Multi-temporal RADARSAT-2 polarimetric SAR for maize mapping supported by segmentations from high-resolution optical image

https://doi.org/10.1016/j.jag.2018.08.021Get rights and content

Highlights

  • Integrates advantages of the microwave and optical satellite data.

  • Optical image is better at representing maize field boundaries than the SAR image.

  • Both off or in-season optical images are useful to increase the flexibility of our proposed method.

  • Time-series PolSAR images are able to distinguish maize from other crops.

  • Random Forest can be used to reduce data redundancy without reducing accuracy.

Abstract

Due to its ability to penetrate the cloud, Synthetic Aperture Radar (SAR) has been a great resource for crop mapping. Previous research has verified the applicability of SAR imagery in object-oriented crop classification, however, speckle noise limits the generation of optimal segmentation. This paper proposed an innovative SAR-based maize mapping method supported by optical image, Gaofen-1 PMS, based segmentation, named as parcel-based SAR classification assisted by optical imagery-based segmentation (os-PSC). Polarimetric decomposition was applied to extract polarimetric parameters from multi-temporal RADARSAT-2 data. One Gaofen-1 image was then used for parcel extraction, which was the basic unit for SAR image analysis. The final step was a multi-step classification for final maize mapping including: the potential maize mask extraction, pure/mixed maize parcel division and an integrated maize map production. Results showed that the overall accuracy of the os-PSC method was 89.1%, higher than those of pixel-level classification and SAR-based segmentation methods. The comparison between optical- and SAR-based segmentation demonstrated that optical-based segmentation would be better at representing maize field boundaries than the SAR-based segmentation. Moreover, the parcel- and pixel-level integrated classification will be suitable for many agricultural systems with small landownership where inter-cropping is common. Through integrating advantages of the SAR and optical data, os-PSC shows promising potentials for crop mapping.

Introduction

Maize (Zea mays L.) is one of the dominant dryland crops growing through summer to fall in Northern China (Wang et al., 2007). The collection of multi-resolution or multi-temporal optical remote sensing images has been demonstrated to be successfully used to map accurate maize distribution (Maxwell et al., 2004, 2006; Avci and Sunar, 2015). These studies provide advanced classification methods for large-scale operation systems such as Crop Data Layer (CDL), a popular product issued annually from National Agricultural Statistics Service, NASS (Boryan et al., 2011). However, optical remote sensing meets inevitable difficulties caused by cloud contamination. It is almost impossible to collect enough cloud-free images for large-scale operational applications due to frequent occurrence of clouds during the crop growing season (Chen et al., 2004; Carrão et al., 2008; Potgieter et al., 2010).

Instead of recording reflected solar radiance by optical sensors, Synthetic Aperture Radar (SAR) is an active technique that emits and receives electromagnetic waves in the microwave range, from K- to P-band, which can penetrate clouds to record backscattered energy reflected from the Earth’s surface. As an alternative technology to detect land surface, SAR shows great potentials to map the crop planting area (Blaes et al., 2005; McNairn et al., 2010). Early orbital radar systems, such as ERS-1/2 and RADARSAT-1, have been used to identify major crops, such as maize, soybean, and wheat (Eric et al., 1997; Ban, 2014). In these studies, SAR systems provided only single-frequency and single-polarization features, and the limited information were insufficient due to similar backscattering characteristics among many crops. Many enhanced SAR sensors have been launched recently to acquire multi-polarization or fully polarimetric SAR (PolSAR) data, recording various backscatter responses to the biophysical structure of the crop canopy. Voluminous studies have applied time-series PolSAR images for crop mapping and results showed that PolSAR data can reach significantly higher classification accuracy than single-polarization SAR (Skriver et al., 1999; Hoekman and Vissers, 2003; Haldar et al., 2012). Moreover, polarimetric decomposition theorems have been developed to separate the polarimetric measurements from a mixture of scatters into independent elements (Cloude and Pottier, 1996, 1997; Freeman and Durden, 1998). McNairn et al. (2009) applied three decomposition approaches (Cloude-Pottier, Freeman-Durden, and Krogager) to PALSAR images for crop detection and obtained results superior to those derived from linear polarization data. Lopez-Sanchez et al. (2014) exploited time-series variation of linear backscattering coefficients and other polarimetric parameters derived from multi-temporal RADARSAT-2 data at multiple phenological stages and proposed a simple approach to estimate crop phenological stages. The good agreement between results and ground measurements verified the capability of C-band SAR imagery to retrieve paddy rice growth phenology. Zhao et al. (2014) tested the sensitivity of high-frequency polarimetric parameters to the crop harvest patterns and produced satisfied overall accuracies that reached 97.20% and 93.81% for wheat and rapeseed, respectively.

