GPCNet: Grid-Parallel Compact Network With Kernel-Adaptive Feature Fusion for Semantic Segmentation of High-Resolution UAV Imagery

Abstract

In recent years, encoder–decoder architectures have driven remarkable progress in remote sensing semantic segmentation. However, achieving both high accuracy and computational efficiency for high-resolution UAV imagery remains challenging. For Vision Transformer (ViT)-based models, the quadratic computational and memory complexity of self-attention with respect to the number of tokens becomes prohibitive for large-scale UAV imagery. In addition, complex scenes with rich textures and fine-grained structures often exhibit significant intraclass appearance variations, which may degrade feature consistency and lead to object misclassification and boundary misalignment. To address these challenges, this article proposes GPCNet, an efficient semantic segmentation framework for high-resolution UAV imagery. First, we introduce foresight encoder sparsification (FES), which constructs a compact sparse encoder by evaluating parameter importance and removing redundant parameters prior to training. Second, we propose a grid-parallel encoding with stage-end semantic exchange (GPE-SSE) strategy to significantly reduce the computational cost of ViT-based encoders for high-resolution images while maintaining long-range semantic dependencies through cross-grid information exchange. Finally, a kernel-adaptive feature fusion (KAFF) module is designed for the semantic decoder, which generates spatially variable low-pass kernels (LPKs), high-pass kernels (HPKs), and dynamic dilated kernels (DDKs) to adaptively modulate high-frequency details and contextual receptive fields, thereby improving intraclass feature consistency and segmentation accuracy. The experimental results on three public high-resolution UAV image datasets demonstrate that our method maintains competitive segmentation accuracy while substantially improving inference efficiency. Specifically, with the compact MiT-B2 as the encoder, our GPCNet-P achieves 72.32% and 81.31% mIoU on the UAVid and VDD test sets, respectively, while reducing encoder parameters by 40% and the total FLOPs by 71.3% compared to the dense encoder. Our code is publicly available at https://github.com/Memoristor/GPCNet