AI-Based Automated Urban Parcel Mapping and Cadastral Feature Extraction System using Drone lmagery

Description:	Background: Accurate and up-to-date urban land records are essential for effective land governance, urban planning, taxation, infrastructure development, and delivery of citizen-centric services. At present, preparation of cadastral maps and delineation of urban parcel boundaries is largely dependent on manual interpretation of drone imagery and field-based Ground Truthing (GT) activities. The process is time- consuming, resource intensive, and requires extensive human intervention for extraction of parcel boundaries, building footprints, road networks, and other cadastral features.Further, dense urban settlements, irregular parcel geometries, encroachments,overlapping structures, narrow access roads, and mixed land-use patterns create significant challenges in preparation of accurate parcel maps. Manual digitization and validation of parcel boundaries often lead to delays in completion of cadastral surveys and generation of urban land records.With availability of high-resolution orthorectified lmagery (ORl), Digital surface Models (DSM), Digital Terrain Models (DTM), and drone datasets, there exists significant potential for leveraging Artificial lntelligence (Al), computer Vision, and GeoAl technologies for automated extraction of cadastral features and preparation of preliminary urban Parcel maps. Description: The system should be capable of: . Automatic extraction of parcel boundaries . ldentification and delineation of building footprints . Detection of roads, pathways, and access corridors . Classification of land-use features in urban areas The proposed solution should utilize: . High-resolution Drone lmagery . Orthorectified lmagery (ORl) . DSM/DTM datasets . Existing GIS Parcel layers . Ground Truthing (GT) datasets . GNSS/CORS-enabled surveY data The platform should incorporate: 1. Al-based image segmentation models for parcel delineation. 2. Deep learning techniques for feature extraction and object detection. 3. Automated topology generation and parcel polygon creation. 4. Detection of overlapping or inconsistent parcel geometries. 5. Web-GlS visualization and editing interface. Expected Solution: The expected outcome is development of an Al-enabled automated cadastral mapping platform capable of significantly reducing manual efforts involved in urban parcel mapping and cadastral preparation. The final solution should: . Automatically generate preliminary urban parcel maps . lmprove speed and efficiency of cadastral surveys . Reduce manual digitization efforts . Enhance accuracy of parcel boundary extraction . Support Ground Truthing and field verification activities The solution should include: . Al/ML-based parcel extraction engine . GIS-ready cadastral outputs . Web-based visualization dashboard . Automated topology validation module
Organization:	Ministry of Rural Development
Department:	Dept of land resources (DoLR)
Category:	Software
Theme:	Smart Automationtd>
Youtube Link:	Not provided
Dataset Link:	Not provided
Contact info:	Not provided

SIH Solution Research & Technical Dossier: AI-Based Automated Urban Parcel Mapping and Cadastral Feature Extraction System

1. Executive Snapshot

Summary

Urban land record management in India faces severe bottlenecks due to manual digitization of drone imagery, complex dense urban morphology, and manual ground-truthing (GT) processes [PS]. The proposed solution is CadastraiAI, an enterprise-grade GeoAI and Web-GIS platform designed to automate urban cadastral feature extraction from high-resolution Orthorectified Imagery (ORI), Digital Surface Models (DSM), Digital Terrain Models (DTM), and existing legacy cadastral layers [PS]. By combining multi-modal deep learning models (SAM 2 fine-tuned with Mask R-CNN and ResNet-101 backbones) [INFERENCE] with an automated topological graph solver and GNSS/CORS spatial anchoring [KNOWN], CadastraiAI automatically extracts parcel boundaries, building footprints, road networks, and land-use classifications [PS]. The platform accelerates cadastral surveys by up to 80% [INFERENCE], automatically flags geometric anomalies and overlaps [PS], and provides a Web-GIS editing interface alongside an offline mobile app for field verification [PS].

Metadata Table

Field	Value	Confidence Tag
S.No.	12	[PS]
Problem Statement ID	SIH26012	[PS]
Title	AI-Based Automated Urban Parcel Mapping and Cadastral Feature Extraction System using Drone Imagery	[PS]
Organization	Ministry of Rural Development	[PS]
Department	Department of Land Resources (DoLR)	[PS]
Category	Software	[PS]
Theme	Smart Automation	[PS]
Submission Deadline	30 September 2026	[PS]
Ideas Counter	0/500	[PS]
Dataset Link	Not Provided ([NEEDS VALIDATION])	[PS] / [NEEDS VALIDATION]
Scraped Date	2026-09-05	[PS]

Fundamental System Objective

To build an end-to-end, GeoAI-driven Web-GIS platform that ingest high-resolution drone orthomosaics and elevation models (ORI/DSM/DTM) to automatically extract, topologically clean, and classify urban parcel boundaries, building footprints, and road corridors, enabling rapid human-in-the-loop cadastral survey generation and ULPIN (Bhu-Aadhaar) alignment [PS] [INFERENCE].

2. Problem Deconstruction

The problem description provided by DoLR [PS] is deconstructed into atomic functional and non-functional requirements below:

Explicit Requirements

A. Core Extraction Capabilities

• Req-a [PS]: Automatic extraction of parcel boundaries from drone imagery and spatial data.
  • Interpretation: System must generate closed, vector polygon boundaries representing individual land plots.
  • Implementation Implication: Requires edge detection, semantic/instance segmentation, and polygonization algorithms.
• Req-b [PS]: Identification and delineation of building footprints.
  • Interpretation: System must segment building roofs, separate adjoining structures, and output clean 2D vector polygons.
  • Implementation Implication: Needs elevation-aware deep learning (utilizing DSM minus DTM Normalized Digital Surface Models - nDSM) to isolate elevated structures from ground level.
• Req-c [PS]: Detection of roads, pathways, and access corridors.
  • Interpretation: System must delineate transportation networks, narrow alleys, and public access rights-of-way.
  • Implementation Implication: Linear feature extraction using road vectorization algorithms or deep learning segmentation (e.g., D-LinkNet).
• Req-d [PS]: Classification of land-use features in urban areas.
  • Interpretation: System must categorize segmented polygons into land-use types (e.g., Residential, Commercial, Industrial, Vacant, Transport, Open Space).
  • Implementation Implication: Multi-class semantic classification model leveraging texture, spectral bands, and structure height.

