Channel–spatial attention-based pan-sharpening of very high-resolution satellite images

The pan-sharpening process aims to generate a new synthetic output image preserving the spatial details of panchromatic and spectral details of the multi-spectral image inputs. Recently, deep learning-based methods show substantial success in the remote sensing field mostly with the application of traditional Convolutional Neural Networks (CNNs). Most of the traditional CNN-based approaches treat all the channels equitably and cannot learn the correlation. Attention mechanism which can learn the correlations among the channels has been proven to be effective in super-resolution and object detection tasks. In this research, we introduced a novel deep learning framework, channel–spatial attention-based method for pan-sharpening (CSAPAN), by designing a Densely residual attention module (RAM). Besides, we train our model in the high-frequency domain and up-sample the low-resolution multispectral images by using the pixel shuffle method before stacking with the panchromatic images for further feature extraction. We evaluated our proposed CSAPAN along with traditional methods and CNN-based methods in reduced and full resolution and obtained satisfactory quantitative and qualitative results on Pleiades, Worldview-2, and QuickBird-2 satellite image datasets.