Niantang Liu

  • PhD candidate 


Room 304, East Quadrangle



Research title: A Deep Convolutional Neural Network-based framework for Crop Classification on High-Resolution Radar & Multispectral Imagery

Research Summary

Research title

 Enhanced Crop Classification through Electrodynamical Modelling of Vegetation Combined with Multi-temporal multi-mode spaceborne Synthetic Aperture Radar (SAR) Data.

Summary of Research

Croplands possess valuable information that can be detected, mapped and monitored over short to long periods using remote sensing technology, with great potential for enhancing sustainable agricultural production and assisting decision-making related to regional planning. Crop identification, however, is regarded as a fundamental step for most agricultural monitoring systems, which leads to relatively accurate estimation of the area allocated to each crop type and then derives relevant statistics for crop management involved with area-based subsidies, on account of crop type being able to be identified (Blaes et al, 2005). Additionally, crop yields can also be predicted through relative forecasting models to which corresponding crop type information is discriminated firstly and then imported as input variable.

Fully polarimetric SAR (PolSAR) systems measure all of orthogonal polarization forms (HH, HV, VH and VV) which provides more insight into the variation of scattering mechanisms between retrieved backscatter values and the surface being imaged. Thus, the improvement of crop classification accuracies has been demonstrated with reference to the analysis of parameters derived by polarimetric decomposition with PolSAR data (McNairn et al, 2009b). Zhang et al (2008) compared two polarimetric decomposition methods including coherent and incoherent approaches using PolSAR data for target classification. However, moisture content also serves as a significant determinant to influence the efficiency of crop identification as it can be highly variable (Saich and Borgeaud, 2000).

Apart from polarimetric decomposition, another way to recognise crop types involves the prediction of backscattering coefficient from vegetation medium based on inherent scattering mechanisms which can be potentially modelled. Many models have been developed in the past decades, such as the model based on Foldy–Lax multiple-scattering equations (Le Toan, 1997), radiative transfer equations (Attema and Ulaby, 1978; Le Toan et al, 1989), the matrix doubling method (Picard et al, 2003; Blaes et al, 2006), the distorted Born approximation, and the Monte Carlo coherent scattering model (Wang et al, 2005). Kozlov et al (2007) presented a pertinent review of existing electrodynamical models of vegetation. These models are mainly concerned with forests and crops including soybean, corn, and wheat. However, those originally proposed vegetation models may encounter some limitations when considering the limited selection of biometrical characteristics of vegetation as input variables to modelling and overlooking complex electromagnetic interactions from modelling the scattering process within an internal vegetation structure of single or double layer, while not taking into account phenological changes of crops, although some statistical models are capable of incorporating temporal dependencies caused by the crop phenology to improve the classification accuracy (Leite et al, 2011; Siachalou et al, 2015; Kenduiywo et al, 2015, 2016).