For this purpose of the rapid and effective scene interpretation for remote sensing big data, a novel method for feature representation is presented. First of all, the corresponding filter banks of local training samples in image are obtained by unsupervised learning. The feature map of scene units is then quantized based on binary hashing coding. Finally, the global feature of scene is obtained from statistical results. The experimental results demonstrate that, combining with both the advantages of filter bank and binary feature descriptor, the proposed feature can greatly enhance the computational efficiency on the assumption of ensuring high accuracy, which has shown its robustness.