A Land Cover Classification Method for High-Resolution Remote Sensing Images Based on NDVI Deep Learning Fusion Network

High-resolution remote sensing (HRRS) images have few spectra, low interclass separability and large intraclass differences, and there are some problems in land cover classification (LCC) of HRRS images that only rely on spectral information, such as misclassification of small objects and unclear boundaries. Here, we propose a deep learning fusion network that effectively utilizes NDVI, called the Dense-Spectral-Location-NDVI network (DSLN). In DSLN, we first extract spatial location information from NDVI data at the same time as remote sensing image data to enhance the boundary information. Then, the spectral features are put into the encoding-decoding structure to abstract the depth features and restore the spatial information. The NDVI fusion module is used to fuse the NDVI information and depth features to improve the separability of land cover information. Experiments on the GF-1 dataset show that the mean OA (mOA) and the mean value of the Kappa coefficient (mKappa) of the DSLN network model reach 0.8069 and 0.7161, respectively, which have good applicability to temporal and spatial distribution. The comparison of the forest area released by Xuancheng Forestry Bureau and the forest area in Xuancheng produced by the DSLN model shows that the former is consistent with the latter. In conclusion, the DSLN network model is effectively applied in practice and can provide more accurate land cover data for regional ESV analysis.

Automated summary

The study introduces a novel deep learning fusion network DSLN that utilizes NDVI to enhance land cover classification of HRRS images, with experiments showing promising results on the GF-1 dataset and good applicability for temporal and spatial distribution.