Land Decoding: A Comparative Study on Image Recognition Using U-Net for Urban Parks

Extracting and analyzing landscape features from aerial imagery has great potential for understanding characteristics and change. The level of spatial detail required for design and planning problems can vary depending on the methods and data resolutions. Although broad-scale data may suffice for landscape characterization in ecologically oriented planning, site-specific design and analysis for urban landscapes require fine-scale methods. Given the complexity and loss of data resulting from the translation of 3D spatial data into 2D representations such as land classification maps, the research aimed to formulate a feature extraction approach to recognize the implicit characteristics of urban parks. The U-net algorithm, a detailed semantic segmentation, was applied and compared with two other commonly referenced alternatives with their relevant data types – Random Forest (RF) and Object-Based Image Analyses (OBIA) – in the land classification literature. The selection of the algorithms and data sources was grounded in considering differentiation between landscape character classification methods, which are highly emphasized and easily applicable, and the spatial feature extraction of fine-scale landscapes as an overlooked field. A new urbanized image segmentation approach was adapted to complex landscapes by exploring the possibilities and drawbacks of methods and medium-resolution data. The study showed that the U-Net algorithm can predict in-between areas of urban parks and give more consistent recognition than OBIA for very high-resolution aerial images. In conclusion, using the U-net algorithm for site-specific and theme-based tacit features, such as detailed spatial compositions of underwood textures regarding urban parks, can be extracted.