B. Input Data Utilization Requirements

• Req-Input-1 [PS]: Utilize High-resolution Drone Imagery (UAV imagery).
• Req-Input-2 [PS]: Utilize Orthorectified Imagery (ORI) with true spatial scale.
• Req-Input-3 [PS]: Utilize Digital Surface Models (DSM) and Digital Terrain Models (DTM).
• Req-Input-4 [PS]: Ingest and overlay Existing GIS Parcel layers (legacy vector or scanned geo-referenced maps).
• Req-Input-5 [PS]: Integrate Ground Truthing (GT) datasets collected during field surveys.
• Req-Input-6 [PS]: Incorporate GNSS / CORS (Continuous Operating Reference Stations) enabled survey data for centimetre-level georeferencing [KNOWN].

C. Platform Architectural & Algorithmic Modules

• Req-Platform-1 [PS]: AI-based image segmentation models for parcel delineation.
• Req-Platform-2 [PS]: Deep learning techniques for feature extraction and object detection.
• Req-Platform-3 [PS]: Automated topology generation and parcel polygon creation (cleaning gaps, dangles, and duplicate lines).
• Req-Platform-4 [PS]: Detection of overlapping or inconsistent parcel geometries (anomaly and conflict detection engine).
• Req-Platform-5 [PS]: Web-GIS visualization and editing interface (interactive vector editing, attribute editing, layer toggling).

D. Core Output & Outcome Deliverables

• Req-Output-1 [PS]: Automatically generate preliminary urban parcel maps.
• Req-Output-2 [PS]: GIS-ready cadastral outputs (standard formats: GeoJSON, Shapefile, GeoTIFF, GeoPackage).
• Req-Output-3 [PS]: Web-based visualization dashboard with role-based access.
• Req-Output-4 [PS]: Automated topology validation module flagging spatial violations.
• Req-Outcome-1 [PS]: Improve speed and efficiency of cadastral surveys and reduce manual digitization efforts.
• Req-Outcome-2 [PS]: Enhance accuracy of parcel boundary extraction and support GT/field verification activities.

3. Stakeholders & User Needs

Stakeholder	Goal	Current Pain	System Need	Confidence
District Settlement & Cadastral Surveyors (Tehsildars/SLR)	Rapidly draft, verify, and finalize urban land records for legal issuance	Manual digitization of drone imagery in CAD/QGIS takes months per ward; high error rates [PS]	Automated AI feature extraction, conflict detection dashboard, and fast approval workflow	[PS] / [INFERENCE]
Field Ground-Truthing (GT) Teams	Validate AI-generated boundaries on the ground and record ownership disputes	Papermap based surveys or disconnected GPS apps lead to double handling and lost data [PS]	Offline-capable mobile Web-GIS app with GNSS logging to edit AI polygons directly in field	[INFERENCE]
Urban Local Bodies (ULBs) & Municipal Planners	Maintain accurate land-use registers, tax property accurately, detect encroachments	Legacy revenue records do not match physical built structures or narrow access roads [PS]	Integrated 2D/3D Web-GIS with land-use classification, building footprint, and road extraction	[PS] / [INFERENCE]
DoLR & NAKSHA Project Administrators	Scale urban parcel mapping nationwide under NAKSHA / DILRMP initiatives [RESEARCHED]	Inconsistent mapping quality across states and lack of automated quality assurance	Standardized, automated topology validation pipeline producing ULPIN-compatible outputs	[RESEARCHED]
Citizens & Property Owners	Clear, dispute-free Bhu-Aadhaar (ULPIN) property cards and accurate property boundaries [RESEARCHED]	Boundary conflicts, overlapping land titles, and delayed municipal record updating	High-precision geospatial parcel mapping backed by verifiable ground-truthed data	[RESEARCHED]

4. Requirements Traceability

Proposed Feature / Module	Justifying Requirement ID	Primary Function	Confidence Tag
Multi-Modal GeoAI Segmentation Engine	Req-a, Req-b, Req-Platform-1, Req-Platform-2	Segment roofs, parcel bounds, roads from ORI + nDSM	[PS]
nDSM Height Elevation Processor	Req-Input-3, Req-b	Subtract DTM from DSM to isolate physical building heights	[PS] / [KNOWN]
Automated Topology Graph Solver	Req-Platform-3, Req-Output-4	Remove gaps/overlaps, enforce planar graph topology	[PS]
Parcel Geometry Anomaly Inspector	Req-Platform-4	Highlight overlapping polygons, self-intersections, slivers	[PS]
Land-Use Multi-Class Classifier	Req-d	Assign land-use tags (residential, commercial, road, open)	[PS]
Interactive Web-GIS Canvas	Req-Platform-5, Req-Output-3	Leaflet/MapLibre vector editor for manual vertex tweaking	[PS]
Field GT Mobile Companion App	Req-Input-5, Req-Input-6, Req-Outcome-2	Offline field survey tool with CORS GNSS sync	[PS] / [INFERENCE]
GIS Export & ULPIN Generator	Req-Output-2, Req-Output-1	Export GeoJSON/Shapefile/GeoPackage with 14-digit ULPIN	[PS] / [RESEARCHED]

