Fig. 2: Land cover classification in Kippa-Ring, Australia (one among the DFC2020 regions), when using one example per class (1-shot), so five examples in total.
From: Meta-learning to address diverse Earth observation problems across resolutions
We show the confusion matrix of the predictions obtained on the 677 remaining images, which have been classified at a 79.6% accuracy in the first (of three) random splits. In 1-shot learning, the choice of training images is especially important, as the representation of classes are solely defined by these single training examples. The second random split is only slightly worse with 78.4%, while the third split is only classified with 47% accuracy. In that last case, several forest images were wrongly classified as grassland (not shown in the figure). To accommodate for this randomness, we average the accuracy of all three random splits leading to a 1-shot accuracy of 68%. The variance between splits decreases with more shots, as can be seen on the quantitative table in Supplementary Table 2.
