An Entropic Optimal Transport loss for learning deep neural networks under label noise in remote sensing images

Abstract

Deep neural networks have established as a powerful tool for large scale supervised classification tasks. The state-of-the-art performances of deep neural networks are conditioned to the availability of large number of accurately labeled samples. In practice, collecting large scale accurately labeled datasets is a challenging and tedious task in most scenarios of remote sensing image analysis, thus cheap surrogate procedures are employed to label the dataset. Training deep neural networks on such datasets with inaccurate labels easily overfits to the noisy training labels and degrades the performance of the classification tasks drastically. To mitigate this effect, we propose an original solution with entropic optimal transportation. It allows to learn in an end-to-end fashion deep neural networks that are, to some extent, robust to inaccurately labeled samples. We empirically demonstrate on several remote sensing datasets, where both scene and pixel-based hyperspectral images are considered for classification. Our method proves to be highly tolerant to significant amounts of label noise and achieves favorable results against state-of-the-art methods.

Unbalanced minibatch Optimal Transport; applications to Domain Adaptation

“…We see that entropy helps getting slightly better results however when the entropic regularization is too high, the accuracy falls. We conjecture that entropic regularized OT regularizes the neural network because the target prediction is matched to a smoothed source label (see a similar discussion in (Damodaran et al, 2019)). And it is well known that label smoothing creates class clusters in the penultimate layer of the neural network (Müller et al, 2019).…”

GEOBIA at the Terapixel Scale: Toward Efficient Mapping of Small Woody Features from Heterogeneous VHR Scenes

“…It is known that this could introduce label noise or mislabeled samples in the training set due to several factors such as misregistration, out-dated maps and databases, etc. Thus, in a future work we would like to consider label noise robust classification methods [42,43] to improve the classification performance, and also to incorporate more valuable information and efficient techniques to further reduce the computation time while still increasing the classification accuracy.…”

Remote Sensing Image Scene Classification with Noisy Label Distillation

“…After observing the features of images and division of scene classes, asymmetric noise was generated by mapping chaparral → agricultural, runway ↔ airplane, tennis court → baseball diamond, river → beach, mobile home park → parking lot, f reeway ↔ overpass, sparse residential → buildings, harbor → mobile home park, medium residential ↔ dense residential as shown in Figure 4. For NWPU-RESISC45, baseball diamond → medium residential, beach → river, dense residential ↔ medium residential, intersection → f reeway, mobile home park ↔ dense residential, overpass ↔ intersection, tennis court → medium residential, runway → f reeway, thermal power station → cloud, wetland → lake, rectangular f arm land → meadow, church → palace, commercial area → dense residential are mapped, following [12]. Figure 5 shows representative images in this dataset.…”

“…For AID, the classes are flipped by mapping bareland ↔ desert; center → storage tank; church → center, storage tank; dense residential ↔ medium residential; industrial → medium residential; meadow → f arm land; play ground → meadow, school; resort → medium residential; school → medium residential, play ground; stadium → play ground, following [12]. Figure 6 shows examples from this dataset.…”

“…Training on noisy labeled datasets become essential and has attracted much attention in recent years [7][8][9]. Furthermore, several approaches learning with noisy labels [10][11][12] have been explored for remote sensing image analysis tasks. Existing methods based on RGB images with noisy labels usually make a strong assumption that all labels are noisy.…”