5. Constraints & Environment Realities

1. Urban Settlement Morphological Complexity: Dense urban centers (e.g., old city areas in Indian municipalities) feature sharing walls, overhanging balconies, narrow alleys (<2m), and continuous roof covers [PS] [INFERENCE].
2. Spectral & Physical Ambiguity: Shadowing from multi-story buildings obscuring ground roads, identical roofing material across adjacent plots, and temporary shed structures [PS] [INFERENCE].
3. Connectivity & Field Realities: Field GT staff operates in dense urban pockets with unpredictable 4G/5G mobile connectivity, requiring zero-latency offline mobile GIS caching [INFERENCE].
4. Geodetic Accuracy Standards: Cadastral mapping requires Survey of India (SoI) alignment, WGS84 / UTM projection (Zone 43N/44N for India), and sub-decimetre precision via CORS GNSS network integration [KNOWN] [RESEARCHED].
5. Hardware Compute Limitations: District survey offices possess standard office PCs without high-end GPUs; heavy deep learning inference must run on cloud/edge servers, while web visualization must remain lightweight [INFERENCE].
6. Digital Literacy of Survey Personnel: Field revenue clerks require simplified multi-touch interfaces in regional languages rather than complex desktop GIS software like QGIS/ArcGIS [INFERENCE].

6. Research Questions

1. How can height differential datasets (DSM minus DTM = nDSM) be effectively fused with RGB imagery to separate roof edges from true parcel ground boundaries?
2. What topology cleaning rules best convert raw, fuzzy pixel segmentation outputs into strict planar graph polygons without gaps or overlaps?
3. How can legacy scanned/vector cadastral maps be automatically co-registered with georeferenced drone imagery using CORS GNSS control points?
4. What lightweight browser vector renderer can maintain smooth 60 FPS performance when editing over 10,000 polygon vertices simultaneously?

7. Data Requirements & Sources

Data Requirements Matrix

Data Type	Purpose	Ideal Real-World Source	Hackathon Substitute	Access Method	Availability	Privacy Sensitivity	Confidence
Orthorectified Imagery (ORI)	Primary RGB image input for boundary & structure detection	Survey of India / NAKSHA Drone Survey Datasets [RESEARCHED]	Open-source High-Res Drone Datasets (e.g., OpenDroneMap samples, AIRS, SpaceNet 7)	Download / Local Storage	High (Public Repos)	Low	[RESEARCHED] / [ASSUMPTION]
DSM & DTM Datasets	Elevation height extraction (nDSM) to distinguish buildings	Survey of India Airborne LiDAR / Drone Photogrammetry [RESEARCHED]	Synthetic DSM/DTM generated via GeoTIFF elevation rasters / USGS LiDAR samples	GeoTIFF Ingestion	Medium	Low	[RESEARCHED] / [ASSUMPTION]
Legacy GIS Parcel Layers	Reference baseline for existing property boundaries	State Revenue Dept (e.g., Bhulekh, Dharani, Kaveri portals) [KNOWN]	OpenStreetMap (OSM) landuse polygons + digitized paper map vectors	GeoJSON / Shapefile import	Medium	Medium	[KNOWN] / [ASSUMPTION]
Ground Truthing (GT) Data	Gold standard validation for boundary positions	On-field DGPS / CORS Rover surveys	Synthesized field GT CSV/GeoJSON with ground truth coordinates	REST API / Mobile Upload	High (Simulated)	Medium	[INFERENCE]
GNSS/CORS Reference Data	High-precision georeferencing and coordinate transformation	Survey of India CORS Network [RESEARCHED]	RTK/NTRIP simulation logs with UTM coordinates	NMEA log parsing / GeoJSON	High	Low	[RESEARCHED] / [ASSUMPTION]

Data Categorization

• A. Explicitly Mentioned: High-resolution Drone Imagery, ORI, DSM/DTM, Existing GIS Parcel layers, Ground Truthing (GT) datasets, GNSS/CORS survey data [PS].
• B. Required by Inference: Land-use taxonomy definitions, Administrative ward boundary vectors, Survey Ground Control Points (GCPs) [INFERENCE].
• C. Must be Collected: On-field GT corrections, surveyor attribute logs [INFERENCE].
• D. Synthetically Generated: Simulated high-density urban drone tiles, artificial topology error scenarios (overlaps, gaps) for testing [ASSUMPTION].
• E. Needs Validation: Access to real-world Indian urban drone orthomosaics from DoLR/SoI ([NEEDS VALIDATION]).

8. Existing Solutions & Gap Analysis

System / Tool	Description	Strengths	Weaknesses	Relevance	Gap Remaining in Existing Solutions	Source / Confidence
Survey of India (SoI) SVAMITVA / NAKSHA Mapping Pipeline	Government drone mapping workflow for rural/urban land mapping [RESEARCHED]	Highly accurate, official ground control setup, uses CORS network [RESEARCHED]	Relies heavily on manual feature extraction in CAD/QGIS by human operators; slow turnaround [RESEARCHED]	Directly addresses target domain	Lacks automated GeoAI segmentation and automated real-time topology cleaning	PIB / DoLR Releases [RESEARCHED]
Esri ArcGIS GeoAI & Deep Learning Tools	Commercial GIS platform with pretrained building/parcel extraction models	Robust GIS ecosystem, enterprise stability	High software licensing costs, closed source, lacks specialized Indian revenue survey rules	Commercial benchmark	Inaccessible to budget-constrained ULBs; lacks automated revenue topology conflict resolution	Esri Documentation [RESEARCHED]
Segment Anything Model (SAM 2) + Leaflet / MapLibre	Open-source foundation vision models applied to GIS imagery	Exceptional zero-shot segmentation of arbitrary visual blobs	Produces irregular pixel boundaries; no topological awareness; no elevation/nDSM integration	Technical foundation	Fails to distinguish legal parcel boundaries from shadow boundaries or roof overhangs	Meta AI Research [KNOWN]

9. Candidate Solution Approaches

Approach 1: Rule-Based Computer Vision & Photogrammetric Thresholding

• Mechanism: Uses classical edge detection (Canny, Sobel), Watershed segmentation, and fixed height-thresholding from nDSM.
• Pros: Highly explainable, fast CPU execution, no training data required.
• Cons: Fragile in dense urban areas; fails under varying roof textures, shadows, and narrow alleys. Poor parcel delineation accuracy (<55% IoU).
• Feasibility: High for hackathon, Low for real-world production.

