Crop Type Identification and Mapping of Pothohar Region using Machine Learning Algorithms and Sentinel 2 Data

Abstract

Mapping crop types and analyzing their characteristics has been a long-standing practice to ensure food security, utilize land effectively, and shape agricultural policy. This has become a crucial task for many nations due to its impact on economics of the country, trade, and global food security. This topic is widely study in policy making, land management and economics. Most of the countries are now utilizing satellite imagery, which offers a comprehensive view, multiple time coverage, and cost efficiency, for crop identification and analysis. Various methods have been tried to gather crops information but now satellite data with its improved accuracy and resolution used to generate crop type maps. Remote sensing techniques using optical sensors have become a significant tool for understanding vegetation structure and properties. This study focuses on using optical imagery from Sentinel-2 to map crop types. The study tells how to use sentinel 2 imagery, pre-processing, and analysis of Sentinel-2 data using the ArcMap, ERDAS Imagine and R code in RStudio. The thesis outlines a plan for categorizing crops using data collection techniques, machine learning algorithms, and validation methods to achieve accurate results on a nationwide scale. We used a machine learning RF classification algorithm and SVM to create accurate crop type maps using (S2) data for growing season of Rabi and Kharif in pothohar region which consists of five districts Attock, Jehlum, Rawalpindi, Islamabad, and Chakwal. The algorithms have shown improved accuracy in previous studies by using sample data and ensemble decision trees. The purpose of the study is to compare the precision of crop classification between Support Vector Machine (SVM) and Random Forest (RF) algorithms. The overall accuracy of Rabi maps for SVM is 96.49% and for RF 91.22% with kappa coefficient 0.94 and 0.86 respectively. The overall accuracy of Kharif maps for support vector machine is 95.38% and for RF 89.23% with kappa coefficient 0.93 and 0.84 respectively. The study compares the two crop classification methods. These results show that with point training data SVM results in better accuracy than random forest. The study found that the machine learning methods (RF & SVM) showed improved robustness in solving multiclassification problems, achieving a user accuracy of greater than 90% in SVM and more than 85% in RF algorithm for each crop type.