A systematic review finds deep learning exceeds 97 percent accuracy on remote sensing benchmarks but struggles with generalization, spectral exploitation and interpretability.