A Booster Analysis of Extreme Gradient Boosting for Crop Classification using PolSAR Imagery

Mustafa Ustuner, Fusun Balik Sanli, Saygin Abdikan, Gokhan Bilgin, Cigdem Goksel

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

7 Citations (Scopus)

Abstract

This study evaluates the impacts of three booster types (two tree-based and one linear model) in extreme gradient boosting (XGBoost) for crop classification using multi-temporal PolSAR (Polarimetric Synthetic Aperture Radar) images. Ensemble learning algorithms have received great attention in remote sensing for classification due to their greater performance compared to single classifiers in terms of accuracy. Extreme gradient boosting is the regularized extension of traditional boosting techniques and could overcome the overfitting constrain of gradient boosting (a.k.a gradient boosting machine). Three types of booster which are linear booster, tree booster and DART (Dropouts meet Multiple Additive Regression Trees) booster were tested on XGBoost for crop classification. From the multi-temporal PolSAR data, two types of polarimetric dataset (linear backscatter coefficients and Cloude-Pottier decomposed parameters) were extracted and incorporated into the classification step. The impacts of polarimetric features for crop classification were also analyzed in detailed besides exploring the boosting types of XGBoost. Our experimental results demonstrated that tree booster and DART booster were found to be superior compared the linear booster in terms of overall classification accuracy for both polarimetric dataset. The highest classification accuracy (87.97%) was achieved by tree booster with linear backscatter coefficients. Furthermore, linear backscatter coefficients achieved higher performance with respect to Cloude-Pottier decomposition in terms of classification accuracy.

Original languageEnglish
Title of host publication2019 8th International Conference on Agro-Geoinformatics, Agro-Geoinformatics 2019
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781728121161
DOIs
Publication statusPublished - Jul 2019
Event8th International Conference on Agro-Geoinformatics, Agro-Geoinformatics 2019 - Istanbul, Turkey
Duration: 16 Jul 201919 Jul 2019

Publication series

Name2019 8th International Conference on Agro-Geoinformatics, Agro-Geoinformatics 2019

Conference

Conference8th International Conference on Agro-Geoinformatics, Agro-Geoinformatics 2019
Country/TerritoryTurkey
CityIstanbul
Period16/07/1919/07/19

Bibliographical note

Publisher Copyright:
© 2019 IEEE.

Funding

Our experimental results demonstrated that tree-based booster types provided similar accuracies to each other and outperformed linear booster. The highest classification accuracy (87.97%) was achieved by tree booster with linear backscatter coefficients When F1-score values were investigated, it was concluded that linear backscatter outperformed the Cloude-Pottier decomposition except sunflower. The main drawback of the XGBoost algorithm is the high number of parameters and their optimization. Authors addresses the need of in-depth parameter analysis of XGBoost to determine the impact of each parameter on classification accuracy. Our future research will focus on testing the model based decomposition with XGBoost for crop classification ACKNOWLEDGMENT This research has been funded by Yildiz Technical University, Scientific Research Projects Office [Project No: FBA-2017-3062]. The authors gratefully acknowledge the data support from TAGEM (Project No: TAGEM/TSKAD/14/A13/P05/03) This research has been funded by Yildiz Technical University, Scientific Research Projects Office [Project No: FBA-2017-3062]. The authors gratefully acknowledge the data support from TAGEM (Project No: TAGEM/TSKAD/14/A13/P05/03)

FundersFunder number
Scientific Research Projects Office
Yildiz Technical University, Scientific Research Projects OfficeFBA-2017-3062, TAGEM/TSKAD/14/A13/P05/03
Yildiz Teknik Üniversitesi

    Keywords

    • Agriculture
    • Crop classification
    • Extreme gradient boosting
    • Polarimetry
    • PolSAR

    Fingerprint

    Dive into the research topics of 'A Booster Analysis of Extreme Gradient Boosting for Crop Classification using PolSAR Imagery'. Together they form a unique fingerprint.

    Cite this