A Comparison of Estimating Crop Residue Cover from Sentinel-2 Data Using Empirical Regressions and Machine Learning Methods

Authors

Yanling Ding
Hongyan Zhang
Zhongqiang Wang
Qiaoyun Xie
Yeqiao Wang, University of Rhode IslandFollow
Lin Liu
Christopher C. Hall

Document Type

Article

Date of Original Version

2020

Department

Natural Resources Science

Abstract

Quantifying crop residue cover (CRC) on field surfaces is important for monitoring the tillage intensity and promoting sustainable management. Remote-sensing-based techniques have proven practical for determining CRC, however, the methods used are primarily limited to empirical regression based on crop residue indices (CRIs). This study provides a systematic evaluation of empirical regressions and machine learning (ML) algorithms based on their ability to estimate CRC using Sentinel-2 Multispectral Instrument (MSI) data. Unmanned aerial vehicle orthomosaics were used to extracted ground CRC for training Sentinel-2 data-based CRC models. For empirical regression, nine MSI bands, 10 published CRIs, three proposed CRIs, and four mean textural features were evaluated using univariate linear regression. The best performance was obtained by a three-band index calculated using (B2 − B4) / (B2 − B12), with an R²_cv of 0.63 and RMSE_cv of 6.509%, using a 10-fold cross-validation. The methodologies of partial least squares regression (PLSR), artificial neural network (ANN), Gaussian process regression (GPR), support vector regression (SVR), and random forest (RF) were compared with four groups of predictors, including nine MSI bands, 13 CRIs, a combination of MSI bands and mean textural features, and a combination of CRIs and textural features. In general, ML approaches achieved high accuracy. A PLSR model with 13 CRIs and textural features resulted in an accuracy of R²_cv = 0.66 and RMSE_cv = 6.427%. An RF model with predictors of MSI bands and textural features estimated CRC with an R²_cv = 0.61 and RMSE_cv = 6.415%. The estimation was improved by an SVR model with the same input predictors (R²_cv = 0.67, RMSE_cv = 6.343%), followed by a GPR model based on CRIs and textural features. The performance of GPR models was further improved by optimal input variables. A GPR model with six input variables, three MSI bands and three textural features, performed the best, with R²_cv = 0.69 and RMSE_cv = 6.149%. This study provides a reference for estimating CRC from Sentinel-2 imagery using ML approaches. The GPR approach is recommended. A combination of spectral information and textural features leads to an improvement in the retrieval of CRC.

Citation/Publisher Attribution

Ding, Y.; Zhang, H.; Wang, Z.; Xie, Q.; Wang, Y.; Liu, L.; Hall, C.C. A Comparison of Estimating Crop Residue Cover from Sentinel-2 Data Using Empirical Regressions and Machine Learning Methods. Remote Sens. 2020, 12, 1470.

Available at: https://doi.org/10.3390/rs12091470

Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.