APPLICATION OF TEMPORAL CONVOLUTIONAL NEURAL NETWORK FOR THE CLASSIFICATION OF CROPS ON SENTINEL-2 TIME SERIES
Keywords: deep learning, multi-temporal classification, sequence data, crop classification, Sentinel-2
Abstract. The recent development of Earth observation systems – like the Copernicus Sentinels – has provided access to satellite data with high spatial and temporal resolution. This is a key component for the accurate monitoring of state and changes in land use and land cover. In this research, the crops classification was performed by implementing two deep neural networks based on structured data. Despite the wide availability of optical satellite imagery, such as Landsat and Sentinel-2, the limitations of high quality tagged data make the training of machine learning methods very difficult. For this purpose, we have created and labeled a dataset of the crops in Slovenia for the year 2017. With the selected methods we are able to correctly classify 87% of all cultures. Similar studies have already been carried out in the past, but are limited to smaller regions or a smaller number of crop types.