Machine Learning Based Hybrid Approach for Estimation of Jute Area Using Optical and Microwave Satellite Imagery

Jute crop holds significant economic importance contributing to environment and ecological balance. In the present study, optical and microwave satellite data was utilized to estimate jute acreage employing three distinct approaches, viz., logical criteria, machine learning and hybrid approach. Harmonized Sentinel-2 optical data products were utilized to generate monthly composite Enhanced Vegetation Index (EVI), whereas Ground Range Detected Sentinel-1 VH polarization Synthetic Aperture Radar data were used for generating monthly VH composite using median filter. Support Vector Machine (SVM), a machine learning algorithm, was adopted for this study. The EVI and VH multi-temporal profiles over the ground truth points were analysed and the criteria for identifying jute crop pixels were developed for different sowing period, i.e., March and April, during 2020 to 2023. The pre-processing of datasets and SVM was executed on the Google Earth Engine platform. The assessment of the three approaches highlighted variations in their predictive capabilities. Instead of its high classification accuracy (~ 0.80), the logical approach has its limitations in terms of complex sets of criteria and adoptability over larger extent. On the contrary, SVM classifier could develop self-learning criteria, but over-estimated the jute area with lower classification accuracy (Kappa ~ 0.45). The present study proposed a novel approach by coupling machine learning with simple logical criteria to address the limitations, while reserving its advantages. The analysis revealed that the capability of hybrid approaches in estimating the jute acreage was comparable with logical approach, with reduced operational complexities and wider adoptability over larger extent.