EVALUATING THE POTENTIAL OF 8 BAND PLANETSCOPE DATASET FOR CROP CLASSIFICATION USING RANDOM FOREST AND GRADIENT TREE BOOSTING BY GOOGLE EARTH ENGINE

Abstract

Abstract. One of the challenging task in agriculture is mapping of crops using satellite images as spectral properties of the crops looks similar to each other, and there are many complexities which are there in the field such as small land holdings, heterogeneous and often distinct field patterns. In past, mapping with high resolution satellite images was not possible due to the non-availability of data and even they were costlier. However, with the free availability of PlanetScope dataset having 8 spectral bands, with daily revisit cycle and resolution of 3m, it is now possible to carry out mapping and monitoring of agriculture crops. The objective of the study is to classify major crops of rabi (December-April) season using single date PlanetScope imagery in the Haridwar district, Uttarakhand, India. Classification of crops has been carried out using an ensemble based machine learning algorithms within Google earth engine. Off late, Google Earth Engine (GEE), a cloud based platform has attracted the attention of remote sensing analyst since it expediates the classification yielding good results of high quality. In this study, Random Forest (RF) and Gradient Tree Boosting algorithm (GTB) have been used after performing hyper-parameter tuning and results shows that overall accuracy obtained by RF and GTB are 88% and 86.5% respectively. According to the results, both classifier performed well but RF achieved 1.5% high accuracy over GTB. Analysis of the results show that highest accuracy was attained by agriculture class (wheat) while other crops class exhibited lower accuracy. In this study, RF was found to be more competent as compared to GTB in classifying the crops and PlanetScope 8 band dataset has also proved its potential in classifying crops in heterogeneous fields.