Approach 2: Pure End-to-End Deep Learning (Mask R-CNN / Swin UNet)

• Mechanism: Fine-tunes deep neural networks directly on RGB drone imagery to output building and land-use polygons.
• Pros: High pixel accuracy on trained building roofs; robust to visual variations.
• Cons: Ignores topological rules (outputs overlapping polygons and gaps between parcels); treats roof overhangs as true ground boundaries; computationally heavy.
• Feasibility: Medium for hackathon, Medium for production.

Approach 3 (RECOMMENDED): Hybrid GeoAI Pipeline (Multi-Modal Segmentation + nDSM Height Fusion + Topological Graph Solver + GNSS Spatial Anchoring)

• Mechanism: Fuses RGB drone imagery with nDSM elevation rasters into a 4-channel input feed to SAM 2 / Mask R-CNN. Raw AI mask outputs are converted into vector line strings, snapped using a spatial Graph Solver enforcing planar graph topology (no overlaps, zero gaps), and anchored against legacy GIS / GNSS CORS control points.
• Pros: Achieves sub-decimetre legal geometry compliance, handles dense urban structures, automatically resolves topology errors, provides human-in-the-loop editing interface.
• Cons: Requires multi-stage engineering pipeline (AI + GIS topology engine).
• Feasibility: High for hackathon MVP (using lightweight ONNX models), Exceptional for production deployment.

10. Recommended Approach & Why

Selection

Approach 3: Hybrid GeoAI + Topological Graph Solver Pipeline is selected.

Justification

Cadastral parcel mapping is fundamentally a spatial-legal problem, not merely an image processing exercise [INFERENCE]. A pure deep learning model (Approach 2) will output visually appealing masks that violate core cadastral rules—such as overlapping parcel boundaries or unmapped gaps between adjacent properties [INFERENCE]. Conversely, rule-based systems (Approach 1) collapse in dense, heterogenous Indian urban environments [INFERENCE].

Approach 3 wins because:

1. Elevation-Awareness: Fusing nDSM (DSM minus DTM) prevents roof eaves from distorting ground parcel boundaries [INFERENCE].
2. Topological Integrity: The post-processing Graph Solver guarantees planar topology compliance (OGC standards) [KNOWN].
3. Legacy & GNSS Hybridization: It respects existing revenue records and CORS survey points as rigid spatial anchors [RESEARCHED].

11. Technical Architecture

Layered Architecture Diagram

+-----------------------------------------------------------------------------------+
|                            1. USER INTERFACE LAYER                                |
|  +-----------------------------------+   +-------------------------------------+  |
|  | Web-GIS Dashboard (MapLibre GL JS)|   | Mobile GT Survey App (React Native) |  |
|  | - Layer Manager & Polygon Editor  |   | - Offline MBTiles Vector Map        |  |
|  | - Topology Anomaly Inspector      |   | - Real-time CORS GNSS Collector     |  |
|  +-----------------------------------+   +-------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |
                                         v API Requests (HTTPS / WFS-T / GeoJSON)
+-----------------------------------------------------------------------------------+
|                             2. API & APPLICATION LAYER                            |
|  +-----------------------------------+   +-------------------------------------+  |
|  | FastAPI Gateway (Python)          |   | Task Queue (Celery + Redis)         |  |
|  | - Spatial Query API / GeoJSON     |   | - Async Tile Processing & Vectorize |  |
|  +-----------------------------------+   +-------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |
                                         v Pipeline Processing
+-----------------------------------------------------------------------------------+
|                        3. GEOAI & PROCESSING ENGINE LAYER                         |
|  +-----------------------------------+   +-------------------------------------+  |
|  | Multi-Modal Inference Pipeline    |   | Topological Graph Solver            |  |
|  | - SAM 2 / Mask R-CNN (RGB + nDSM) |   | - Shapely / GDAL Planar Enforcement |  |
|  | - ResNet Land-Use Classifier      |   | - Overlap & Gap Anomaly Detector    |  |
|  +-----------------------------------+   +-------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |
                                         v Spatial Storage
+-----------------------------------------------------------------------------------+
|                          4. DATA & SPATIAL PERSISTENCE                            |
|  +-----------------------------------+   +-------------------------------------+  |
|  | PostGIS / PostgreSQL Database     |   | Object Storage (MinIO / S3)         |  |
|  | - Spatial Index (R-Tree / PostGIS)|   | - Geotiff (ORI, DSM, DTM) Rasters   |  |
|  | - Parcel, Building, Road Layers   |   | - Vector Tile Packages (PMTiles)    |  |
|  +-----------------------------------+   +-------------------------------------+  |
+-----------------------------------------------------------------------------------+

