Abstract
Airborne remote sensing scene classification is important for land-use surveys, urban planning, and environmental monitoring. However, large intra-class variation, strong inter-class similarity, and heterogeneous spatial structures make it difficult to coordinate representations extracted by different network pathways. To address this problem, this paper proposes a dual-branch split-attention network (DSBF-Net) with spatial bilinear fusion. EfficientNet and MobileNetV3 are employed as structurally heterogeneous backbones, Split-Attention is applied before and after fusion, and a spatial bilinear fusion module (SBFM) is introduced to align the two feature streams and model position-wise diagonal bilinear interactions. SBFM further incorporates signed square-root transformation, normalization, Coordinate Attention refinement, and residual aggregation. Experiments on the UC Merced Land-Use (UCM) dataset and Aerial Image Dataset (AID) show that DSBF-Net outperforms concatenation, addition, direct multiplication, and cross-attention fusion strategies, achieving classification accuracies of 99.07% and 97.58%, respectively. Component ablations further support the contribution of heterogeneous dual-branch learning, Split-Attention, and structured cross-branch fusion. The evaluation framework additionally examines backbone complementarity, hyperparameter sensitivity, corruption robustness, computational efficiency, interpretability, and limited-label behavior.

This publication has 36 references indexed in Scilit: