1. Introduction
Euphrates or desert poplars (
Populus euphratica) are medium-sized deciduous trees mainly found in fluvial and floodplain areas in arid and semi-arid regions, where the groundwater is close to the surface. Their unique physiological traits make them tolerant of saline and brackish water and resistant to droughts and sand storms. They form the near top-level natural arbor community in the desert environment and are a significant component of Tugai floodplain ecosystems in arid climates. Desert poplars play an important role in maintaining ecological balance, establishing windbreaks for the protection from wind and the encroachment of sands, fixing shifting sands, regulating the climate and improving the health of ecosystems [
1,
2]. However, poplar plantations in China have been constantly threatened by defoliating insects, particularly the poplar looper
Apocheima cinerarius Erschoff (Lepidoptera: Geometridae). In recent years, there are widespread outbreaks of insect pests in desert poplar plantations in northwest China due to the weakened resistance of water-stressed poplars to pests and diseases resulted from excessive unregulated water diversion and climate change. Poplar looper is the most serious insect pest affecting desert poplars in the Tarim Basin, ** strategies to protect desert poplars.
Poplar looper is a geometrid moth, distributed in the arid climate zone in the East Asia monsoon region, including the entire area of northern China [
5]. A lot of studies have been done on this moth species, including its habitat, physiology, life cycle, evolutionary history, geographical distribution, population dynamics, the nature of its attack and damage it caused, as well as forecasts and control of its outbreaks [
6,
7,
8,
9,
10,
11,
12]. However, Euphrates are distributed in deserts, which are usually remote and poorly accessible by transport. It is difficult to carry out large-scale and in-depth studies on the pest infestation using traditional field survey techniques. So far, there has been a lack of information on the extent and damage level of the poplar looper infestation in Euphrates forests in China.
Remote sensing is a proven technology for pest monitoring and damage assessment over large areas. Many studies have used remotely sensed data with various spatial resolutions to detect and assess forest insect infestations, based on the analysis of spectral responses of trees to the biophysical phenomena such as water stress caused by insect attack [
13,
14,
15,
16,
17,
18,
19]. Over the years, a number of spectral indices (such as the Normalised Difference Vegetation Index or NDVI, the Normalised Burn Ratio or NBR, the Moisture Stress Index or MSI, the Enhanced Wetness Difference Index or EWDI, and the Leave Area Index or LAI) have been used for the detection and differentiation of insects and diseases using multispectral and hyperspectral data with different spatial resolutions. According to a recent review by Senf et al. [
20], medium and coarse spatial resolution remote sensing data and NDVI were most often used for map** insect induced broadleaved defoliation, and intra-annual time series analysis was frequently applied in the analysis of these data sets. As many insects are active for a short period when an outbreak is detectable, most studies used very dense time series of remote sensing data such as the Advanced Very High Resolution Radiometer (AVHRR), the Moderate Resolution Imaging Spectroradiometer (MODIS), and SPOT VEGETATION to achieve daily temporal resolution. MODIS data have both high spectral and temporal resolutions, and have been more frequently employed [
21]. For example, they have been used to detect forest defoliation induced by the European gypsy moth in North America [
22], the autumnal moth and winter moth in northern Sweden [
23], the European pine sawfly in southeastern Norway [
24], and by the bark beetle in North America [
25]. These studies achieved various levels of success. For example, Olsson et al. [
26] reported that with MODIS data only 50% of damaged Scots pine forest stands caused by pine sawfly in eastern Finland were detected and 22% of healthy stands were misclassified, while in northern Sweden 75% of damaged birch forest stands were detected and 19% of healthy birch forest stands were misclassified. However, these applications have shown potential of MODIS data for large area monitoring and early detection [
27]. Our research used MODIS NDVI data to detect defoliation of poplar trees induced by poplar looper and assess the severity of the insect outbreak.
In recent years, for large-scale pest infestation monitoring, Landsat and Sentinel-2 data with a higher spatial resolution have been widely used [
28,
29,
30,
31]. Since poplar forests are most vulnerable to poplar looper attacks in less than two months [
32], it requires monitoring data to be updated daily to quickly monitor the status of poplar forests damaged by the insect. Although the spatial resolution of MODIS data is coarser than Landsat and sentinel-2, its long timespan and high temporal resolution are more suitable for this monitoring task. With the temporal resolution of 16 days, Landsat images could miss the foliation peak starting of defoliation, reducing the information about the pest infestation. Though Sentinel-2 data have higher spatial and temporal resolutions, hence a higher monitoring accuracy than Landsat 8 [
33,
34], the data with a 5-day temporal resolution only started to become available in 2017, which limits their use in the analysis and modelling of the long-term trend. Although the spatial resolution of MODIS data is coarser than Landsat and Sentinel-2, its long timespan and high temporal resolution are more suitable for the monitoring of poplar looper attacks on poplar forests.
Time series fitting techniques have been used to analyse MODIS time series data to detect insect disturbance [
20]. For example, TIMESAT, a computer program for time series analysis of satellite sensor data, fits smoothed functions (including Savitzky-Golay fitering, least-squares fitted asymmetric Gaussian functions and double logistic functions) to time series data and extracts seasonality parameters [
35], which was applied in the studies by Eklundh et al. [
24], Anees and Aryal [
25] and Olsson et al. [
23,
26]. Jia [
36], Huang [
37] and Wang [
38] used the S-G method to filter the MODIS NDVI time series and combined it with the data on the poplar phenology and poplar looper life history to extract the poplar looper hazard information in a poplar forest area using a change detection method. The wavelet transform is a signal processing tool for analysing both the local frequency and temporal behaviour of signals by decomposing a signal into frequency components and then studying each component with a resolution matched to its scale [
39]. Wavelet analysis has the merit of decomposing the signal into components of different scales (or frequencies) thus facilitating the detection of subtle signals. An NDVI time series can be considered as a signal. The wavelet transform was used in this study to decompose the NDVI time series into components of different scales, facilitating the detection of trends and subtle signals. Discriminant analysis is a statistical technique for analysing a data set with individuals classified into particular groups, and then using the results to classify new individuals that are not included in the above data set [
40].
In this paper, we present an innovative and effective method for detecting and monitoring the poplar looper infestation of dense desert poplar forests using MODIS time series data through the wavelet transform and discriminant analysis. The objectives of this study can be formulated as the following:
- (1)
Proposing a method of decomposition of mixed signals from the NDVI time series to isolate and enhance the pest infestation signals; and
- (2)
Building a predictive model for identifying severity levels and dates of pest outbreaks.
4. Discussion
4.1. Poplar Looper Pest Monitoring
Remote sensing allows non-contact and spatially continuous monitoring of forest pest infestation [
66]. In this paper, we presented an approach to the detection and severity classification of the infestation of poplar looper on desert poplars using MODIS NDVI time series through the wavelet transform and discriminant analysis. Gärtner et al. [
67] did a similar research. They combined RapidEye high-resolution and Landsat 8 medium resolution data to produce synthetic images using the enhanced spatial and temporal adaptive reflectance fusion model (ESRARFM) [
67] to detect poplar forest disturbance caused by poplar looper in a riparian Tugai forest at the Arghan forest station in **njiang. Their purpose was to assess whether the synthetic images could achieve a better disturbance detection accuracy than that achieved by using RapidEye data alone. They used the relative NDVI difference to measure forest defoliation and recovery, and classify the severity of defoliation based on the ratio of defoliation and recovery [
68]. Eight RapidEye images (5 m resolution) and ten Landsat 8 images were acquired in the 2003 growing season from the end of March to mid-September with large time windows. One limitation with the study is that the acquisition dates for the high resolution images could miss the foliation peak starting of defoliation, reducing the information about the disturbance. The second limitation is that the relative NDVI differences might not capture defoliation caused by insects. The third limitation is the high cost of high resolution image acquisition, which would prevent the use of the method for real-time monitoring of forest insect infestation and for studying insect dynamics over a long term. MODIS data can overcome these limitations and allow for analysing long time series to capture the natural fluctuations that are inherent to insect dynamics [
21,
22,
23,
25,
26]. Huang [
32], Wang [
38] and Qiu [
69] used MOD13Q1 and satellite HJ A/B data to study poplar looper infestation. The main limitation of their research is that they only studied whether an infestation had occurred using very short time series data and did not assess the severity of pest infestation and detect the time of its occurrence. Jia [
36] and Huang [
37] used the same method as that used in Huang [
32] to obtain poplar looper infestation information in different years in our study area, simulated the spatial spread of insect infestation, and identified the factors affecting the poplar looper outbreaks. Liu [
70] used the MODLT1D temperature product to study the response of the poplar looper in the desert poplar forests to surface temperature in this area. These studies including ours analysed and modelled the infestation of poplar looper in desert poplar forests using remote sensing data from different perspectives, and provided new tools for future large-scale remote sensing monitoring of poplar looper infestation.
