Fig 1.
Representative samples from the TV-RSI dataset.
All sampling, annotations, panel composition, and graphics are original works by the authors and are released under CC BY 4.0.
Fig 2.
Regional distribution of semantic objects.
Fig 3.
Heat map of the object’s position.
Fig 4.
Overall architecture diagram of MPLNet.
Fig 5.
Structure of the Mamba Fusion Module (MFM).
Table 1.
Experimental results on the TV-RSI dataset (pixel accuracy Acc and IoU per class).
Fig 6.
Qualitative comparisons on TV-RSI.
Ground truth masks and model outputs are shown for representative tiles. Base imagery (where visible) was obtained from the Geospatial Data Cloud (GF-2) for non-commercial academic use. All overlays, annotations, and panel layouts are original works by the authors and are released under CC BY 4.0.
Fig 7.
Ablation of the Mamba Fusion Module (MFM).
Visual comparisons among Backbone, MPLNet (w/o MFM), and MPLNet demonstrate sharper boundaries and reduced omissions with MFM. Base imagery (where visible) was obtained from the Geospatial Data Cloud (GF-2) for non-commercial academic use. All derived visualizations are author-original and released under CC BY 4.0.
Table 2.
Ablation results of MFM.
Fig 8.
Ablation of prompt learning (PL).
Compared with the student baseline, MPLNet with PL recovers fine structures and reduces confusion in mixed land-cover zones. Base imagery (where visible) was obtained from the Geospatial Data Cloud (GF-2) for non-commercial academic use. All annotations and compositions are original works by the authors and are released under CC BY 4.0.
Table 3.
Ablation results for prompt learning.
Table 4.
Quantitative comparison results for the Vaihingen dataset.
Table 5.
Quantitative comparison results for the Potsdam dataset.
Table 6.
Comparison of model parameters and computational complexity.