Development of a Low-Cost UAV-based Land Use Classification System: Design, Implementation, Evaluation

Thesis Abstract

The rapid advancement of affordable unmanned aerial vehicles (UAVs) and open-source geoinformatics tools has energized the potential for high-resolution land-use classification at scales suitable for local planning and environmental monitoring, yet the practical deployment of reliable, low-cost UAV-based systems remains constrained by data quality, processing workflows, and transferability across heterogeneous landscapes. This study addresses the problem of achieving accurate land-use classification using a cost-effective UAV platform, with an emphasis on designing an end-to-end workflow that balances hardware affordability, methodological rigor, and evaluative robustness for real-world applications in urban fringe and peri-urban settings. The aim is to develop, implement, and evaluate a low-cost UAV-based land-use classification system that delivers parcel- and plot-level insights suitable for district-level decision making. The specific objectives are (i) to design a modular data acquisition and processing pipeline using off-the-shelf UAV hardware (DJI Mini 3 Pro) and open-source software (OpenDroneMap, QGIS, and Python-based scikit-learn pipelines); (ii) to implement a multi-spectral and RGB imaging protocol with calibrated radiometric corrections and standardized flight planning to enhance feature extraction for land-use classes including residential, commercial, agriculture, water, forest, and transportation; (iii) to develop and compare machine-learning classifiers (Random Forest, Support Vector Machines, and Gradient Boosting) for pixel- and object-based approaches, integrating texture, spectral, and contextual features; (iv) to validate classification outputs against ground-truth data collected through stratified random sampling of 500 plots with in-situ surveys and high-resolution reference data, and to assess transferability across two study sites with differing land-use patterns; and (v) to evaluate system performance against conventional high-cost UAV and satellite-derived baselines using accuracy metrics, kappa statistics, and cost-benefit analysis. The methodology adopts a pragmatic sequential mixed-methods design. The population comprises land-use parcels within two demographically distinct districts. A stratified random sample of 500 ground-truth plots (250 per site) is collected through field surveys and high-resolution orthophotos, with each plot annotated for primary land-use category. Data collection instruments include a calibrated consumer-grade UAV platform with a multispectral camera, a temporal flight schedule spanning pre- and post-vegetation periods, ground-truth survey forms, and labeling tools in ENVI/GRASS for reference datasets. The validity and reliability of instruments are ensured through radiometric calibration using reflected-light targets, cross-validation of ground-truth labels by two independent observers, and inter-class labeling consistency checks. Data analysis employs a two-stage process (1) feature extraction and image pre-processing (radiometric correction, DSM/orthomosaic generation, and object-based image analysis with segmentation) to produce a feature set of spectral indices (NDVI, NDBI, NDWI), texture measures (GLCM features), and contextual attributes; (2) supervised learning model comparison with nested cross-validation to prevent overfitting, followed by statistical comparison using McNemar’s test and paired t-tests for accuracy differences. An analytical framework based on the Theory of Planned Behavior is integrated to interpret adoption and deployment outcomes among local planning practitioners. Hypotheses test whether the low-cost UAV-derived classifier achieves non-inferior overall accuracy relative to the high-cost reference system and whether feature-rich, object-based approaches outperform pixel-based methods in heterogeneous landscapes. Expected findings indicate that the proposed system attains overall classification accuracies above 85% with a kappa exceeding 0.80 across study sites, with negligible degradation when transferred between sites due to robust texture and contextual features. The research is anticipated to reveal that radiometric calibration and object-based segmentation substantially improve discrimination among closely related land-use classes, such as mixed-use and residential areas. The study contributes to knowledge by delivering a validated, replicable, low-cost methodological blueprint for UAV-based land-use classification, including open-source processing pipelines, a replicable ground-truth framework, and transferable performance metrics. It also provides practical guidance on cost-benefit trade-offs for local governments and researchers seeking affordable yet rigorous geospatial decision-support tools. The main conclusion will articulate the feasibility of deploying a low-cost UAV-based land-use classification system for urban and peri-urban applications, with recommendations for improving classifier generalizability, integrating temporal analysis for dynamic land-use monitoring, and extending the framework to multi-temporal and multi-sensor datasets to enhance policy-relevant land-use intelligence.