In this paper, we study the downlink channel state information (CSI) feedback problem in frequency-division duplexing (FDD) mode for massive multi-input multi-output (MIMO) systems. We proposed a joint convolution-coordinate attention neural networks algorithm based on deep learning. to address the shortcomings of insufficient performance, long operation time and high channel sparsity requirement of existing CSI feedback methods. The algorithm is based on the idea of autoencoder network, and converts the CSI feedback problem into an encoding and decoding problem. We proposed a new Joint Convolution-Coordinate Attention Mechanism Module (JCCAM), and a JCCAM-CsiNet neural network model is proposed based on CsiNet neural network model. Simulation results show that the algorithm proposed in this paper has better performance compared with existing channel estimation algorithms.