DBMSL-Net: A Dual-Branch Multi-Source Network for Lithological Classification in Vegetated Terrain

Under dense vegetation cover, deep networks for lithological classification must resolve small spectral differences between rock–vegetation mixtures. To address this issue, we propose the Dual-Branch Multi-Source Lithological Classification Network (DBMSL-Net), which uses Gaofen-5 Advanced Hyperspectral Imager (GF-5 AHSI) hyperspectral imagery, Phased Array type L-band Synthetic Aperture Radar-2 (PALSAR-2) data, and terrain data. Specifically, one branch encodes pixel spectra, and the other encodes spatial features from radar and terrain observations in the surrounding neighborhood. The spectral branch retains the overall spectral shape and local responses within intervals derived from absorption features in laboratory rock spectra. Laboratory pretraining initializes this branch, and source-domain class prototypes provide additional spectral references. A reliability gate adjusts the prototype contribution using target-pixel features and their similarity to the prototypes. The two branches are fused using weights that depend on the normalized difference vegetation index and pixel features. The core classifier, the differentiable binning residual module (DBRM), is trained jointly with both branches. It learns piecewise feature–class relationships through soft binning and refines class scores through successive residual updates. During inference, full-spectrum and feature-interval predictions are combined using class-specific weights. Neighborhood adjustment accounts for target extent. Across runs, boundary-aligned minimum-margin consensus is used for the spatially restricted target, whereas probability averaging is used for the extensive target. We evaluate DBMSL-Net in Lannitang, Yunnan, China, where quartz monzonite porphyry and quartz diorite porphyrite have similar mixed-pixel spectra and markedly different labeled sample sizes. On the within-area test subset, DBMSL-Net achieves a Target Macro F1 of 0.8301. Average accuracy, macro-averaged F1, and mean intersection over union reach 0.9388, 0.8780, and 0.7897, respectively. All four metrics exceed those of the nine comparison methods under the shared evaluation framework. Ablation experiments show that neither branch alone approaches the complete system (0.5727 and 0.7135, compared with 0.8301), and that the residual stage adds a measurable gain in both the initial and the complete configuration. The classification maps preserve the distributions of both the local and the extensive target and support lithological discrimination and field verification in this vegetated study area.