Latest remote sensing sensors are capable of acquiring high spatial and spectral Satellite Image Time Series (SITS) of the world. These image series are a key component of classification systems that aim at obtaining up-to-date and accurate land cover maps of the Earth’s surfaces. More specifically, current SITS combine high temporal, spectral and spatial resolutions, which makes it possible to closely monitor vegetation dynamics. Although traditional classification algorithms, such as Random Forest (RF), have been successfully applied to create land cover maps from SITS, these algorithms do not make the most of the temporal domain. This paper proposes a comprehensive study of Temporal Convolutional Neural Networks (TempCNNs), a deep learning approach which applies convolutions in the temporal dimension in order to automatically learn temporal (and spectral) features. The goal of this paper is to quantitatively and qualitatively evaluate the contribution of TempCNNs for SITS classification, as compared to RF and Recurrent Neural Networks (RNNs) —a standard deep learning approach that is particularly suited to temporal data. We carry out experiments on Formosat-2 scene with 46 images and one million labelled time series. The experimental results show that TempCNNs are more accurate than the current state of the art for SITS classification. We provide some general guidelines on the network architecture, common regularization mechanisms, and hyper-parameter values such as batch size; we also draw out some differences with standard results in computer vision (e.g., about pooling layers). Finally, we assess the visual quality of the land cover maps produced by TempCNNs.

Supervised classification systems used for land cover mapping require accurate reference databases. These reference data come generally from different sources such as field measurements, thematic maps, or aerial photographs. Due to misregistration, update delay, or land cover complexity, they may contain class label noise, i.e., a wrong label assignment. This study aims at evaluating the impact of mislabeled training data on classification performances for land cover mapping. Particularly, it addresses the random and systematic label noise problem for the classification of high resolution satellite image time series. Experiments are carried out on synthetic and real datasets with two traditional classifiers: Support Vector Machines (SVM) and Random Forests (RF). A synthetic dataset has been designed for this study, simulating vegetation profiles over one year. The real dataset is composed of Landsat-8 and SPOT-4 images acquired during one year in the south of France. The results show that both classifiers are little influenced for low random noise levels up to 25%–30%, but their performances drop down for higher noise levels. Different classification configurations are tested by increasing the number of classes, using different input feature vectors, and changing the number of training instances. Algorithm complexities are also analyzed. The RF classifier achieves high robustness to random and systematic label noise for all the tested configurations; whereas the SVM classifier is more sensitive to the kernel choice and to the input feature vectors. Finally, this work reveals that the cross-validation procedure is impacted by the presence of class label noise.