Detection of trend and seasonal changes in non-stationary remote sensing data: Case study of Tunisia vegetation dynamics

Highlights

• Comparison between four decomposition methods for non-stationary time series analysis.

• Modelling and monitoring NDVI time-series changes due to inter- and intra-annual variations.

• Vegetation change analysis in northwest of Tunisia.

Abstract

The availability of long-term time series (TS) derived from remote sensing (RS) images is favorable for the analysis of vegetation variation and dynamics. However, the choice of appropriate methods is a challenging task. This article presented an experimental comparison of four methods widely used for the detection of long-term trend and seasonal changes of TS, with a case study in north-western Tunisia. The four methods are the Ensemble Empirical Mode Decomposition (EEMD), Multi-Resolution Analysis-Wavelet transform (MRA-WT), Breaks for Additive Season and Trend (BFAST), and Detecting Breakpoints and Estimating Segments in Trend (DBEST). Their efficiencies were compared by analysing Normalized Difference Vegetation Index (NDVI) TS from 2001 to 2017 in the study area, obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) observations. The variations of long-term NDVI trends were analysed using non-parametric statistical tests. Results indicated that MRA-WT gave efficient results for both trend and seasonal changes, especially in forest area. Moreover, it exhibited the fastest efficiency in terms of time of execution and thus recommended for detecting detailed features (such as forest fire detection). DBEST also showed a good performance for trend detection in forest area as MRA-WT, however, it was more constrained to a longer computational time of execution. BFAST and EEMD exhibited a better performance in bare soil and cropland areas, and the latter can be taken as an appropriate and fast alternative for a general long-term trend overview with long TS.

Introduction

Vegetation is one of the most important components of the world's ecosystems. It plays a key role in global climate and in regulation of various biochemical cycles such as water and carbon (Ferchichi et al., 2022; Zhong et al., 2010). Vegetation dynamics have been investigated widely regarding land-use/land-cover (LULC), grassland change, deforestation, land degradation, forest fire, and crop monitoring (Lu et al., 2004). Over the last decades, remote sensing (RS) data have been widely used to monitor vegetation dynamics (Dubovyk et al., 2015; Martnez and Gilabert, 2009) due to the availability of key vegetation-related variables over a long temporal period and large spatial scales.

The Normalized Difference Vegetation Index (NDVI) is a proven index of satellite imagery and has been widely used as a substitute for biomass, cover and vitality of vegetation (Coppin et al., 2004; de Jong et al., 2012; Essid et al., 2012; Fensholt et al., 2006; Forkel et al., 2013; Huang et al., 2018; Xue et al., 2014). It is favorable and advisable to investigate various ecosystem changes from the NDVI time series. For example, annual mean or peak NDVI values provide an integrated view on photosynthetic activity; seasonal NDVI amplitude is related to the composition of evergreen and deciduous vegetation; and the length of NDVI growing season can be related to phenological changes (Forkel et al., 2013). Generally, these studies on vegetation changes mainly concern linear trends of NDVI time series (TS). For a long TS, linear fitting could completely mask internal trend shifts and hinder the identification of underlying causes (Wei et al., 2018). To overcome the problem, more effective algorithms have been proposed to detect the time-varying trends of vegetation growth (e.g. piecewise linear regression, polynomial fitting, etc.). These algorithms involve setting a prior-function form and defining key parameters subjectively to describe the trend with implicit stationary assumptions. NDVI TS are usually non-stationary, which may not only be limited to the mean of the series but also affect its overall structure of variance (Ben Abbes et al., 2018; Martnez and Gilabert, 2009; Verbesselt et al., 2010a, Verbesselt et al., 2010b). It makes the analysis of vegetation changes more complicated compared to a monotonic trend at large timescales (Cohn and Lins, 2005). Generally, NDVI TS present seasonal, gradual, and abrupt changes. Gradual changes refer to long-term small magnitude date-to-date changes. Trend detection in NDVI TS can help to recognize recent changes from local to global scales in ecosystems. However, seasonal changes are related to plant cycle and short-term changes.

The development of efficient methods for handling RS data with non-stationary characteristics at multi-temporal scales is a difficult issue (Bruzzone et al., 2003; Martnez and Gilabert, 2009). Some of the most commonly used methods are summarized in Fig. 1. They can be divided into two categories: time-frequency methods, and statistical methods (Ben Abbes et al., 2018). Various time-frequency methods have been proposed and applied in RS field. These include fast Fourier transform (FFT) Azzali and Menenti (2000), Hilbert-Huang transform (Huang et al., 2001), wavelet transform (WT) (Martnez and Gilabert, 2009), ensemble empirical mode decomposition (EEMD) (Kong et al., 2015), and others. The FFT decomposes TS using a superposition of sine and cosine functions, being only localized in frequency. The WT decomposes TS in both time and frequency domains, and has thus been more commonly used in different fields compared with FFT (Rhif et al., 2019). EEMD is based on local characteristics and decomposes TS into a collection of intrinsic mode functions (IMF), as a self-adaptive signal processing system.

