Using 3LC’s visualization dashboard, we can quickly recognize patterns within the wildfire detection dataset and address ground-truth labeling inaccuracies. Firstly, by visualizing how images are distributed in the model’s latent embedding space, we can intelligently remove redundant scenes that are feeding the model duplicate information. Next, by applying filters to key prediction metrics we can quickly find and fix false negatives in the ground-truth labels. This brief human-in-the-loop workflow produces a smaller and more accurate dataset, leading to improved classification performance from less data and a smaller model.
