Assessing the Sentinel-2 Capabilities to Identify Abandoned Crops Using Deep Learning

Abstract

The termination or interruption of agro-forestry practices for a long period gradually results in abandoned land. Abandoned land parcels do not match the requirements to access to the basic payment of the European Common Agricultural Policy (CAP). Therefore, the identification of those parcels is key in order to return fair subsidies to farmers. In this context, the present work proposes a methodology to detect abandoned crops in the Valencian Community (Spain) from remote sensing data. The approach is based on the assessment of multitemporal Sentinel-2 images and derived spectral indices, which are used as predictors for training machine learning and deep learning classifiers. Several classification scenarios, including both abandoned and active parcels, were evaluated. The best results (98.2% overall accuracy) were obtained when a bi-directional Long Short Term Memory (BiLSTM) network was trained with a multitemporal dataset composed of twelve reflectance time series, and a derived bare soil spectral index (BSI). In this scenario we were able to effectively distinguish abandoned crops from active ones. The results revealed Sentinel-2 features are well suited for land use identification including abandoned lands, and open the possibility of implementing this type of remote sensing based methodology into the CAP payments supervision.