Abstract: ‍ ‍Wildfire has caused heavy damage to the environment and human in recent years. However， commonly used detection methods are based on the satellites or UAVs， which was inefficient and high-cost. Massive sensor network settled in remote area can not afford the big data transmission because of the limited capability， especially images transmission， which can provide useful and clear information about the change of surrounding during the disaster. To solve the above question， this paper proposed a GAN（Generative Adversarial Networks） based framework for image compression using deep learn and semantic communication， by extracting the semantic information of the captured photos and transmitting the semantic label. The generator of GAN reconstructed the corresponding image based on the received semantic information. To avoid learning the complex physical channel feature， channel coding took the convention solution like LDPC code in our method. The experimental results show that the proposed method achieves lower bitrate and less distortion than BPG， the state of the art coding format in traditional image compression.