Few-shot Classification of Aerial Scene Images via Meta-learning

Abstract

CNN-based methods have dominated the field of aerial scene classification for the past few years. While achieving remarkable success, CNN-based methods suffer from excessive parameters and notoriously rely on large amounts of training data. In this work, we introduce few-shot learning to the aerial scene classification problem. Few-shot learning aims to learn a model on base-set that can quickly adapt to unseen categories in novel-set, using only a few labeled samples. To this end, we proposed a meta-learning method for few-shot classification of aerial scene images. First, we train a feature extractor on all base categories to learn a representation of inputs. Then in the meta-training stage, the classifier is optimized in the metric space by cosine distance with a learnable scale parameter. At last, in the meta-testing stage, the query sample in the unseen category is predicted by the adapted classifier given a few support samples. We conduct extensive experiments on two challenging datasets: NWPU-RESISC45 and RSD46-WHU. The experimental results show that our method outperforms three state-of-the-art few-shot algorithms and one typical CNN-based method, D-CNN. Furthermore, several ablation experiments are conducted to investigate the effects of dataset scale and support shots; the experiment results confirm that our model is specifically effective in few-shot settings.

RS-SSKD: Self-Supervision Equipped with Knowledge Distillation for Few-Shot Remote Sensing Scene Classification

“…Backbone. Following recent works [ 50 , 51 , 52 , 59 ], we use Resnet-12 [ 9 ] as the backbone in both the SSKD module and the meta-learning module. A GAP layer is added to the last ResNet Block, which outputs 512-dimensional embeddings; the details are introduced in Section 4.2.1 .…”

“…Table 2 and Table 3 show the results of several FSL approaches on the NWPU-RESISC45 dataset, where both 5-way 1-shot and 5-way 5-shot classification performance are reported. The main results of comparison approaches are cited from a recent study [ 50 ] on few-shot classification of RS scenes. The methods marked with an asterisk indicate that the backbone of the original method is replaced with Resnet-12.…”

“…Most of the comparison methods evaluate the models on 600 tasks sampled from the UNSEEN categories, which leads to high variance. We follow the more reliable evaluation setting suggested by Zhang et al [ 50 ], evaluating our method on 8000 sampled tasks. The mean accuracy (in %) with 95% confidence interval is reported for comparison.…”

“…Instead of learning from scratch, we note that Zhang et al [ 50 ] utilize a pre-train stage to classify all SEEN categories with the cross-entropy loss (e.g., 25 categories in the NWPU-RESISC45 Dataset). The feature maps generated by the last residual block are fed to the GAP layer, whose weights are then used as initialization in the meta-training stage.…”

“…Li et al [ 49 ] introduce an approach based on Protypical Networks (ProtoNet) [ 40 ] to explore RS scene few-shot classification. The authors of [ 50 ] provide a testbed for few-shot classification of RS scene and re-implemented several well-known few-shot learning approaches with a deeper backbone Resnet-12 for a fair comparison. While many existing FSL models [ 45 , 51 , 52 ] focus on devising different architectures, we argue that feature embedding are overlooked.…”