A SwiGLU enhanced ConvNeXt framework for multispectral Sentinel-2 land use and land cover classification

Land use and land cover (LCLU) classification from multispectral satellite imagery is important for applications such as urban planning, environmental monitoring, agriculture, and disaster management. Existing deep learning based approaches have improved LCLU; however, most of them utilize only the red, green, and blue (RGB) bands of multispectral sentinel-2 imagery, thereby underutilizing the complementary spectral information available in the near-infrared (NIR), short-wave infrared (SWIR), and derived spectral indices such as normalized difference built-up Index (NDBI), Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI). In this work, we propose a convolutional neural network (CNN), SwigNeXT, based on a modified ConvNeXt architecture for the classification of multispectral Sentinel-2 imagery using spectral band combinations which are RGB, NIR–Red–Green, SWIR–NIR–Red, and derived index combination NDBI–NDVI–NDWI rather than using only the RGB. The SwigNeXT incorporates SwiGLU-based gated feature modulation to improve spectral spatial feature representation. The final classification is obtained by ensembling the predictions of the three spectral band combinations and the derived index combination. Experimental results show that the proposed ensemble framework achieved an overall accuracy of 99.19% with a Kappa coefficient ( \(\kappa \) ) of 0.9909, outperforming conventional CNNs, transfer learning based CNNs and transformer based approaches. A gradient based SHapley Additive exPlanations framework was further integrated to analyse spatial and spectral feature contributions. The proposed framework was further validated on Sentinel-2 imagery from the Delhi–NCR region, to examine its applicability under geographically distinct real world imaging conditions.