Alter-CNN: An Approach to Learning from Label Proportions with Application to Ice-Water Classification

We present an approach to train a model for classifying ice and open water directly using the polygon-wise ice concentration available from ice charts. This can be considered as a “learning from label proportions” (LFLP) problem which has been studied in the last decade and applied to many real-world applications. Our approach is based on convolutional neural networks (CNNs), which have been shown to capture representative features and achieve impressive classification results provided a large number of labeled training samples. We provide a probabilistic formulation to learn from label proportions while considering the proportion bias, and an expectationmaximization (EM) approach is employed to estimate both the CNN model parameters and infer the per-pixel labels. Experiments on a large-scale satellite image dataset1 show that our proposed approach achieves better results than previous approaches for LFLP problems.