DenseNetx: Efficient DenseNets for Remote Scene Classification without Pretraining

Remote sensing scene classification is growing fast in demand and application within the Earth Observation domain. Satellite Image data are usually high resolution but low in number. DenseNet architectures are quite powerful and achieve good accuracy in this task even without large-scale pretraining from ImageNet-like datasets. But, DenseNet lacks efficiency and is considered a quite heavy model by modern standards. We propose DenseNetx, a family of efficient densenet architecture which can dramatically reduce computation costs while outperforming the baseline model. In short, we use a larger input size while aggressively downsampling in the stem block using two $3\times 3$ convolutions of stride 2, and use large-kernel depthwise-separable convolution in the denselayer to achieve higher efficiency. Our results on the WHU-RS19 and Optima1-31 scene classification datasets show that our model can outperform the baseline at 20% reduced parameters and 53% fewer flops, while achieving up to 4.5% increased accuracy with a larger input while retaining efficiency.