Buildings are not only central to the day-to-day activities but also serve as critical indicators of urban development and transformation. The automatic extraction of building footprints from high-resolution Remote Sensing Imagery (RSI) has emerged as an important and popular tool in urban studies. It helps to enhance the understanding and management of urban sprawl, urban planning, population estimation, resource allocation, and post-disaster damage assessment. In this context, having an automated and robust building detection model is crucial. Deep Learning (DL) model and Object-Based Image Analysis (OBIA) techniques are the main and commonly used for automated building detection. This study investigates the efficacy of a pre-trained DL model and a rule-based model OBIA techniques in building detection across varied resolutions and geographic settings. Employing orthophotos from Luleå, Gävle, and Stockholm, the research assesses the adaptability and robustness of these methods under image properties and urban densities.

The DL model was initially trained on 0.25m resolution data of Sweden by Lantmäteriet (Sweden mapping agency). The rule-based model was developed by applying OBIA techniques on behalf of this study. Models were analyzed through six feature agreement statistics including Critical Success Index (CSI), Precision, and Detection Probability (POD). The findings reveal that the DL model consistently outperformed the OBIA approach across all study areas, particularly at the original 25 cm resolution. Gävle showed superior precision with a CSI of 0.8139 for the DL model against a CSI of 0.7493 for OBIA at 25 cm. 

The evaluation was improved by considering 50*50 sq. m subsets and building sizes. These evaluations highlight that building size and urban density significantly influence detection accuracy. Larger (> 2500 sq. m) buildings and less dense areas tend to yield higher accuracy across both detection methods. The DL model exhibited high CSI values for very large buildings (>5500 sq. m) in Gävle, surpassing 0.8, while the detection of very small (< 50 sq. m) buildings remained challenging for both methods.

Overall, the pre-trained DL model is very sensitive to resolution changes compared to OBIA. Importantly, both give their best performance at the original resolution while DL is superior than OBIA. A rule-based OBIA model is affected by the geographical characteristics more heavily than a DL model. Both models have their best performance in the area with medium building density when medium to very large buildings exist. This study highlights how big the impact of building size, geographic characteristics, and image resolution on the performance of DL and OBIA techniques. However, further investigation is recommended to draw a strong conclusion regarding the impact of resolution on the model performance.