Crop mapping using machine learning algorithms and dual-polarized indices derived from multi-temporal Sentinel-1 data

ABSTRACT

Introduction: In recent years, with increasing pressure on natural resources and the need for sustainable use of agricultural land, precision agriculture has gained growing importance as an efficient approach for smart resource management. One of the fundamental challenges in this field is the accurate identification of crop types and monitoring their growth stages at appropriate temporal and spatial scales. Remote sensing technology, particularly Synthetic Aperture Radar (SAR), enables the extraction of precise ground information under various weather conditions. Unlike optical data, which are highly dependent on illumination and often limited in cloudy regions, radar datasuch as Sentinel-1 imagery provide effective tools for agricultural monitoring with their all-weather, day-and-night imaging capability. In this context, the present study aims to investigate the potential of Sentinel-1 radar time series data and its VV and VH polarizations, along with derived polarimetric indices, for accurate crop classification in an agricultural area located in the suburbs of Mashhad. The main focus of this research is to evaluate the capability of different machine learning algorithms combined with radar data to enhance classification accuracy and achieve more precise crop identification.

Materials and Methods: For this study, Sentinel-1 radar time series data with VV and VH polarizations, covering the period from winter 2021 to spring 2022, were used. The study area includes agricultural lands in Mashhad County, characterized by a variety of crops such as wheat, chickpea, alfalfa, as well as non-cropland areas. Through radar data processing, four key polarimetric indices NRPB (Noise-to-Backscatter Ratio), DPDD (Dual-Polarization Difference), IDPDD (Integrated Dual-Polarization Differential), and VDDPI (Temporal Variation Index for VV and VH) were extracted and employed in combination with the original data for classification purposes.

Three powerful machine learning algorithms XGBoost, Random Forest (RF), and Support Vector Machine (SVM) were applied for crop classification. Training samples were collected for seven defined classes within the study area, and classification accuracy was evaluated using the error matrix, Kappa coefficient, and overall accuracy.

Results and Discussion: The results of modeling and classification, which were validated using field data associated with the coordinates of each plot, showed that the XGBoost and RF algorithms performed significantly better than the SVM algorithm. The overall accuracy and Kappa for the XGBoost model were 83.48% and 0.78, respectively, and for the RF model were 82.27% and 0.78, whereas the SVM algorithm achieved an overall accuracy and Kappa of 61.46% and 0.51. This performance difference is primarily attributed to the superior ability of tree-based algorithms to model complex and nonlinear relationships between features and classes.

Among the polarimetric indices, DPDD and IDPDD demonstrated distinct temporal behaviors during different crop growth stages, proving highly valuable for phenological crop discrimination. Crops such as alfalfa, chickpea, and wheat were classified with higher accuracy and less confusion by XGBoost and RF, whereas SVM struggled to separate classes with similar vegetation cover, leading to substantial overlaps between certain crop types. Among the polarimetric indices, DPDD and IDPDD demonstrated distinct temporal behaviors during different crop growth stages, proving highly valuable for phenological crop discrimination. Crops such as alfalfa, chickpea, and wheat were classified with higher accuracy and less confusion by XGBoost and RF, whereas SVM struggled to separate classes with similar vegetation cover, leading to substantial overlaps between certain crop types.

Conclusion: This study clearly demonstrated that Sentinel-1 radar data particularly VV and VH polarizations combined with time-series derived polarimetric indices, hold strong potential for accurate crop classification. The integration of these data with advanced machine learning algorithms, especially XGBoost and RF, can provide reliable alternatives to traditional optical-based methods, particularly in cloudy regions or areas with limited access to optical data. Moreover, the findings, in line with similar international studies, highlight the importance and effectiveness of polarimetric indices as key tools for periodic crop monitoring. Utilizing these indices alongside time-series data represents a significant step toward optimizing agricultural land management, enhancing productivity, and promoting sustainable development in the agricultural sector. Consequently, the adoption of modern remote sensing technologies and machine learning will play a pivotal role in shaping the future of smart and sustainable agriculture.