Fueled by the boom of deep learning, electric power maintenance has turned from human inspection to intelligent era. Massive image data will be collected and analyzed by machines. A flexible and efficient image compression scheme is vital especially in varied communication channel. To address this requirement, we propose a scalable image compression model utilizing features at different scales, which breaks the limitation in preserving fine spatial details caused by too many steps of down-sampling. Specifically, a pyramidal encoder is applied to generate small-scale and large-scale representations which correspond to base layer and the enhancement layer respectively. A decomposition module is developed to explore and reduce the redundancy between intra-layers and intra-layers. In addition, channel autoregressive model is introduced to further squeeze the statistical redundancy of the latent representations. The experiments show that our model outperforms all the comparative scalable methods without significant increase in complexity. In terms of MS-SSIM, about 38% BD-rate reduction and 1.8 dB BD-PSNR gain over the VVC can be achieved.