Large scale mapping: an empirical comparison of pixel-based and object-based classifications of remotely sensed data

Abstract

The study rationale  is based on the backdrop of the rigours involved in geo-dataset collection, time taken to process them, huge financial cost of purchase and processing, and higher accuracy requirements in analysing and mapping geographic data for reliable decision making. Using a case study of Aba city in south eastern Nigeria, the paper examined both methods of classifying high resolution satellite images in mapping using ENVI 4.8 and eCongnition software for data classifications and ArcGIS 10.1 software for cartographic map composition. It further evaluates the results of the raster maps generated in terms of classification accuracy (using error matrix) and equally test classification results’ agreement to geographic reality using Kappa Coefficient statistical analysis. Analysing 2012 QuickBird image as a proof of concept, the study reveals that the object-based approach provides a higher overall accuracy (OA= 98.75%) than the pixel-based approach (OA=79.44%). In addition, object-based results also show a higher overall producer accuracy (PA=98.42% > PA=85.37) and user accuracy (UA=96.70 > UA=81.04%) respectively. With a Kappa Coefficient of K=0.97 (very good) for object-based approach and K=0.62 (good) for pixel-based, the object-based approach shows a higher class separability between and among geographic objects especially water and bareland as evidenced in the Golf Course under re-construction. The study further proves to be the best for extracting roads and buildings especially when producing a large scale map such as cadastral or property map. This study is recommended for rapid large scale mapping of urban landscape that requires limited cost.