A Progressive Crop-Type Classification Using Multitemporal Remote Sensing Data and Phenological Information

Abstract

Local parameters for climate modelling are highly dependent on crop types and their phenological growth stage. The land cover change of agricultural areas during the growing season provides important information to distinguish crop types. The presented progressive classification algorithm identifies crop types based on their phenological development and their corresponding reflectance characteristics in multitemporal satellite images of the four sensors Landsat-7 and -8, Sentinel-2A and RapidEye. It distinguishes crop types not only retrospectively, but progressively during the growing season starting in early spring. Binary fuzzy c-means clustering differentiates seven crop types in eight decisions at particular time periods. These decisions are based on expert knowledge about plant characteristics in different phenological stages. The unsupervised classification approach and previously defined decisions enable the algorithm to work independently of training data. The fuzzy approach provides certainties of crop-type existence and generates first results in early spring. The accuracy and reliability of the classification results improve with increasing time. The method is developed at the German Durable Environmental Multidisciplinary Monitoring Information Network (DEMMIN). The study area is an official test site of the Joint Experiment of Crop Assessment and Monitoring (JECAM) and is located in an intensely agricultural used area in Northern Germany. Classification results were produced for the growing seasons 2015 and 2016. The overall accuracy in 2015 amounted to 89.49%. A challenge remains the distinct separation of wheat and rye, whereas barley, rapeseed, potato, corn and sugar beet are classified with high accuracies. The overall accuracy in 2016 was lower (77.19%) due to unfavourable weather conditions.