A Comparative Study and Machine Learning Enabled Efficient Classification for Multispectral Data in Agriculture

Abstract

Reliable and accurate crop maps are required for food security from regional to global scale. The increased availability of satellite imagery leads to a “Big Data” problem while producing crop maps. Now, cloud-based platforms have gained a lot of attention for crop classification over large regions. The main goal of the research is to analyze crop classification using various machine learning (ML) such as Support Vector Machine (SVM), Gradient Tree Boosting (GTB), Random Forest (RF), Decision Tree (DT) as well as Classification and Regression Trees (CART) on Google Earth Engine platform. The aim is to explore the Google Earth Engine’s efficiency (GEE) when classification different crops using multi- spectral datasets of Sentinel 2 MSI and Landsat 8 OLI satellites for crop mapping of Mathura district of Uttar Pradesh, India. The best cloud free image (less than 5%) of Landsat 8 OLI and Sentinel 2 MSI datasets ("2020-12-26","2020-12-30") were used for crop classification with the help of automatic filtering i.e. percentage cloud property on the GEE platforms. Moreover that GEE platform perform, acquiring, clarifying as well as preprocessing of satellite dataset could be organized very powerfully. Points as feature spaces were used like training datasets. Furthermore confusion matrixes are used for accuracy assessment (producer and user accuracy) and kappa coefficient. Additionally compare the outcome of the dataset on the basis of overall accuracy (OA), F1 score as well as kappa coefficient. The highest OA is found using GTB (86.7%) followed by RF (82.5%), CART (81.0%), DT (78.1%) and SVM (66.5%) for Landsat 8 OLI image. For the Sentinel 2 image, GTB achieved the highest OA of 84.2% followed by SVM (84%), RF (82.3%), DT (75.2%), and CART (75. 0%) respectively. On the basis of research, found that GTB performed well among all the classifiers to crop mapping using both multi-spectral datasets.

Keywords

• Google Earth Engine (GEE), Machine learning (ML), Remote Sensing (RS), Satellite imagery