From the abundance perspective: Multi-modal scene fusion-based hyperspectral image synthesis