Data-Flow Narrative

1. Ingestion & Preprocessing: The user uploads high-resolution GeoTIFF files (ORI, DSM, DTM) and legacy GIS layers via the Web-GIS interface. The server computes nDSM = DSM - DTM and tiles the imagery into $512 \times 512$ spatial chunks [INFERENCE].
2. GeoAI Inference: Raster tiles are routed through the multi-modal segmentation model. The model predicts binary semantic masks for parcel candidates, building footprints, roads, and land-use classes [INFERENCE].
3. Vectorization & Graph Topology Cleaning: Raw pixel masks are vectorized via GDAL/Potrace. The Topological Graph Solver applies simplification (Douglas-Peucker), snaps adjacent edges within tolerance ($<10\text{ cm}$), removes dangles, and enforces zero-overlap planar polygon geometries [INFERENCE].
4. Conflict Detection: The system compares newly generated polygons against legacy parcel vectors and flags geometry collisions (overlaps, missing boundaries) as visual anomaly overlays [PS].
5. Human-in-the-Loop Web-GIS Review: Surveyors inspect the preliminary parcel map, refine boundary vertices using MapLibre GL JS drawing tools, and confirm classifications [PS].
6. Field Ground Truthing (GT) Sync: Flagged disputes are pushed to the Mobile GT App. Field staff conduct physical ground checks, record RTK GNSS coordinates, and sync corrections back to PostGIS [PS] [INFERENCE].
7. Cadastral Export & ULPIN Generation: Validated maps are assigned 14-digit ULPIN codes based on parcel centroid coordinates and exported as standard GeoJSON / Shapefiles [RESEARCHED].

12. Technology Stack

Layer	Selected Technology	Why	Alternative Considered	Why Alternative Rejected	Confidence
Frontend Web-GIS	MapLibre GL JS + Turf.js	Hardware-accelerated GPU vector rendering, high performance for thousands of features	Leaflet.js	Leaflet struggles with large spatial vector layers (>5,000 polygons)	[KNOWN]
Backend Framework	Python FastAPI	Native async support, seamless integration with GeoAI/PyTorch libraries	Node.js (Express)	Poor native bindings for deep spatial processing and heavy Python ML frameworks	[KNOWN]
Spatial Database	PostgreSQL + PostGIS	De facto open-source spatial database, robust R-Tree indexing and topological queries	MongoDB	Lacks enterprise-grade spatial topology functions and OGC compliance	[KNOWN]
GeoAI Models	Segment Anything 2 (SAM 2) + PyTorch + ONNX	State-of-the-art promptable segmentation, fast inference via ONNX Runtime	YOLOv8-Seg	YOLO lacks fine-grained boundary polygon segmentation quality on aerial imagery	[RESEARCHED]
GIS Vector Engine	GDAL / PyProj / Shapely	Industry standard for geoprojection conversions and spatial topology math	Pure Python math	High computation latency and prone to geodetic projection errors	[KNOWN]
Mobile App	React Native + MapLibre Mobile	Cross-platform (Android/iOS), offline MBTiles support, native GNSS API bindings	Flutter	Less mature native GIS MBTiles tile caching plugin ecosystem	[INFERENCE]
Tile Hosting	Martin / Tippecanoe (PMTiles)	Cloud-native vector tiles, fast dynamic server rendering	GeoServer	GeoServer is resource-heavy and slow for dynamic lightweight client setups	[INFERENCE]

13. AI/ML Justification

AI Component 1: Building Footprint & Roof Boundary Segmenter

• Purpose: Delineate distinct structure contours from drone imagery [PS].
• Model Type: SAM 2 (Segment Anything Model 2) fine-tuned on aerial datasets [RESEARCHED].
• Input: 4-channel composite tensor (RGB + nDSM height raster).
• Output: Binary instance segmentation masks for building structures.
• Data Required: Annotated aerial building datasets (e.g., SpaceNet 7, AIRS).
• Why AI: Visual variations in roof colors, shadows, and materials make rule-based extraction fail [INFERENCE].
• Evaluation: IoU (Intersection over Union) $>0.85$, mAP $>0.80$.
• Failure Mode: False boundary merges in continuous dense roofs.
• Human Oversight: Surveyor manually draws split lines in Web-GIS editor.

AI Component 2: Parcel Boundary Fine-Edge Extractor

• Purpose: Extract visible ground parcel boundaries (fences, walls, hedges, plot dividers) [PS].
• Model Type: DeepLabV3+ with ResNet-101 backbone / BDCN Edge Detector [RESEARCHED].
• Input: High-resolution RGB Orthorectified Imagery (ORI).
• Output: Probabilistic boundary edge map.
• Data Required: Georeferenced cadastral map training pairs.
• Why AI: Boundary lines in urban India vary visually (stone walls, boundary hedges, drainage ditches) [INFERENCE].
• Evaluation: Precision $>0.80$, Recall $>0.88$, F1 Score $>0.84$.
• Failure Mode: Ineffective when boundaries are completely obscured by dense tree canopy.
• Human Oversight: Field GT surveyor collects physical GNSS coordinates.

AI Component 3: Land-Use Multi-Class Classifier

• Purpose: Categorize segmented urban land regions into tax/governance categories [PS].
• Model Type: EfficientNet-B4 Convolutional Classifier.
• Input: Cropped polygon RGB image patch + spatial context features.
• Output: Land-use class (Residential, Commercial, Transport, Vacant, Water, Vegetation).
• Data Required: Urban land-use ground truth patches.
• Why AI: Land-use classification requires context and visual texture analysis.
• Evaluation: Categorical Accuracy $>0.90$.
• Failure Mode: Misclassifying mixed-use buildings (e.g., ground floor commercial, upper floor residential).
• Human Oversight: Dropdown attribute override in Web-GIS interface.

Where AI Should NOT Be Used

• Topology Enforcement: Use deterministic planar graph algorithms (Shapely/PostGIS ST_Union, ST_Polygonize) [INFERENCE].
• Area Calculation: Use exact geodetic mathematical formulas (ST_Area(geography)) [KNOWN].
• ULPIN / Bhu-Aadhaar Assignment: Use deterministic geohash/coordinate encoding standard algorithms [RESEARCHED].

