In recent years, there has been a growing interest in the use of deep learning-based hyperspectral unmixing methods. However, most existing networks only extract either spectral information or local spatial-spectral correlations of a single pixel, and their performance is often unsatisfactory. To address this issue and further extract spatial-spectral information from hyperspectral images, In this paper, a novel hyperspectral unmixing method based on spatial-spectral dual branch autoencoder and adaptive convolution is proposed. This method utilizes the autoencoder architecture with different branches designed to extract both spatial and spectral features. The extracted features are then stacked and propagated through the network, followed by mapping to obtain the abundance percentage. Additionally, We have incorporated sparsity constraint on the loss function as a way to exploit the correlation between neighboring pixels and further improve the unmixing performance. We evaluated the performance of our proposed method on two real hyperspectral datasets, and the results demonstrate significant accuracy compared to existing methods.