Satellite remote sensing-based crop cover classification over Europe

Abstract

Crop maps play an important role in a variety of applications, from calculating crop areas and forecasting food production quantities to the analysis of agri-environmental interactions, highlighting the necessity of timely and accurate information on agricultural land use. The availability of remote sensing data has permitted numerous crop classification studies, which have investigated a variety of methods to improve classification performance, such as the selection of remote sensing sources, classification algorithms, and preprocessing methods. This paper compares these approaches with respect to classification accuracy in a European context. The study also investigates aspects such as classification level, study area division, and class granularity. The review shows that optical products provide more information for crop identification than radar products, however, combining optical data with radar backscatter increases accuracy. Classification accuracy benefits from specific features such as red-edge and spectral indices for optical products and Haralick textures for radar. Compared to traditional machine learning and distance-based classification methods, deep learning algorithms have been shown to achieve superior performance. Nevertheless, random forest's comparative accuracy at relatively low computational cost makes it a viable alternative for largescale applications. Finally, preprocessing methods and data on topography, climate, and crop growth patterns appear to improve accuracy.