Search

This article presents TimeSen2Crop, a pixel-based dataset made up of more than 1 million samples of Sentinel 2 time series (TSs) associated to 16 crop types. This dataset, publicly available, aims to contribute to the worldwide research related to the supervised classification of TSs of Sentinel 2 data for crop type mapping. TimeSen2Crop includes atmospherically corrected images and reports the snow, shadows, and clouds information per labeled unit. The provided TSs represent an agronomic year (i.e., period from one year's harvest to the next one for agricultural commodity) ranging from September 2017 to August 2018. To generate the dataset, the publicly available Austrian crop type map based on farmer's declarations has been considered. To ensure the selection of reliable labeled units from the map (i.e., pure pixels correctly associated to their labels), an automatic procedure for the extraction of the training set based on a multitemporal deep learning model has been defined. TimeSen2Crop also includes a TS of Sentinel 2 images acquired in the following agronomic year (i.e., from September 2018 to August 2019). These data are provided with the aim of attract more research activities for solving a typical challenge of the crop type mapping task: adapting multitemporal deep learning models to different year (domain adaptation). The design of the dataset is described along with a benchmark comparison of deep learning models for crop type mapping.

This work presents a system for multi-year crop type mapping based on the multi-temporal Long Short-Term Memory (LSTM) Deep Learning (DL) model and Sentinel 2 image Time Series (TS). The method assumes the availability of a pre-trained LSTM model for a given year and aims to update the corresponding crop type map fora different year considering a small amount of recent reference data. To this end, the proposed approach combines Self-Paced Learning (SPL) and fine-tuning (FT) techniques. While the SPL technique gradually incorporates samples from crop types that can be classified with high-confidence by the pre-trained model, the FT strategy adapts the network to those classes having low-confidence accuracy. This condition allows us to reduce the labeled samples required to achieve accurate classification results. The experimental results obtained on three tiles of the Austrian country on TSs of Sentinel 2 data acquired in 2019 and 2020 (considering a model pre-trained on images of 2018) demonstrate the capability of the LSTM to adapt to TS of images with different temporal and radiometric characteristic with respect to the one used to pre-train the model, with a relatively small number of training samples. As expected, by directly applying the model without performing any adaptation, we obtain a mean F-score (F1%) of 64% and 62% compared to 76% and 70% achieved by the proposed technique with only 1500 samples for 2019 and 2020, respectively.

The regular monitoring of agricultural areas is extremely important for mitigating food insecurity risks and for planning government interventions. In the literature, several deep learning algorithms have been recently proposed to perform land cover/ land use classification by using multispectral optical images. However, most of the considered deep learning models, such as the standard Convolutional Neural Networks (CNN), rely on mono-temporal images, focusing on spectral and textural features while discarding the temporal component, which is crucial for the accurate crop type mapping. In this work, we exploit a Long Short Term Memory (LSTM) deep learning classification architecture to characterize agricultural area dynamics by using the multitemporal multispectral information provided by satellite multispectral sensor Sentinel 2. Instead of considering a pre-trained network and applying to it a fine-tuning, the proposed architecture is trained from scratch in order to be tailored to the specific properties of the long time series of Sentinel 2 multispectral images. To face the lack of labeled training database, existing crop type maps available at the country level are used to generate a large set of weak reference data. First, the proposed method automatically extracts a large training dataset from existing crop type maps, by detecting those samples having the highest probability of being correctly classified. Then, the weak labeled samples extracted are used to train the deep LSTM architecture on a time series of Sentinel 2 images acquired over an entire year. The preliminary results obtained demonstrate the effectiveness of the proposed approach, which is promising at large scale.