Crop Type Classification based on Machine Learning with Multitemporal Sentinel-1 Data

Abstract

The amount of open satellite data has increased tremendously in recent years, simultaneously with a continuing decrease in the price of high performance cloud computing. This can be combined with machine learning methods to perform crop type classification in the agricultural sector. In this paper, we propose a data processing chain for processing multitemporal Sentinel-1 SAR data, and show how the temporal patterns of agricultural fields can be visualized to provide a valuable overview prior to classification. We then investigate the performance of 6 machine learning methods for crop type classification of 12 crop types based on 44333 fields, and achieve an overall accuracy of (94.02 ± 0.25)% with an RBF SVM classifier. The dataset used is a subset of all the fields of the chosen crop types in Denmark in 2019, which comprises a total of 289810, or 49.34% of all fields in the country for the 2019 season. The entire data processing chain is based on open data and free open source software, thereby minimizing the cost of practical applications and future work for both industry and academia. All code used for the paper is available on GitHub.