Deep Aggregation Net for Land Cover Classification

Publication - IEEE CVPRW 2018

Tzu-Sheng Kuo, Keng-Sen Tseng, Jia-Wei Yan, Yen-Cheng Liu, and Yu-Chiang Frank Wang, “Deep Aggregation Net for Land Cover Classification,” IEEE International Conference on Computer Vision and Pattern Recognition Workshop on DeepGlobe (CVPRW 2018), 2018.

Figure 1: Architecture of our deep aggregation net. The rectangular boxes represent tensor features and the arrows denote operations. Blocks 1 to 4 are residual convolutional blocks, and ASPP indicates Atrous Spatial Pyramid Pooling. Each tensor feature is specified with its output-stride (os), which denotes the ratio of input image spatial resolution to the feature resolution.

Land cover classification aims at classifying each pixel in a satellite image into a particular land cover category, which can be regarded as a multi-class semantic segmentation task. In this paper, we propose a deep aggregation network for solving this task, which extracts and combines multi-layer features during the segmentation process. In particular, we introduce soft semantic labels and graph-based fine tuning in our proposed network for improving the segmentation performance. In our experiments, we demonstrate that our network performs favorably against state-of-the-art models on the dataset of DeepGlobe Satellite Challenge, while our ablation study further verifies the effectiveness of our proposed network architecture.

Figure 2. Example segmentation results using different models. Note that our model was able to accurately classify pixels over different scales/regions.