4.2. Identification of Outbreaks of Pests
Remote sensing images can capture the changes in vegetation canopy caused by pest infestation [
71]. In this study, we applied the wavelet transform to filter the noise and graft the original NDVI values from March to April. The D1 detail component of an NDVI time series obtained after decomposition and reconstruction can highlight the sudden changes in NDVI values. As shown in
Figure 9, when the NDVI time series curve suddenly increases or decreases, the D1 curve also fluctuates strongly. Because the poplar looper infestation starts in late March and ends in middle or late April in the study area, the maximum value on the D1 curve from March to April each year can be used to determine whether the poplar looper infestation occurred. For example, substituting the maximum values of D1 in March and April of 2009, 2010 and 2012 respectively (reference lines A, B, C in
Figure 9) into the Equations (6)–(9) to determine the severity level of pest infestation at Sampling Site 4 (see
Figure 1), the results show that poplar looper infestation occurred at this site in early March 2009, 2010 and late April 2012, which correspond to our field observations.
The outbreak time determined by the method and the outbreak time observed through field surveys match with an accuracy of 94.37%, which shows that the methodology is able to reliably determine the outbreak time.
4.3. Wavelet Analysis
Wavelet analysis has been applied in the field of radio [
72] (such as noise reduction for acoustic or video signals), meteorological [
73,
74,
75] and hydrological information analysis [
76,
77]., image enhancement [
78] and image fusion [
79]. Hyperspectral images have a very high spectral resolution with hundreds of bands forming an approximately continuous spectral curve [
80]. Some scholars have tried to use continuous wavelet analysis for monitoring wheat yellow rust and powdery mildew for winter wheat based on hyperspectral images [
81,
82,
83]. They expanded the spectral data of each pixel in a single hyperspectral image to form a spectral curve, which is similar to an acoustic audio signal, and then applied wavelet analysis to detect or extract specific information about the outbreaks of the fungal diseases in the crops.
In this study, we modelled the numerical curve of each image element of the long time series of NDVI images as a special signal curve based on the characteristics of the periodic and regular changes of the vegetation canopy NDVI, which contains information about the seasonality and trends of the poplar tree growth, as well as localized abrupt changes caused by disturbance events such as droughts, fires, pests, diseases, and leave and timber harvesting or caused by clouds and snows; decomposed NDVI time series data over multiple levels via the wavelet transform; and constructed a smoothed NDVI time series to represent the poplar tree growth under normal conditions. We then added the March and April NDVI data to the smoothed NDVI time series to extract the high-frequency components. The level 1 detail components of the blended time series provide the best indicators of the pest infestation. The result shows that we have successfully detected the pest information. This is a new attempt to apply the wavelet analysis method.
4.4. Limitations
Although the research focused on the poplar looper infestation on desert poplars, the proposed approach can be used to detect and assess forest insect disturbances in general. However, our approach has its limitations.
As we blended the smoothed NDVI time series with the subset of the original time series containing the pest infestation signals, other noises such as those created by clouds and dust storms in the same period remained in the blended time series, which may reduce the detection and severity classification accuracy. It is necessary to identify and remove those noises not caused by pest induced forest disturbances by comparing the original time series and the detail components of the blended time series.
In addition, MODIS NDVI data have a moderate temporal resolution of 16 days and a coarse spatial resolution of 250 m. The presented approach works well in dense poplar forests. But both temporal and spatial resolutions may not be sufficient for accurately determine when and where pest infestations occurred in the areas where poplars are sparsely distributed. Further research will be conducted to fuse MODIS NDVI time series data with other remote sensing data with higher temporal and spatial resolutions to improve the accuracy.
Poplar looper has one generation a year in the Tarim Basin. Its outbreak occurs in March and April. Our approach can be applied directly to an area with the similar natural environment and pest infestation dynamics. Tests are required to verify whether it is applicable to other regions where pests may live for more than one generation, such as the areas south to the Yangtze River in China where pests may have three generations a year and the areas north to the Yangtze River where pests may have two generations a year.
Detection and monitoring of poplar looper infestations on desert poplars by means of remote sensing is still in the stage of exploration and development. Further studies will be conducted to improve the proposed approach with the aim to more accurately detect the poplar looper infestation signals from remote sensing data and enable early detections of poplar forest pest-induced disturbance, which can provide timely information and early warning for effective poplar forest pest prevention and control. Without the disturbance of pests, the growth of desert poplars will be getting better, thereby improving the ecological conditions of its habitat in the Taklimakan Desert.