Dryland Crop Classification Combining Multitype Features and Multitemporal Quad-Polarimetric RADARSAT-2 Imagery in Hebei Plain, China.

Di Wang,Tian Tian,Zheng Sun,Chang-An Liu,Yan Zeng

doi:10.3390/s21020332

Abstract

The accuracy of dryland crop classification using satellite-based synthetic aperture radar (SAR) data is often unsatisfactory owing to the similar dielectric properties that exist between the crops and their surroundings. The main objective of this study was to improve the accuracy of dryland crop (maize and cotton) classification by combining multitype features and multitemporal polarimetric SAR (PolSAR) images in Hebei plain, China. Three quad-polarimetric RADARSAT-2 scenes were acquired between July and September 2018, from which 117 features were extracted using the Cloude–Pottier, Freeman–Durden, Yamaguchi, and multiple-component polarization decomposition methods, together with two polarization matrices (i.e., the coherency matrix and the covariance matrix). Random forest (RF) and support vector machine (SVM) algorithms were used for classification of dryland crops and other land-cover types in this study. The accuracy of dryland crop classification using various single features and their combinations was compared for different imagery acquisition dates, and the performance of the two algorithms was evaluated quantitatively. The importance of all investigated features was assessed using the RF algorithm to optimize the features used and the imagery acquisition date for dryland crop classification. Results showed that the accuracy of dryland crop classification increases with evolution of the phenological period. In comparison with SVM, the RF algorithm showed better performance for dryland crop classification when using full polarimetric RADARSAT-2 data. Dryland crop classification accuracy was not improved substantially when using only backscattering intensity features or polarization decomposition parameters extracted from a single-date image. Satisfactory classification accuracy was achieved using 11 optimized features (derived from the Cloude–Pottier decomposition and the coherency matrix) from 2 RADARSAT-2 images (acquisition dates corresponding to the middle and late stages of dryland crop growth). This study provides an important reference for timely and accurate classification of dryland crop in Hebei plain, China.

Highlights

IntroductionSatellite-based remote-sensing techniques have proven very effective for mapping crops owing to their advantages of wide spatial coverage, periodic observation, high efficiency, and low cost [3,4,5]
In terms of the classification accuracy, the overall accuracy (OA) and Kappa coefficient of each of the four typical ground objects are not high on the three different dates, i.e., they vary in the ranges of 77.31–81.20% and 0.619–0.687%, respectively, whichever algorithm was used for the classification
An in-depth investigation was conducted into the accuracy of the classification of two typical dryland crops in Jizhou county, Hebei plain, China using satellite-based polarimetric SAR (PolSAR) imagery

Summary

Introduction

Satellite-based remote-sensing techniques have proven very effective for mapping crops owing to their advantages of wide spatial coverage, periodic observation, high efficiency, and low cost [3,4,5]. Optical imagery such as that acquired by the Landsat-8, Sentinel-2, and Gaofen-1 satellites has become the main data source for classifying crop types and determining their spatial distribution [6]. Cultivated land is a predominant land use type and dryland crops, such as maize and cotton, have been widely planted in Hebei plain for a long time

Objectives

Methods

Results

Discussion

Conclusion