Recent satellite missions have led to a huge amount of Earth observation data, most of them being freely available. In such a context, satellite image time series have been used to study land use and land cover information. However, optical time series, such as Sentinel-2 or Landsat ones, are provided with an irregular time sampling for different spatial locations, and images may contain clouds and shadows. Thus, preprocessing techniques are usually required to properly classify such data. The proposed approach is able to deal with irregular temporal sampling and missing data directly in the classification process. It is based on Gaussian processes and allows to perform jointly the classification of the pixel labels as well as the reconstruction of the pixel time series. The method complexity scales linearly with the number of pixels, making it amenable in large-scale scenarios. Experimental classification and reconstruction results show that the method does not compete yet with state-of-the-art classifiers but yields reconstructions that are robust with respect to the presence of undetected clouds or shadows and does not require any temporal preprocessing.
5 January 2023 in IEEE Transactions on Geoscience and Remote Sensing
In this article, we propose an approach based on Gaussian Processes (GP) for large scale land cover pixel-based classification with Sentinel-2 satellite image time-series (SITS). We used a sparse approximation of the posterior combined with variational inference to learn the GP’s parameters. We applied stochastic gradient descent and GPU computing to optimize our GP models on massive data sets. The proposed GP model can be trained with hundreds of thousands of samples, compared to few thousands for traditional GP methods. Moreover, we included the spatial information by adding the geographic coordinates into the GP’s covariance function to efficiently exploit the spatio-spectro-temporal structure of the SITS. We ran experiments with Sentinel-2 SITS of the full year 2018 over an area of 200 000 km 2 (about 2 billion pixels) in the south of France, which is representative of an operational setting. Adding the spatial information significantly improved the results in terms of classification accuracy. With spatial information, GP models have an overall accuracy of 79.8. They are more than three points above Random Forest (the method used for current operational systems) and more than one point above a multi-layer perceptron. Compared to a Transformer-based model (which provides state of the art results in the literature, but are not applied in operational systems), GP models are only one point below.
Mapping forest composition using multiseasonal optical time series remains a challenge. Highly contrasted results are reported from one study to another suggesting that drivers of classification errors are still under-explored. We evaluated the performances of single-year Formosat-2 time series to discriminate tree species in temperate forests in France and investigated how predictions vary statistically and spatially across multiple years. Our objective was to better estimate the impact of spatial autocorrelation in the validation data on measurement accuracy and to understand which drivers in the time series are responsible for classification errors. The experiments were based on 10 Formosat-2 image time series irregularly acquired during the seasonal vegetation cycle from 2006 to 2014. Due to lot of clouds in the year 2006, an alternative 2006 time series using only cloud-free images has been added. Thirteen tree species were classified in each single-year dataset based on the Support Vector Machine (SVM) algorithm. The performances were assessed using a spatial leave-one-out cross validation (SLOO-CV) strategy, thereby guaranteeing full independence of the validation samples, and compared with standard non-spatial leave-one-out cross-validation (LOO-CV). The results show relatively close statistical performances from one year to the next despite the differences between the annual time series. Good agreements between years were observed in monospecific tree plantations of broadleaf species versus high disparity in other forests composed of different species. A strong positive bias in the accuracy assessment (up to 0.4 of Overall Accuracy (OA)) was also found when spatial dependence in the validation data was not removed. Using the SLOO-CV approach, the average OA values per year ranged from 0.48 for 2006 to 0.60 for 2013, which satisfactorily represents the spatial instability of species prediction between years.
18 December 2023 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
In this article, we propose a method exploiting irregular and unaligned Sentinel-2 satellite image time series (SITS) for large-scale land cover pixel-based classification. We perform end-to-end learning by combining a time and space informed kernel interpolator with a Sparse Variational Gaussian Processes (SVGP) classifier. The interpolator embeds irregular and unaligned SITS onto a fixed and reduced size latent representation. The spatial information is taken into account by using a spatial positional encoding . The obtained latent representation is given to the SVGP classifier and all the parameters are jointly optimized w.r.t. the classification task. We run experiments with irregular and unaligned Sentinel-2 SITS of the full year 2018 over an area of 200 000 $\rm{km}^{2}$ (about two billion pixels) in the south of France (27 MGRS tiles). Such experimental condition exacerbates the irregular and unaligned issues of SITS. In terms of overall accuracy, with the learned latent representation instead of linearly interpolated SITS, the results of the SVGP classifier are improved by about 10 points. Moreover, with the learned latent representation, the SVGP classifier outperforms the main state-of-the-art methods from the literature at large scale (e.g., seven points for the Multi-layer Perceptron) and is robust to the available timestamps used for training and testing.
3 February 2020 in IEEE Transactions on Geoscience and Remote Sensing
Hyperspectral unmixing aims at identifying a set of elementary spectra and the corresponding mixture coefficients for each pixel of an image. As the elementary spectra correspond to the reflectance spectra of real materials, they are often very correlated, thus yielding an ill-conditioned problem. To enrich the model and reduce ambiguity due to the high correlation, it is common to introduce spatial information to complement the spectral information. The most common way to introduce spatial information is to rely on a spatial regularization of the abundance maps. In this article, instead of considering a simple but limited regularization process, spatial information is directly incorporated through the newly proposed context of spatial unmixing. Contextual features are extracted for each pixel, and this additional set of observations is decomposed according to a linear model. Finally, the spatial and spectral observations are unmixed jointly through a cofactorization model. In particular, this model introduces a coupling term used to identify clusters of shared spatial and spectral signatures. An evaluation of the proposed method is conducted on synthetic and real data and shows that results are accurate and also very meaningful since they describe both spatially and spectrally the various areas of the scene.