Article: Fine-grained classification and health assessment of urban green spaces using U-net and transfer learning Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.94 pp.42 - 64 Abstract: This study develops a U-Net with transfer learning framework for fine-grained classification and health assessment of urban green spaces using GF-2 remote sensing imagery. The two-level classification includes pixel-level land cover (arbour, shrub, herb, water, impervious surface) and patch-level functional types (park, protective, residential, roadside green space). Health assessment quantifies vegetation coverage, species richness, vitality, landscape pattern, and ecological service potential from vegetation structure to ecological function. The U-Net encoder incorporates MobileNetv2 and an attention mechanism to improve boundary recognition accuracy. Transfer learning from ImageNet pretrained models addresses insufficient samples and cross-domain adaptation, reducing convergence epochs from 120 to 35 and total training time from 84 to 24.5 hours under identical experimental conditions. Comparative experiments on 1,369 preprocessed 256 × 256-pixel samples demonstrate 89.5% overall accuracy, representing a 12.7 percentage point improvement over baseline U-Net, with significantly enhanced mean intersection-over-union across all evaluated categories. Inderscience Publishers - linking academia, business and industry through research

Title: Fine-grained classification and health assessment of urban green spaces using U-net and transfer learning

Authors: Haibo Wang; Yafei Wang; Song Wang; Xue Huang; Shiying Jia

Addresses: Department of Resources and Environment, Shandong Water Conservancy Vocational College, Rizhao, 276800, China ' Department of Resources and Environment, Shandong Water Conservancy Vocational College, Rizhao, 276800, China ' College of Biology Engineering, BinZhou Polytechnic University, Binzhou, 256603, China ' College of Biology Engineering, BinZhou Polytechnic University, Binzhou, 256603, China ' Department of Computer Science, Tianjin City Vocational College (Hexi Branch), Tianjin, 300000, China

Abstract: This study develops a U-Net with transfer learning framework for fine-grained classification and health assessment of urban green spaces using GF-2 remote sensing imagery. The two-level classification includes pixel-level land cover (arbour, shrub, herb, water, impervious surface) and patch-level functional types (park, protective, residential, roadside green space). Health assessment quantifies vegetation coverage, species richness, vitality, landscape pattern, and ecological service potential from vegetation structure to ecological function. The U-Net encoder incorporates MobileNetv2 and an attention mechanism to improve boundary recognition accuracy. Transfer learning from ImageNet pretrained models addresses insufficient samples and cross-domain adaptation, reducing convergence epochs from 120 to 35 and total training time from 84 to 24.5 hours under identical experimental conditions. Comparative experiments on 1,369 preprocessed 256 × 256-pixel samples demonstrate 89.5% overall accuracy, representing a 12.7 percentage point improvement over baseline U-Net, with significantly enhanced mean intersection-over-union across all evaluated categories.

Keywords: U-Net; transfer learning; green space classification; health assessment; remote sensing imagery.

DOI: 10.1504/IJICT.2026.156127

International Journal of Information and Communication Technology, 2026 Vol.27 No.94, pp.42 - 64

Received: 02 Mar 2026
Accepted: 16 Apr 2026
Published online: 04 Sep 2026 *

Title: Fine-grained classification and health assessment of urban green spaces using U-net and transfer learning

Keep up-to-date