.
🗺️
Cadastral Mismatch Detection
Candidate boundaries compared with existing cadastral geometry to detect overlaps, gaps, shifts. Supports prioritization of field verification resources in fast-changing peri-urban zones.
📄
Document Classification & Extraction
LayoutLM and Donut (OCR-free transformer) classify land records and extract structured fields - parcel IDs, party names, transaction types - from scanned archives, even with degraded scan quality.
🔗
Entity Resolution & Record Linkage
Probabilistic similarity scoring links records across spelling variants and transliteration differences. All accepted matches logged with audit trails and reversible correction procedures.
🛡️
Quality Assurance & Anomaly Detection
Topology checks (no overlaps/gaps), attribute validation, workflow completeness checks, and transaction log anomaly flags. Flags are risk indicators, not accusations - reviewed with due process.

04 LADM Alignment & Data Framework

The framework is conceptually grounded in the Land Administration Domain Model (LADM · ISO 19152) - the international standard for representing relationships between people, parcels, and rights. AI modules produce candidate evidence that can be inserted into LADM-aligned database fields only after human verification, preserving legal meaning.

Four categories of data materials are identified: spatial imagery (satellite, aerial, UAV orthophotos), cadastral and base-map layers (parcel polygons, control points, admin boundaries), documents (ownership certificates, deeds, mutation records), and transaction logs (mutation events, edit histories).

05 Key Risks & Controls

RiskHow it AppearsControl
Boundary InvisibilityAI misses socially defined boundaries, produces misleading candidate linesParticipatory verification; treat outputs as candidates only
Extraction ErrorsWrong party names or parcel IDs extracted; rework increasesHuman confirmation gate; source region evidence; continuous monitoring
Bias & ExclusionLower performance in informal settlements or minority language contextsDiverse validation sets; fairness audits; complaint channels
Privacy BreachUnauthorized access to personal and property dataRole-based access; encryption; data minimization policy
Automation OverreachStaff accept AI outputs without review; legal harm embeddedMandatory review gates; training; reversible correction procedures

06 Nepal-Specific Context

Nepal's land administration operates in a mixed digital–paper environment. Centralized web-based transaction systems (LRIMS) already exist, providing integration points for AI as assistive services - without replacing legal decision-making. Documents exist in Nepali script and mixed languages, requiring training data and validation procedures aligned with local forms.

The recommended adoption sequence: 1) Document AI for triage and indexing (lowest risk, highest return). 2) Imagery-based mismatch detection in selected municipalities. 3) Entity resolution and anomaly support as internal QA tools. All stages require mandatory verification gates and staff training.