Uncertainty-Guided Representation Learning in Local Climate Zone Classification

Abstract

A significant leap forward in the performance of remote sensing models can be attributed to recent advances in machine and deep learning. Large data sets particularly benefit from deep learning models, which often comprise millions of parameters. On which part of the data a machine learner focuses on during learning, however, remains an open research question. With the aid of a notion of label uncertainty, we try to address this question in local climate zone (LCZ) classification. Using a deep network as a feature extractor, we identify data samples that are seemingly easy or hard to classify for the model and base our experiments on the relatively more uncertain samples. For training of the network, we make use of distributional (probabilistic) labels to incorporate the voter confusion directly into the training process. The effectiveness of the proposed uncertainty-guided representation learning is shown in context of active learning framework where we show that adding more certain data to the training pool increases model performance even with the limited data.

Can Land Cover Classification Models Benefit From Distance-Aware Architectures?

“…We closely follow the code implementation of the authors of [11], which can be found in [25]. We expand the experiments conducted by [19] by adding another deterministic uncertainty quantification technique, DUE. It is here denoted by GP-IP, which stands for a Gaussian Process (GP) output mapping with the inducing points (IP) approximation by [11].…”

“…The OOD detection is carried out by a binary classifier defined by the predictive uncertainties, on which the area under the receiver operator curve (AUROC) and under the precision recall curve (AUPR) are derived. The Dempster-Shafer metric (DSM) and 1 minus the maximum softmax probability (MSP) have already been previously used as uncertainty metric in [19], here the predictive entropy of the softmax prediction (Pred.Ent.) is added specifically for GP-IP approaches.…”

“…Adding to the experimental setup presented by [19], we conduct OOD detection on another remote sensing image classification benchmark dataset, namely on the So2Sat LCZ42 [26] dataset. This dataset contains labeled Sentinel-2 imagery of 42 urban conglomerates around the world.…”

“…Also, c denotes a hyper-parameter to practically control the exact upper bound depending on the input data. As a novelty compared to a previous study [19], we here follow the convention of [11] and apply the spectral normalization also to the batch normalization. The Lipschitz constant of the batch normalization operator is given by max…”

“…We refer the interested reader to [24]. Note that the SNGP approach proved to effectively increase the model's predictive uncertainty quality in an OOD detection setting for remote sensing image classification [19].…”