Title: DMAConv: Dual Mask-Adaptive Convolution for Remote Sensing Pansharpening

Abstract:

The process of pansharpening involves combining a high-resolution panchromatic image with a lower-resolution multispectral image. However, current deep learning approaches, even those utilizing recent adaptive convolution techniques, face significant hurdles in handling the regional heterogeneity inherent in remote sensing data, often resulting in excessive computational demands. To overcome these limitations, we introduce Dual Mask-Adaptive Convolution (DMAConv), an innovative operator designed to dynamically distribute computational resources according to specific feature traits.

DMAConv utilizes a lightweight module to produce both hard and soft masks. The hard mask functions to divide features into two distinct pathways: a compact branch responsible for globally processing redundant information, and a focused branch that dedicates higher computational effort to modeling complex, heterogeneous areas. Subsequently, the soft mask applies preliminary modulation to the input features for both branches. This architecture, characterized by its dual-branch and mask-adaptive nature, substantially improves feature representation while keeping computational overhead to a minimum. Comprehensive experimental results show that our approach attains state-of-the-art performance across a wide range of quantitative benchmarks, boasting significantly fewer parameters and the lowest computational cost among existing adaptive convolution models.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC