<p>Recent advances have witnessed that deep learning-based joint source-channel coding (DeepJSCC) methods can achieve noise-resiliency performances for wireless image transmission tasks. Among them, a method named nonlinear transform source-channel coding (NTSCC) achieves superior performance by incorporating the nonlinear transform as a prior to extract the source semantic features and developing a hyperprior-aided codec refinement mechanism. However, the NTSCC framework still cannot achieve adaptive code rates for different channel signal-to-noises (SNRs), which reduces its flexibility and bandwidth efficiency. Additionally, the entropy model in the NTSCC inadequately captures the channel and spatial correlation between latent features, which leads to inaccurate rate transmission. In this paper, we propose an adaptive nonlinear deep coding (ANDC) framework for realizing flexible code rate optimization and improving accuracy of rate transmission. ANDC is realized by a hybrid attention multi-reference entropy model (HA-MREM) that captures the correlation of latent features in channel and spatial to improve the guidance of rate allocation, and adaptive hybrid attention (AHA) module that combines the SNR in both channel and spatial to adapt to different SNRs to flexibly adjust the transmission strategy. Combining these two structures, ANDC enables flexible and efficient image transmission. Simulation results show that the proposed model achieves equal or even better performance results in several metrics compared to the existing NTSCC model.</p>