Comparing the Effectiveness of Different Machine Learning Algorithms for Crop Cover Classification Using Sentinel 2

Abstract

Crop cover mapping is an essential tool for controlling and enhancing agricultural productivity. By determining the spatial distribution of different crop types, solidified judgements regarding crop planning, crop management, and risk management can be made. Crop cover classification using optical data pose constraints in terms of spatial and spectral resolution. With Sentinel – 2 data providing the ground information at 10m resolution, users may choose the best spectral band combinations and temporal frame by analysing the spectral-temporal information of different crops. The crop categorization map for the Kallakurichi and Villupuram districts were created in this study using the Random Forest (RF) and Decision tree (C5.0) classifiers. The study mainly focuses on comparing the classification accuracy of two classifiers and figuring out the best classifiers for crop cover mapping with respect to the study area. The ground truth information collected, were partitioned into calibration and validation datasets and the validation resulted with the Overall Accuracy (OA) and kappa coefficient of 66%; 0.63 and 60%; 0.67 for RF and C5.0 algorithms, respectively. From the results, it could be concluded that the RF classifier performed comparatively better than C5.0, thus making it suitable for crop cover classification.