A Comparison Of Fully Convolutional And Recurrent Networks For Multi-Temporal Crop Recognition Using SAR Images

With population and food consumption continuously growing, the demand for efficient agricultural crop monitoring systems has been increasing in the last years. Crop dynamics are inherently complex and to model them both spatial and temporal context have to be considered. The increasing availability of timely, precise and cost-effective Remote Sensing data along with the recent development of deep learning techniques for image analysis open up new possibilities for crop monitoring. Motivated by this context, this work presents a comparative analysis of three deep learning architectures for crop recognition: Fully Convolutional Networks, Recurrent Neural Networks and Convolutional Recurrent Neural Networks. The paper reports the results of experiments performed over two datasets: a temperate region near Hanover, Germany and a sub-tropical region in Campo Verde, Brazil. Only SAR data from Sentinel-1 satellite was considered because it is marginally affected by atmospheric conditions. The experiments showed that the tested models achieved state-of-the-art performance.