GreenNet: A dual-encoder network for urban green space classification using high-resolution remotely sensed images

This is the authors' abstract. We don't add key points for this paper.

• Propose a dual-encoder neural network with a transformer-based block for urban green space classificaition. • Creat a new benchmark dataset for urban green space classificaition from remote sensing images. • Conduct compartive experiments and ablation studies to indicate the superior performance of the proposed method. Accurate classification of urban green spaces from high-resolution remotely sensed images is critical for ecological environment planning, construction, and utilization. However, existing deep learning networks for large-scale high-resolution remote sensing images often face limited receptive fields and insufficient extraction of global information, making it challenging to achieve satisfactory performance on urban green space classification tasks. To address these issues, this paper presents a novel dual-encoder network, termed GreenNet, specifically designed for urban green space classification from high-resolution remotely sensed images. GreenNet features a unique dual-encoder structure. i.e., an inside encoder for efficiently extracting interior intra-image (i.e., local and global) features of urban green spaces from the small-sized images cropped from raw input remote sensing images, and an outside encoder for modeling long dependencies (i.e., external inter-image features) from the large-sized images cropped from raw input images. Additionally, a transformer-based outside-global–local attention block (OGLAB) is developed to fuse the intra-image and inter-image features from the dual-encoder to effectively capture inherent semantic representations of urban green spaces. Finally, to ensure classification consistency along class boundaries, a boundary loss is computed using edge-defined images, which are generated by a pre-trained Segmenting Anything Model (SAM) from the raw input image. The proposed GreenNet was evaluated on a self-built urban green space dataset, covering the whole area of Nanshan District, Shenzhen City, China, achieving an overall accuracy (OA) of 88.88 %, a mean F1-score (mF1) of 74.06 %, and a mean Intersection over Union (mIoU) of 60.77 %, respectively, demonstrating its superior performance to state-of-the-art networks on green space classification tasks.