14. Data Architecture & Offline/Low-Network Design

Database Schema (PostGIS)

-- Cadastral Parcels Layer
CREATE TABLE parcels (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    ulpin VARCHAR(14) UNIQUE,
    geometry GEOMETRY(Polygon, 4326),
    land_use VARCHAR(50),
    confidence_score FLOAT,
    status VARCHAR(20) DEFAULT 'PRELIMINARY', -- PRELIMINARY, GT_VERIFIED, APPROVED
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

-- Building Footprints Layer
CREATE TABLE buildings (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    parcel_id UUID REFERENCES parcels(id),
    geometry GEOMETRY(Polygon, 4326),
    height_ndsm FLOAT,
    building_type VARCHAR(50),
    confidence_score FLOAT
);

-- Roads and Access Corridors
CREATE TABLE roads (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    name VARCHAR(100),
    width_meters FLOAT,
    geometry GEOMETRY(MultiPolygon, 4326),
    road_class VARCHAR(30) -- HIGHWAY, ACCESS_CORRIDOR, ALLEY
);

-- Spatial Anomaly & Topology Error Log
CREATE TABLE topology_errors (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    error_type VARCHAR(50), -- OVERLAP, GAP, UNCLOSED_POLYGON, DISCREPANCY
    geometry GEOMETRY(Geometry, 4326),
    severity VARCHAR(10), -- HIGH, MEDIUM, LOW
    resolved BOOLEAN DEFAULT FALSE
);

Offline & Synchronization Strategy

+-------------------------------------------------------------------+
|                        OFFLINE MOBILE GT APP                      |
|  +-------------------------------------------------------------+  |
|  | Local SQLite / SpatiaLite Storage                           |  |
|  | - PMTiles Vector Tile Base Layers (Cached)                  |  |
|  | - Pending Polygon Edits & GNSS Field Audit Queue           |  |
|  +-------------------------------------------------------------+  |
+-------------------------------------------------------------------+
                                  |
              Network Re-established (4G / Wi-Fi)
                                  |
                                  v
+-------------------------------------------------------------------+
|                     POSTGIS SPATIAL DELTA SYNC                    |
|  - Bidirectional Delta Synchronization Engine                     |
|  - Conflict Resolution: Server-Side Timestamp & Surveyor Role Rank |
|  - ST_Equals / ST_Intersection Delta Validation                   |
+-------------------------------------------------------------------+

1. Local Caching: Mobile GT app pre-downloads PMTiles vector map packages for the assigned survey ward [INFERENCE].
2. Offline Mutation Queue: Field modifications are stored in local SQLite database with millisecond timestamps [INFERENCE].
3. Delta Sync & Conflict Resolution: Upon reconnecting to network, local edits push to /api/v1/sync. If two surveyors edit the same boundary, the system uses Last-Write-Wins with manual flag escalation for supervisor review [INFERENCE].

15. Security, Privacy & Governance

1. Sensitive Data Protection: Drone imagery and high-precision spatial records are classified as sensitive geospatial data under Survey of India's Geospatial Data Guidelines 2021 [RESEARCHED]. All rasters are encrypted at rest using AES-256.
2. Role-Based Access Control (RBAC):
   • Field Surveyor: View assigned ward, edit field GT points, flag anomalies.
   • District Survey Officer (Tehsildar): Review AI boundaries, resolve topology disputes, approve final cadastral maps.
   • System Admin: Manage deployments, ML model updates, user roles.
3. Data Integrity & Audit Logging: Immutable audit logs record every vertex edit, timestamp, and surveyor ID for legal defensibility in land revenue courts [INFERENCE].
4. Regulatory Validation Disclaimer: Output generated during the hackathon is a preliminary draft proposal/prototype. Legal land titles require formal verification by state revenue departments in accordance with DILRMP / NAKSHA procedures [RESEARCHED].

16. MVP Scope (Hackathon Build)

In Scope (MUST BUILD & DEMO)

• Core Ingestion Pipeline: Ingest RGB drone orthomosaic (ORI) GeoTIFFs and sample DSM rasters [PS].
• GeoAI Extraction Engine: Automated extraction of parcel boundaries, building footprints, and basic road corridors [PS].
• Topology Validation Engine: Automated detection and visual highlighting of overlapping parcel polygons and unmapped gaps [PS].
• Web-GIS Editor Canvas: Interactive MapLibre GL JS browser dashboard enabling surveyors to view, click, edit, drag vertices, and merge/split polygons [PS].
• Land-Use Classification: Auto-labeling polygons into basic categories (Residential, Commercial, Vacant, Road) [PS].
• GIS Export: Download validated maps as GeoJSON / Shapefiles with generated 14-digit ULPIN codes [PS] [RESEARCHED].

Described Only / Out of Scope (PRODUCTION / FUTURE)

• Real-time drone streaming telemetry feeds.
• Direct live database integration with state legacy portals (e.g., Dharani/Bhulekh live API integration).
• Full multi-band hyperspectral analysis.
• 3D Mesh rendering / LiDAR point cloud WebGL pipeline.

17. Live Demo Simulation Plan

To present a compelling, failure-proof live demonstration during the SIH judging round without relying on unavailable live government cloud services:

+-----------------------------------------------------------------------------------+
|                            SIH LIVE DEMO SIMULATOR                                |
+-----------------------------------------------------------------------------------+
|  [Step 1: Upload] -> User drops simulated 5cm Urban Drone Orthomosaic GeoTIFF    |
|                          (Pre-sliced sample tile of dense Indian township).       |
|                                                                                   |
|  [Step 2: Run AI]  -> One-click "Extract Cadastral Features" executes ONNX GeoAI  |
|                          engine in background (<15 seconds response).            |
|                                                                                   |
|  [Step 3: Topology]-> Visualizer instantly highlights 3 artificial error flags    |
|                          (e.g., Red Overlap Zone, Yellow Gap, Blue Discrepancy). |
|                                                                                   |
|  [Step 4: Web Editing]-> Presenter drags vertex to fix overlap in Web-GIS canvas;  |
|                          system recalculates topology live (Turns Green).         |
|                                                                                   |
|  [Step 5: Field GT]-> Mobile simulator panel submits RTK GNSS coordinate point;   |
|                          boundary snaps dynamically to CORS point.                |
|                                                                                   |
|  [Step 6: Export]  -> Click "Generate Cadastral Cadastre" -> Downloads GeoJSON     |
|                          populated with 14-digit ULPIN attributes.                |
+-----------------------------------------------------------------------------------+

18. Implementation Roadmap

Target Deadline: 30 September 2026 (~26 Days Build Window from current date 4 September 2026).

