Phenology Alignment-Based PolSAR Crop Classification Considering Polarimetric Statistical and Time-Varying Curve Characteristics

The uncertainty of crop phenological cycle is an important issue in crop classification with time series polarimetric synthetic aperture radar (PolSAR) data. The time series alignment algorithm represented by dynamic time warping (DTW) can supply a potential solution, which realigns curves based on shape matching, dealing with the distortion of feature curves caused by uncertain crop phenological development. However, previous studies mainly focused on shape characteristics of time-varying feature curves, which is hard to comprehensively evaluate the similarity degree of crop phenological cycles. Furthermore, it ignored the differences in scattering signal and polarimetric statistical distribution of crops, which limited the accuracy of crop classification. In this letter, a novel crop classification method based on phenology alignment is proposed. First, the dual-branch time series alignment method is proposed, including the time-weighted DTW (TWDTW) alignment and the Wishart distance-based TWDTW (WD-TWDTW) alignment, which combines the feature curve characteristics and the polarimetric statistical information to correctly describe the similarity degree of phenological cycles. Second, a multi-similarity measure (including shape similarity, feature similarity, and polarimetric similarity) is defined to improve capacity of crop discrimination. The multisimilarity measure can describe the differences of crop types from three aspects, including crop growth trend, growth status, and statistical distribution. The proposed method is evaluated with time series full-polarization Radarsat-2 data in Flevoland area. The results show that our method is superior to traditional method with single TWDTW alignment and shape similarity, and the corresponding overall accuracy (OA) is improved by 6%.

I. Introduction

Polarimetric synthetic aperture radar (PolSAR) sensors can provide a wealth of continuous time series observations, which can reflect time series change of crop growth. It is essential for growth monitoring, yield prediction, and agricultural management [1]. Furthermore, time series PolSAR data have been widely applied in crop classification to avoid the low separation degree of crop types on single date [2]. Nowadays, there are three main kinds of crop classification methods as follows. 1)

The methods based on the simple stack of feature vector, which treat each temporal data as independent feature, and discriminate crop types using different machine learning classifiers [2], [3].

2)

The methods based on temporal context between observations, such as hidden Markov model (HMM) [4] and recurrent neural network (RNN) [5]. These algorithms specifically model the temporal relationship of time series data to detect specific crop growth patterns.

3)

The methods based on time series feature curve alignment, which compare feature curves, and realign temporal sequences to generate time-varying similarities [6], [7], [8]. The classification label is designated to the class with the highest similarity.