Smartphone-based Crowd-Sourced Land-Use Mapping and Validation Network

Thesis Abstract

The rapid proliferation of smartphones and location-based services has enabled scalable, participatory approaches to land-use data collection, yet current methods face challenges in spatial bias, data validation, and temporal consistency. This study addresses the problem of unreliable and spatially biased land-use maps produced from crowd-sourced inputs by proposing a Smartphone-based Crowd-Sourced Land-Use Mapping and Validation Network (SCU-LUMAN) that integrates mobile crowdsourcing with automated validation and cross-referencing against official cadastral and high-resolution satellite data. The aim is to develop a cost-effective, scalable framework for accurate, real-time land-use classification through community participation while ensuring data quality and temporal coherence. The specific objectives are (1) to design a mobile application workflow that enables residents to capture geotagged land-use observations with standardized metadata, (2) to develop an ensemble validation mechanism that combines peer-confirmation, expert verification, and automated image-based classification using convolutional neural networks, (3) to implement a data fusion model that reconciles crowd-sourced inputs with satellite imagery (Landsat 8/9 and Sentinel-2) and cadastral records, (4) to evaluate spatial and temporal accuracy using reference datasets and statistical tests, and (5) to assess user engagement and data quality dynamics over a six-month data collection period. The methodology adopts a mixed-methods research design grounded in the Theory of Planned Behavior and Social Cognitive Theory to model participation and data quality dynamics. The population comprises city-wide land-use observations from residents in a mid-sized metropolitan area with an installed a priori reference map. A stratified random sample of 600 participants will be invited, aiming for at least 400 active contributors and 2,400 validated observations per month, totaling approximately 24,000 observations over six months. Data collection instruments include a cross-platform Android/iOS application for geolocated land-use tagging, a structured metadata schema (location accuracy, timestamp, confidence level, device model, and user training level), and an integrated image capture module. Validation tools consist of a triage workflow (i) community peer-verification through a consensus scoring mechanism, (ii) expert review by urban-planning professionals, and (iii) automated classification using a CNN trained on a labeled land-use dataset and transfer learning from a pre-trained EfficientNet backbone. Additional reference data sources include high-resolution satellite imagery (?0.5 m where available), official cadastral polygons, and municipal land-use records. Data analysis proceeds in multiple layers. Descriptive statistics summarize participation, data completeness, and metadata quality. Spatial accuracy is assessed via point-in-polygon and buffer-based methods comparing crowd-derived labels with reference land-use classes, using metrics such as user’s accuracy, producer’s accuracy, and Kappa. Temporal consistency is examined through time-series cross-correlation and change-detection tests to identify label drift. The data fusion framework employs a Bayesian hierarchical model to integrate crowd input, CNN-derived predictions, and satellite/cadastral references, producing a probabilistic land-use map with quantified uncertainty. Hypotheses regarding the relationship between participant training level and data accuracy, and between validation method and label reliability, will be tested using mixed-effects logistic regression and ANOVA where appropriate. Thematic analysis will interpret qualitative feedback from participant surveys to elucidate factors influencing data quality and engagement. Expected findings include (i) improved spatial accuracy of crowd-sourced land-use labels when augmented with expert validation and CNN-based classification, (ii) demonstrable reduction in label drift through the proposed validation framework, and (iii) positive correlations between structured training, feedback loops, and data reliability. The study is anticipated to yield a scalable, transferable framework for participatory land-use mapping and real-time validation with quantified uncertainty, contributing to knowledge in GIScience, urban informatics, and participatory sensing. The theoretical contribution lies in integrating the Theory of Planned Behavior with a data-quality-oriented validation architecture and Bayesian data fusion, offering a novel model for citizen-driven geospatial information ecosystems. Practical implications include improved metropolitan land-use inventories, enhanced urban planning decision support, and a replicable protocol for similar contexts globally. Recommendations include expanding to multilingual training modules, incorporating offline data capture for connectivity-constrained environments, and establishing governance protocols for data quality assurance and user privacy. The study concludes that a carefully designed smartphone-based crowd-sourced network, augmented by automated and expert validation, can produce timely, accurate land-use maps at scale, with transparent uncertainty quantification and strong potential to inform policy and planning.