Several statistical approaches were developed in the literature such as season trend loess (STL) (Jacquin et al., 2010), break for additive season and trend (BFAST) (Verbesselt et al., 2010a), detecting breakpoints and estimating segments in trend (DBEST) (Jamali et al., 2015), Singular spectrum analysis (Mahecha et al., 2010), sub-annual change detection (Cai and Liu, 2015), seasonal auto-regressive integrated moving average (Jiang et al., 2010), and Landsat-based detection of trends in disturbances and recovery (Kennedy et al., 2010). These approaches usually decompose TS into different variations (e.g., seasonal variations, gradual trends and sudden shifts) at multi-timescales. Both BFAST and DBEST were proposed exclusively for RS datasets. BFAST uses an additive algorithm to account for seasonality and detect gradual (inter-annual) and abrupt (intra-annual) changes within the trend component. However, a segmentation algorithm was used by DBEST to simplify the trend into linear segments.

Many studies investigated the different performances of these methods for RS data analysis. For example, De Oliveira et al. (2009) compared FFT and WT for vegetation classification by means of artificial neural networks (ANN) and verified the better efficiency of the latter. Wang et al. (2016) used FFT, WT, and EEMD for rice phenology extraction using NDVI data in Jiangsu Province (China). They concluded that WT provided the best estimation for phenology analysis, followed by EEMD and FFT. A comparative study between STL, BFAST, and WT using NDVI TS were studied by Ben Abbes et al. (2018). They concluded that BFAST provided the most accurate results. For the multi-level decomposition, MRA-WT was reported to be more insightful. Jamali et al. (2015) reported that DBEST performed better in time compilation compared to BFAST. Cai and Liu (2015) compared the sub-annual change detection (SCD) method and BFAST, proving that both methods can detect NDVI change with equal precision, but the former is more computationally efficient.

Based on previous studies, EEMD, WT, BFAST, and DBEST proved their effectiveness for NDVI TS analysis. For example, Kong et al. (2015) decomposed the NDVI TS based on EEMD to detect the trend, seasonal and abrupt changes. They concluded that EEMD is fast, easy for computation and robust for trend and seasonal analysis. In fact, both components are not distorted by transient, aberrant behaviours. Also, Wei et al. (2018) investigated vegetation trends in East Africa using EEMD, which presented a new insight into the nonlinear dynamics of vegetation growth and characterized the spatiotemporal dynamics of NDVI trends, including instantaneous trend, accumulated trend, and turning points. Recently, EEMD was used for the detection of monotonic trend of vegetation in China (Tang and Yang, 2020).

Regarding vegetation dynamics, Martnez and Gilabert (2009); Martnez et al. (2011) applied MRA-WT to explore vegetation dynamics in Spain and Senegal by decomposing the TS into different temporal scales. Campos and Di Bella (2012) used MRA-WT to analyse land cover changes in different areas and to assess their impacts, such as seasonal inundation in Cambodia and Vietnam, fire damage in Spain, farming in Argentina and urbanization in China. Recently, the effectiveness of WT for non-stationary TS analysis was reviewed in (Rhif et al., 2019). Recently, Nourani et al. (2021) investigated the WT to analyse land cover changes using the hydro-climatic, SST and NDVI time series.

BFAST proves its effectiveness to account for gradual (inter-annual), seasonal and abrupt changes (de Jong et al., 2012; Verbesselt et al., 2010b). Verbesselt et al. (2012) used BFAST to detect drought-related vegetation disturbances in Somalia. Schucknecht et al. (2013) investgated vegetation variability and trends in north-eastern Brazil using AVHRR and MODIS NDVI TS. Zewdie et al. (2017) identified the driving factors of dryland changes using MODIS NDVI, precipitation and temperature data in Ethiopia. They used BFAST to iteratively detect the number of changes and their timings. Recently, Fang et al. (2018) described large-scale vegetation dynamics in Quebec (Canada) using MODIS images by BFAST.

DBEST was developed by Jamali et al. (2015) to detect and characterize changes of NDVI over large areas. It is a flexible method that allows the control of fine details or general features of the trend to be captured. Recently, Shen et al. (2018) detected the vegetation changes in China using DBEST and MODIS NDVI dataset.

Although EEMD, WT, BFAST, and DBEST have been widely used, the selection of the most effective method for the detection of long-term trend and seasonal changes in non-stationary NDVI TS is still missing and has not been justified. To our knowledge, there still lacks a comparison of the four methods based on regional and local RS analysis. The objective of this paper is to explore four decomposition methods in order to compare their features and, particularly, their performance in modelling and monitoring NDVI time-series changes due to inter- and intra-annual variations. For this purpose, the MODIS NDVI time series at 250 m is assessed for the period 2001–2017 in northwest Tunisia. The inter-annual trend and seasonal variations are analysed at regional and local scale for each method. Finally, results using non-parametric tests are analysed.

To this end, the study area and the data used are introduced, followed by the descriptions of the four methods and the statistical tests used to detect the vegetation changes in the study area; results are discussed and concluded finally.