With more satellites launched, the combination of optical and SAR images is widely suggested to improve land surface monitoring. The complementary information from each source enhance the discrimination among various land-covers. Laurin et al. (2013) integrated Landsat TM and ALOS PALSAR images for forest mapping in west Africa, both maximum likelihood (MLC) and neural networks (NN) classifiers produced high accuracies. Also, optical and SAR data were fused in the urban context for impervious surface estimation using random forest algorithm (Zhang et al., 2014), which reduced the confusions between impervious surface and bare soil, as well as shaded area and water surface. Inglada et al. (2016) applied time-series SAR and optical image to improve early crop type identification, results indicated that the use of SAR imagery allows users to use optical data without gap-filling yielding results.

Despite these achievements, most classification algorithms are implemented at pixel-based level and reach low classification accuracies (Niu and Ban, 2012). The issue is caused by the inherent noise in SAR images from wave interference. Layerstack between SAR and optical images at pixel or feature-level also cannot avoid this issue. Therefore, object-oriented classification (OOC) has been used to reduce radar speckle by introducing more spectral and spatial features such as mean of backscattering coefficients, texture and contexture information (Qi et al., 2012). Qi et al. (2012) demonstrated that the object-based method could reduce speckle effects substantially by combining textural information, getting a satisfactory result with an overall accuracy of 86.6%. Mitchell et al. (2014) tested the capability of C- and L-band SAR to produce a comparable estimate of forest/non-forest cover by integrating multi-scale segmentations. So far, there are numerous algorithms for SAR image segmentation, including clustering-, edge-, and region-based methods (Germain and Refregier, 2001; Zhang et al., 2008; Carvalho et al., 2010; Li et al., 2010; Wang et al., 2010). In Ban and Jacob (2013), the Sobel filter was performed on SAR and optical data individually, and a majority voting rule was applied to integrate all edge images. These results showed that the proposed method achieved better performance on segmentation than eCognition software, particularly for built-up and linear features. Deng et al. (2014) performed hierarchical segmentation on multi-temporal RADARSAT-2 data utilizing integration Stationary Wavelet Transform (SWT) and the Algebraic Multi-Grid (AMG) method, comparisons between segmentations from proposed method and eCognition exhibited that SWT- and AMG-based method produced integrated segments and accurate borders, particularly in the urban area. Nevertheless, speckle noise inherent in SAR is still a challenge to generate accurate segmentation (Li et al., 2010). Different land-cover type pixels may be merged into one segment when the differences in the scattering mechanism among these land cover types are smaller than thresholds used to merge adjacent pixels or objects. Meanwhile, one type of land cover parcel is easily divided into finer segments due to the strong spatial heterogeneity of internal SAR backscattering coefficients, which results in low-quality segmentation. The over-segmentation or under-segmentation at the first stage of object-oriented classification will degrade the performance for subsequent crop identification.

Fortunately, the impact of speckle on the SAR segmentation can be solved by object-oriented optical-segmentation. The optical remote sensing image, especially for high-resolution image, is capable of delineating meaningful land-cover objects effectively. A novel method presented in this paper attempts to map the maize distribution through incorporating multi-temporal SAR images with high-resolution optical imagery. RADARSAT-2 provides the backscatter features to describe the maize growth characteristics across the maize growing season and the high-resolution optical imagery used for segmentation provides the basic processing units for mapping maize. We called it as an optical image supported parcel-based SAR classification, briefly abbreviated as os-PSC. Furthermore, a parcel- and pixel-level integrated classification scheme was developed for many agricultural systems with small landownership where inter-cropping is common. We carried out an experiment in Shenzhou District, Hengshui in Northern China, a representative maize planting region, to test the robustness and applicability of the proposed method.

Section snippets

Study area

Shenzhou, located in Hengshui, Northern China, was selected as study area (central longitude: 115°33′12′'E, latitude: 38°0′3′'N), as illustrated in Fig. 1. This area is characterized by predominantly flat terrain, fertile soil, sufficient solar illumination and precipitation, thus providing optimal conditions for crop growth. The major land-cover is fragmented cultivated land which intermixed with trees, built-up and water body, forming a complex mosaic of typical agricultural landscapes. Maize

Methodology

The os-PSC procedure was listed in Fig. 4, mainly including polarimetric features extraction from PolSAR images, optical image parcels segmentation which was used to calculate mean polarimetric and NDVI value within parcel, potential maize mask determination calculated from NDVI threshold, parcel-level SVM classification and a key step of pure and mixed parcel separation based on entropy value and final maize map production by mosaicking the pure-parcel and pixel-based maize parts.

