 We propose a full convolutional network based on DenseNet for remote sensing scene classification that solves the problems of overfitting and lack of depth by using dense connections, adaptive average 3D pooling, and a convolutional layer instead of a fully connected layer. Our network has more than 100 layers but only about 7 million parameters, significantly improving classification performance on several datasets compared to state-of-the-art algorithms. This article was authored by Jian Mingzhuang, Chao Kuan Lu, Su Dongli, and others.