A Discriminative Long Short Term Memory Network with Metric Learning Applied to Multispectral Time Series Classification

Abstract

In this article, we propose an end-to-end deep network for the classification of multi-spectral time series and apply them to crop type mapping. Long short-term memory networks (LSTMs) are well established in this regard, thanks to their capacity to capture both long and short term temporal dependencies. Nevertheless, dealing with high intra-class variance and inter-class similarity still remain significant challenges. To address these issues, we propose a straightforward approach where LSTMs are combined with metric learning. The proposed architecture accommodates three distinct branches with shared weights, each containing a LSTM module, that are merged through a triplet loss. It thus not only minimizes classification error, but enforces the sub-networks to produce more discriminative deep features. It is validated viaBreizhcrops, a very recently introduced and challenging time series dataset for crop type mapping.