Land cover classification of the Shendong mining area based on reflectance–emissivity–texture fusion

Purpose In Shendong mining area, land cover types are interspersed, and features such as dark bare land, impervious surfaces, and vegetation are easily confused in reflectance spectra, which affects classification accuracy.

Methods To improve the accuracy and robustness of land cover classification in complex mining areas, a synergistic classification approach integrating reflectance, emissivity, and textural features is developed. Based on Landsat-9 imagery, land surface reflectance, land surface emissivity, and gray-level co-occurrence matrix (GLCM) texture features were constructed. Three progressive input configurations were designed, namely reflectance only, reflectance plus emissivity, and reflectance plus emissivity and texture. A spectral–spatial network combining three-dimensional convolution and attention mechanisms was established, and comparative experiments were conducted with Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP).

Results and Conclusions  The results show that the introduction of land surface emissivity enhances the separability between spectrally similar ground objects, particularly alleviating confusion between dark bare land and impervious surfaces. With the further integration of textural features, the spatial continuity of classification results is improved, and local salt-and-pepper noise is reduced. Compared with random forest, support vector machine, and multi-layer perceptron, the proposed spatial–spectral network achieves higher classification accuracy and stability under different feature configurations, indicating that the model can effectively exploit the complementary information among reflectance, emissivity, and textural features. Under the full-feature input condition, the model achieves an overall accuracy of 87.71% and a Kappa coefficient of 0.617, representing a 3.05% improvement in overall accuracy over the best-performing traditional machine learning model. This study demonstrates that the synergistic expression of reflectance, emissivity, and textural features, combined with spatial–spectral joint modeling, can effectively enhance the accuracy of land cover classification in complex mining areas.