Very-high　resolution　(VHR)　remote　sensing　images　(RSIs)　have　significantly　larger　spatial　size　compared　to　typical　natural　images　used　in　computer　vision　applications.　Therefore,　it　is　computationally　unaffordable　to　train　and　test　classifiers　on　these　images　at　a　full-size　scale.　Commonly　used　methodologies　for　semantic　segmentation　of　RSIs　perform　training　and　prediction　on　cropped　image　patches.　Thus,　they　have　the　limitation　of　failing　to　incorporate　enough　context　information.　In　order　to　better　exploit　the　correlations　between　ground　objects,　we　propose　a　deep　architecture　with　a　two-stage　multiscale　training　strategy　that　is　tailored　to　the　semantic　segmentation　of　large-size　VHR　RSIs.　In　the　first　stage　of　the　training　strategy,　a　semantic　embedding　network　is　designed　to　learn　high-level　features　from　downscaled　images　covering　a　large　area.　In　the　second　training　stage,　a　local　feature　extraction　network　is　designed　to　introduce　low-level　information　from　cropped　image　patches.　The　resulting　training　strategy　is　able　to　fuse　complementary　information　learned　from　multiple　levels　to　make　predictions.　Experimental　results　on　two　data　sets　show　that　it　outperforms　local-patch-based　training　models　in　terms　of　both　accuracy　and　stability.