A Comparative Analysis of Machine Learning Algorithms for Crop Type Mapping Using Sentinel-2 Imagery: Assessment of Spatial Distribution Patterns of Paddy and Maize in Semi-Arid Region

Abstract

Remote sensing-based classification of paddy and maize crops is challenging due to spectral similarities and complex cropping systems, particularly in semi-arid regions. This study investigates the potential of multiple machine learning algorithms i.e., Forest (RF), Support Vector Machine (SVM), XGBoost, and Light GBM for paddy and maize crops classification in Mahabubabad district of Telangana, India, using Sentinel-2 imagery. Different remote sensing datasets including multi-temporal Sentinel-2 data acquired during the kharif and rabi of 2023-2024 year andSRTM digital elevation are used for training the machine learning models. Sentinel-2 data derived vegetation indices and phenological metrics, and SRTM DEM derived slope parameters are used along with 850 georeferenced crop sites for training and validation of the models.The performance of these methods are assessed using different accuracy measures. The results indicate, XGBoost outperformed other machine learning models with overall accuracy 92.3%, followed by RF and other methods. The spatial pattern analysis of classification accuracy depicts classification errors are mainly related to field size and crop phenology. This study highlights the usefulness of machine learning approaches in classifying staple crops such as paddy and maize. Additionally provides insights on the impact of different auxiliary parameters on classification accuracy.Overall the framework implemented in this study can be useful for enhancing the accuracy of crop type mapping in other regions with similar agro-ecological conditions.