MDCA-Net: Remote Sensing Image Classification Based on Multiscale Dynamic Convolution and Attention Enhancement

In response to the challenges faced by traditional deep learning models in adaptively extracting multi-scale features from remote sensing images and establishing long-range dependencies across both channels and spatial dimensions—issues that contribute to a decline in classification performance—this study introduces a novel network architecture that integrates dynamic multi-scale attention with spatiotemporal fusion, designated as (Multi-scale Dynamic Convolution and Attention Fusion Network) MDCA-Net.The research presents a Dynamic Multi-scale Large Kernel Attention, termed D-MLKA, which is engineered to dynamically weight and amalgamate multi-scale convolutional features derived from 7×7 and 5×5 kernels, thereby facilitating a spectral-spatial adaptive perception mechanism. Subsequently, a Spectral–Spatial Convolution and Attention Fusion Module, SS-CAFM, is developed to leverage depthwise separable convolutions for the extraction of cross-channel spectral correlations, in conjunction with multi-head attention to effectively capture long-range spatial dependencies. Moreover, a pseudo-temporal enhancement unit is introduced to bolster the model’s robustness against interannual variations by generating virtual temporal features. Comparative experiments conducted on the Indian Pines and Houston University public datasets yielded overall accuracies (OA) of 77.38% and 83.34%, respectively. Furthermore, when validated against the 2024 multispectral dataset from Rong’an County, Guangxi Autonomous Region, the model achieved a classification accuracy of 98.66%, representing a 2.15% improvement over existing methodologies such as Morphformer. Ultimately, the model demonstrated Average Accuracy (AA) accuracies of 98.12% and 97.79% on datasets from 2020 and 2022, respectively, thereby sustaining high performance levels while mitigating the risk of performance degradation.