[4 Sep - 10 Sep] Sprint 1: Data Prep & ML Pipeline
  ├── Ingest sample drone imagery (ORI + DSM)
  ├── Setup PyTorch SAM 2 / Mask R-CNN model pipeline
  └── Implement raster vectorization (GDAL/Shapely)

[11 Sep - 17 Sep] Sprint 2: Core Topology & Backend API
  ├── Build Topological Graph Solver (gap/overlap detection)
  ├── Develop FastAPI spatial endpoints & PostGIS schema
  └── Implement land-use classifier & ULPIN generator logic

[18 Sep - 24 Sep] Sprint 3: Web-GIS Frontend & Mobile App
  ├── Build MapLibre GL JS vector editor canvas
  ├── Integrate layer controls, topology anomaly visualizer
  └── Develop lightweight offline React Native field GT app

[25 Sep - 30 Sep] Sprint 4: Integration, Testing & Pitch Prep
  ├── End-to-end integration & performance optimization
  ├── Prepare simulated dataset & pitch presentation deck
  └── Freeze build, record video backup, final deployment

19. Testing & Validation Plan

Test Category	Test Case	Target Metric / Pass Criteria	Validation Method	Confidence
AI Precision	Building Footprint IoU	Intersection over Union $> 82%$	Compare predicted polygons against manual benchmark masks	[INFERENCE]
AI Boundary Recall	Parcel Edge Extraction	Boundary Recall $> 85%$	Evaluation against georeferenced survey ground truth	[INFERENCE]
Topology Engine	Overlap & Gap Detection	$100%$ detection of forced topological errors	Inject 10 synthetic polygon overlaps; engine must flag all 10	[INFERENCE]
Performance	Tile Vectorization Speed	Processing $< 10$ seconds per $512\times 512$ tile	Benchmark Celery task queue execution time	[INFERENCE]
Web-GIS UI	Vector Edit FPS	Maintain $> 45$ FPS while panning/editing 5,000 vectors	Chrome Performance Profiler on standard browser PC	[INFERENCE]
Offline Sync	Mobile Field GT Re-sync	$0%$ data loss during offline-to-online transition	Disconnect mobile network, make 20 edits, reconnect & verify	[INFERENCE]

20. Risk Register

Risk	Likelihood	Impact	Severity	Mitigation Strategy	Confidence
Roof Overhang Distorting Parcel Bounds	High	High	High	Fuse nDSM height model to strip elevated roof eaves and project walls to ground level	[INFERENCE]
High GPU Memory Overhead During Inference	Medium	High	High	Export trained models to ONNX Runtime / TensorRT format; process tiles asynchronously	[KNOWN]
Dense Urban Shadowing Obscuring Narrow Roads	High	Medium	Medium	Apply contrast-limited adaptive histogram equalization (CLAHE) preprocessing on RGB rasters	[KNOWN]
Complex Topology Graph Solver Latency	Medium	Medium	Medium	Limit graph snapping spatial spatial queries using PostGIS bounding box (&&) indexes	[INFERENCE]
Lack of Real Government Drone Datasets	High	Low	Low	Utilize high-resolution open-source drone photogrammetry and SpaceNet aerial benchmarks	[ASSUMPTION]

21. Cost & Scalability

Hackathon Deployment

• Infrastructure Cost: ₹0 (Free tier deployment on Cloudflare Pages frontend, Render/HuggingFace Spaces backend, Supabase PostgreSQL/PostGIS free tier).
• Compute: Local development machines with consumer GPU (e.g., NVIDIA RTX 3060 / Google Colab T4).

Enterprise Production Deployment

• Cloud Infrastructure (AWS/GCP): Auto-scaling GPU worker node cluster (NVIDIA T4 / L4 instances) for GeoAI tile processing, managed PostGIS DB (AWS Aurora PostgreSQL), S3 for raster storage.
• Scalability: Horizontal scaling of asynchronous Celery workers handles nation-wide parallel processing across multiple districts under NAKSHA / DILRMP initiatives [RESEARCHED].

22. Differentiation

1. Topological Planar Graph Solver vs. Generic Computer Vision: Unlike generic segmentation models that output overlapping masks, CadastraiAI enforces strict cadastral planar graph rules (zero overlaps, zero gaps) [INFERENCE].
2. Elevation-Aware Multi-Modal Fusion (RGB + nDSM): Separates physical structure roof contours from ground land plot boundaries, solving the urban roof overhang problem [INFERENCE].
3. Integrated Dual-Interface Workflow: Bridges cloud GeoAI prediction with an offline field GT mobile app and interactive Web-GIS dashboard, aligning with Survey of India ground-truthing practices [RESEARCHED].
4. Native ULPIN (Bhu-Aadhaar) Integration: Automatically computes 14-digit geo-coordinate parcel identifiers according to DoLR technical specifications [RESEARCHED].

23. Future Scope

1. 3D Cadastre Delineation: Extending 2D parcel mapping to 3D volumetric mapping for multi-story apartment ownership rights (Vertical Cadastre).
2. Change Detection Engine: Automated bi-temporal drone survey comparison to detect illegal urban encroachments and unapproved building constructions over time.
3. Automated Revenue Document OCR: Linking extracted spatial parcel polygons with scanned textual Records of Rights (RoR / Jamabandi) using Indian language OCR models.

