Temporal Vegetation Modelling Using Long Short-Term Memory Networks for Crop Identification From Medium-Resolution Multi-Spectral Satellite Images

Marc Russwurm, Marco Korner; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 11-19

Land-cover classification is one of the key problems in earth observation and extensively investigated over the recent decades. Usually, approaches concentrate on single-time and multi- or hyperspectral reflectance space- or airborne sensor measurements observed. However, land-cover classes, e.g., crops, change their reflective characteristics over time complicating classification at one observation time. Contrary, these features change in a systematic and predictive manner, which can be utilized in a multi-temporal approach. We use long short-term memory (LSTM) networks to extract temporal characteristics from a sequence of Sentinel-2 observations. We compare the performance of LSTM and other network architectures and a SVM baseline to show the effectiveness of dynamic temporal feature extraction. A large test area combined with rich ground truth labels was used for training and evaluation. Our LSTM variant achieves state-of-the art performance opening potential for further research.