Comparison of Long Short-Term Memory Networks and Random Forest for Sentinel-1 Time Series Based Large Scale Crop Classification

Abstract

To ensure future food security, improved agricultural management approaches are required. For many of those applications, precise knowledge of the distribution of crop types is essential. Various machine and deep learning models have been used for automated crop classification using microwave remote sensing time series. However, the application of these approaches on a large spatial and temporal scale is barely investigated. In this study, the performance of two frequently used algorithms, Long Short-Term Memory (LSTM) networks and Random Forest (RF), for crop classification based on Sentinel-1 time series and meteorological data on a large spatial and temporal scale is assessed. For data from Austria, the Netherlands, and France and the years 2015–2019, scenarios with different spatial and temporal scales were defined. To quantify the complexity of these scenarios, the Fisher Discriminant measurement F1 (FDR1) was used. The results demonstrate that both classifiers achieve similar results for simple classification tasks with low FDR1 values. With increasing FDR1 values, however, LSTM networks outperform RF. This suggests that the ability of LSTM networks to learn long-term dependencies and identify the relation between radar time series and meteorological data becomes increasingly important for more complex applications. Thus, the study underlines the importance of deep learning models, including LSTM networks, for large-scale applications.

Keywords

• crop classification
• Sentinel-1
• synthetic aperture radar (SAR)
• recurrent neural network
• long short-term memory network
• random forest