Princecharles C. Anyadiegwu (AUTHOR)
Department of Surveying and Geoinformatics, Imo State University, Owerri, Nigeria
Corresponding author: cpcanyadiegwu@gmail.com
ABSTRACT: Reliable subnational land-cover statistics are essential for spatial planning, infrastructure prioritisation, environmental management and transparent allocation of development resources. This study presents and stress-tests a reproducible remote-sensing and geographic information system framework for producing land-cover indicators for the 27 local government areas (LGAs) of Imo State, Nigeria. Because no verified satellite scenes, reference samples, or official administrative boundary dataset were supplied for this computational experiment, all numerical inputs and outputs are explicitly synthetic and are not presented as observations of current ground conditions. A deterministic simulation generated 5,530 km² of illustrative LGA area and five land-cover classes: built-up, cropland, forest/woodland, wetland/water, and bare/open land. The proposed operational workflow combines Sentinel-2 surface reflectance, spectral indices, terrain and texture variables, a random-forest classifier, stratified validation, and LGA-level zonal aggregation. In the simulated scenario, cropland accounted for 34.2% of the area, followed by forest/woodland (24.0%), built-up land (20.1%), bare/open land (12.3%), and wetland/water (9.4%). The synthetic confusion matrix produced 84.3% overall accuracy and a Cohen’s kappa of 0.800. Injecting 30% label noise reduced mean simulated accuracy from 84.3% to 65.8%, demonstrating that reference-data quality is a dominant operational risk. The contribution is therefore a transparent, auditable prototype—not an empirical map—showing how a future field-validated assessment can be designed, tested, and reported without confusing model output with measured evidence.
Keywords: Land-Cover Classification; Sentinel-2; Random Forest; GIS; Accuracy Assessment; Imo State