Accuracy assessment for maize

In order to test the performance of the os-PSC method for mapping maize, a comparison between os-PSC and two other methods, pixel-level (Pi_C) and PolSAR-based segmentation classifications (Pa_C), was conducted. Training samples were selected for each method separately due to various mapping units, however, we reduced to a large extent the impact of training samples on classification results by only choosing pure land-cover parcels or pixels within user-defined areas as training samples. Pi_C,

Conclusions

We proposed in this paper an innovative approach which integrates the superiority of PolSAR data for capturing temporal development of maize and the strengths of high-resolution optical images for delineating the land surface boundaries as the basic processing units. A generic framework of multi-step classification is an effective procedure to progressively eliminate the influence of non-maize classes. The comparison between the proposed method and both pixel-based (Pi_C) or SAR segmentation

Acknowledgement

This research is supported by ‘Major project of high resolution earth observation system (civil part)(09- Y20A05-9001-17/18)’.

References (61)

  • F. Löw et al.

    Impact of feature selection on the accuracy and spatial uncertainty of per-field crop classification using support vector machines

    ISPRS J. Photogramm. Remote. Sens.

    (2013)
  • S.K. Maxwell et al.

    An automated approach to mapping maize from landsat imagery

    Comput. Electron. Agric.

    (2004)
  • S.K. Maxwell et al.

    A method for mapping maize using the US geological survey 1992 national land cover dataset

    Comput. Electron. Agric.

    (2006)
  • A. Mitchell et al.

    C- and L-band SAR interoperability: filling the gaps in continuous forest cover mapping in Tasmania

    Remote Sens. Environ.

    (2014)
  • A. Potgieter et al.

    Early-season crop area estimates for winter crops in NE Australia using MODIS satellite imagery

    ISPRS J. Photogramm. Remote. Sens.

    (2010)
  • Z. Qi et al.

    A novel algorithm for land use and land cover classification using RADARSAT-2 polarimetric SAR data

    Remote Sens. Environ.

    (2012)
  • C.J. Tucker

    Red and photographic infrared linear combinations for monitoring vegetation

    Remote Sens. Environ.

    (1979)
  • Y. Zhang et al.

    Improving the impervious surface estimation with combined use of optical and SAR remote sensing images

    Remote Sens. Environ.

    (2014)
  • Z.D.U. Avci et al.

    Process-based image analysis for agricultural mapping: a case study in Turkgeldi region, Turkey

    Adv. Space Res.

    (2015)
  • Y. Ban

    Synergy of multitemporal ERS-1 SAR and landsat TM data for classification of agricultural crops

    Can. J. Remote. Sens.

    (2014)
  • Y. Ban et al.

    Object-based fusion of multitemporal multiangle ENVISAT ASAR and HJ-1B multispectral data for urban land-cover mapping

    IEEE Trans. Geosci. Remote. Sens.

    (2013)
  • Y. Bazi et al.

    Toward an optimal SVM classification system for hyperspectral remote sensing images

    IEEE Trans. Geosci. Remote. Sens.

    (2006)
  • C. Boryan et al.

    Monitoring US agriculture: the US department of agriculture, national agricultural statistics service, cropland data layer program

    Geocarto Int.

    (2011)
  • E. Box et al.

    Accuracy of the AVHRR vegetation index as a predictor of biomass, primary productivity and net CO2 flux

    Vegetatio

    (1989)
  • L. Breiman

    Random forests

    Mach. Learn.

    (2001)
  • S.R. Cloude et al.

    A review of target decomposition theorems in radar polarimetry

    IEEE Trans. Geosci. Remote. Sens.

    (1996)
  • S.R. Cloude et al.

    An entropy based classification scheme for land applications of polarimetric SAR

    IEEE Trans. Geosci. Remote. Sens.

    (1997)
  • R.G. Congalton et al.

    Assessing the Accuracy of Remotely Sensed Data: Principles and Practices

    (2008)
  • Cortes, C., and V.N. Vapnik “Support-vector networks.” Machine Learning 20(3) 273–297. doi:...
  • A.M. Dean et al.

    An evaluation of per-parcel land cover mapping using maximum likelihood class probabilities

    Int. J. Remote Sens.

    (2003)
  • Cited by (0)

    View full text