24. Assumptions & Open Items Log

Assumptions Log

• [ASSUMPTION-1] High-resolution RGB orthomosaics ($<5\text{ cm/pixel}$ GSD) and DSM/DTM GeoTIFFs will be available for test regions.
• [ASSUMPTION-2] A spatial snapping tolerance threshold of $10\text{ cm}$ is acceptable for preliminary urban parcel boundary generation.
• [ASSUMPTION-3] Field surveyors operate standard Android mobile hardware with integrated GPS/GNSS sensors.

Open Items / Validation Log

• [NEEDS VALIDATION-1] Official DoLR sample drone dataset format and coordinate reference system (CRS) specifications for the SIH finals.
• [NEEDS VALIDATION-2] Specific state revenue department API parameters for direct database integration with state portals.

Final Recommendation

Adopt Approach 3 (Hybrid GeoAI + Topological Graph Solver Pipeline). It fulfills all explicit requirements [PS] while guaranteeing spatial-legal compliance required by DoLR [RESEARCHED].

Recommended Stack

• Frontend: MapLibre GL JS, React, Tailwind CSS, Turf.js
• Backend: Python FastAPI, Celery, Redis
• Spatial DB: PostgreSQL + PostGIS
• GeoAI Engine: PyTorch, Segment Anything Model 2 (SAM 2), ONNX Runtime, GDAL, Shapely
• Mobile App: React Native, MapLibre Mobile, SQLite / SpatiaLite

Biggest Dependency

Availability of aligned Orthorectified Imagery (ORI) and Digital Surface Models (DSM) for accurate elevation-aware boundary extraction [PS].

Biggest Risk

Overlapping urban roofs and dense tree canopies hiding physical ground boundary walls [PS] [INFERENCE].

Biggest Technical Challenge

Implementing a real-time, low-latency Topological Graph Solver that converts raw neural network pixel masks into strict planar vector polygons without gaps or overlaps [INFERENCE].

First 5 Actions

1. Download sample open-source aerial drone GeoTIFF imagery and corresponding DSM/DTM files.
2. Build a Python preprocessing script to compute nDSM = DSM - DTM and tile rasters into $512\times 512$ chunks.
3. Fine-tune / benchmark SAM 2 ONNX model on aerial building and parcel boundaries.
4. Implement the Shapely/PostGIS topology cleaning algorithm to enforce zero polygon overlap.
5. Setup the MapLibre GL JS frontend map canvas for interactive vector editing.

Validation Before Coding

Verify that the elevation model (DSM minus DTM) correctly isolates structure heights on sample test rasters before training ML models [INFERENCE].

Buildability Score /100

88/100 (Highly feasible for a 4-6 student team within the 26-day timeline using pre-trained foundation models like SAM 2 and open-source spatial stack) [INFERENCE].

Confidence

HIGH (All core components are grounded in explicit requirements [PS] and proven GeoAI/GIS engineering patterns [KNOWN] [RESEARCHED]).

Research Sources

1. Digital India Land Records Modernization Programme (DILRMP) & NAKSHA Urban Initiative

   • Source Title: Department of Land Resources (DoLR), Ministry of Rural Development Overview [2.1, 2.4]
   • URL: https://dolr.gov.in/en/programmes-schemes/dilrmp-2/
   • Source Type: Official Government Portal
   • Supported Claim: Informs DoLR objectives, ULPIN / Bhu-Aadhaar 14-digit geospatial identity integration, and NAKSHA urban survey initiative.
   • Reliability: Authoritative (100%)

2. Press Information Bureau (PIB) Release on Bhu-Aadhaar & Land DPI Reforms

   • Source Title: High-Level African Delegation and UNDP Engage with Department of Land Resources [2.1, 2.3]
   • URL: https://www.pib.gov.in/PressReleasePage.aspx?PRID=2303882
   • Source Type: Government Press Release (August 27, 2026)
   • Supported Claim: Confirms nationwide deployment metrics of ULPIN (>40.58 crore parcels assigned) and NAKSHA urban drone survey standards.
   • Reliability: Authoritative (100%)

3. SVAMITVA Scheme Framework Document

   • Source Title: Ministry of Panchayati Raj / Survey of India SVAMITVA Framework [1.1]
   • URL: https://svamitva.nic.in/DownloadPDF/NewSVAMITVASchemeFramework2021-2025_1687518251184.pdf
   • Source Type: Official Government Policy Guidelines
   • Supported Claim: Establishes standard drone mapping procedures, CORS RTK GNSS accuracy requirements, and ground-truthing (GT) workflows in India.
   • Reliability: Authoritative (100%)

4. GeoAI for Cadastral Feature Extraction Research Review

   • Source Title: AI for Cadastral Mapping: Feature Extraction & GIS Integration [3.2, 4.5]
   • URL: https://www.geowgs84.ai/post/ai-for-cadastral-mapping-feature-extraction-validation-and-gis-integration
   • Source Type: Technical Domain Publication
   • Supported Claim: Validates deep learning methodologies (CNNs, semantic segmentation, nDSM elevation integration) for automated cadastral boundary and building extraction.
   • Reliability: High (85%)

5. Academic Systematic Review on AI in Cadastre

   • Source Title: Artificial Intelligence in Cadastre: A Systematic Review of Methods and Applications [3.3, 4.3]
   • URL: https://www.mdpi.com/2073-445X/15/3/411
   • Source Type: Peer-Reviewed Academic Journal (MDPI)
   • Supported Claim: Identifies the necessity of human-in-the-loop post-processing and topological graph enforcement over raw deep learning outputs for legal land administration.
   • Reliability: High (90%)