Supporting the Common Agricultural Policy with Sentinel-2 data and deep recurrent networks

The 2020+ Common Agricultural Policy encourages the use of Copernicus remote sensing data for the monitoring of agricultural parcels. In this work, a procedure for automatic identification of land use from remote sensing data is proposed. The approach includes the use of spectral information of Sentinel-2 time series over the Valencia province (Spain) during the agronomic year 2027/2018, and deep learning recurrent networks. In particular, a bi-directional Long Short Term Memory (Bi-LSTM) network was trained to classify active land uses and abandoned lands. A comparison exercise was undertaken to assess the classification power of the Bi-LSTM as compared to the random forest (RF) algorithm. The Bi-LSTM network outperformed the RF, and provided and overall accuracy of 97.5% when discriminating eleven land uses including abandoned lands. The results suggest the proposed methodology could potentially be implemented in an automated procedure to supervise the CAP requirements to access subsidies. In addition, the classification process also supports the continuous update of the Land Parcel Identification System (LPIS), which allows paying agencies to uniquely identify land parcels in space, store records of land uses (and assess its evolution), and ultimately ease the declaration procedure to both farmers and paying agencies.