1. Introduction
As an important characteristic of forest ecosystems, forest aboveground biomass (AGB) provides a basis for ecosystem and forestry research; AGB estimation further provides data critical to estimating the forest carbon sink [
1,
2]. In recent years, accurate and rapid AGB estimation has, therefore, become crucial for forest ecosystem and global climate change research.
Traditionally, high precision AGB field measurement methodologies have involved extensive field surveys [
3]. However, these methods are time-consuming, laborious and destructive; in addition, they cannot be used to analyze the spatial distribution and dynamic change of AGB on a large scale [
4]. Today, remote sensing-based methodologies are more commonly used to estimate AGB as they rapidly provide near real-time, dynamic and regional scale data, and the characteristics of the obtained images are strongly correlated with AGB [
5]. Remote sensing data can be divided into two categories: Passive remote sensing (optical sensors, thermal and microwave) and active remote sensing (radar and light detection and ranging (LiDAR)) [
5,
6,
7]. Optical sensors such as Landsat, Systeme Probatoire d’Observation de la Terre (SPOT), the moderate-resolution imaging spectroradiometer (MODIS), QuickBird and the Advanced Very High-Resolution Radiometer (AVHRR) have been widely used for AGB estimation because of their coverage, repetitive observation and cost-effectiveness [
6,
8]. Of these sensors, Landsat images are the most commonly used for remote sensing-based AGB estimations because the sensors provide a long-term data record and have medium spatial resolution, wide spatial coverage and high spectral sensitivity [
9]. In many countries, the spatial resolution obtained using Landsat is similar to the size of sample plots in national forest inventories; therefore, using Landsat to estimate AGB can reduce spatial errors associated with matching pixels to sample plots [
10].
The information derived from Landsat images significantly correlates with AGB because these images provide valuable information regarding the forest canopy [
11]. In fact, previous studies have shown that individual spectral bands, vegetation indices, transformed images (using principal component analysis (PCA)) and textural images are strongly correlated with AGB and can, therefore, be used to effectively estimate AGB [
12,
13,
14,
15]. Furthermore, many statistical models can be used in developing remote sensing-based AGB models. These models can be divided into two categories: (i) Parametric models (linear, nonlinear, etc.) [
16,
17,
18] and (ii) nonparametric models (random forest, RF; artificial neural networks, ANN; support vector machines, SVM; etc.) [
14,
19,
20,
21]. Multiple linear regression models, however, are most frequently used in AGB research.
Optical sensors mainly provide information about the forest canopy [
11]. The canopy structure of subtropical forests significantly varies between seasons, and even between months [
6,
22,
23]. These variations can cause differences in remote sensing data [
24]. Therefore, AGB estimation can vary widely when time-series images are used to model AGB in the same study area [
25]. Previous studies have used a single Landsat image (taken during the peak growing season or at a time close to when the ground survey of national forest inventory plots took place) to estimate AGB [
21,
26,
27,
28]. These images, however, do not always accurately reflect forest characteristics. For example, dense canopy cover during the peak growing season often results in extremely saturated images [
25,
29,
30], which ultimately affects AGB estimation accuracy. Some studies have, therefore, utilized time-series of Landsat images to estimate AGB, e.g., Zhu and Liu [
25], Safari et al. [
31] and Powell et al. [
32]. These studies, however, focused on particular forest type or a regional forest. Therefore, there exists a knowledge gap regarding whether time-series Landsat images affect the accuracy of AGB estimations in different forest types and whether the estimations differ among forest types.
Given this gap in knowledge, this study explores the use of seasonal Landsat 8 Operational Land Imager (OLI) images in estimating AGB in a subtropical forest in northern Hunan, China, using stepwise regression. The main objectives of this study were to: (1) Explore the potential variables of seasonal time-series data for different forest types when estimating AGB; (2) investigate the potential of seasonal time-series data in improving the accuracy of AGB estimations in different forest types; and (3) investigate the uncertainties associated with using seasonal time-series data to estimate AGB.
4. Discussion
Forests are complex ecosystems containing variable species composition and structure; therefore, the image information (especially textural information) of these ecosystems also varies considerably [
55,
56]. Previous studies utilizing Landsat images to estimate AGB were unable to determine which spectral variables were best able to predict AGB [
6,
57]. In this study, the selected spectral variables used for AGB models of different months and different forest types varied. Nonetheless, we found that for all forest types, textural images played an important role in AGB estimation, in accordance with previous research [
58]. The selected variables belonged to various original bands (bands 2 to 7), indicating that all original bands can be used to estimate AGB in this study. These results differed from earlier research in which the shortwave infrared (SWIR) bands (e.g., Landsat TM spectral bands 5 and 7) were more important in AGB estimation than other bands [
59,
60,
61]. In addition, in previous research utilizing Landsat imagery, spectral information (e.g., vegetation index, original band) was often selected to estimate the AGB of coniferous forest given that the structure of coniferous forest was simple and the importance of spectral information over textural information. On the other hand, textural information has often been used in the study of broadleaf forest and mixed forest given that those forests often have multiple canopy layers and more complex structures. In our study area, because of the low level of forest management, the forest structure was complex; therefore, in this study, for each forest type, textural information was mostly used to estimate AGB, regardless of which seasonal image was utilized.
In this study, stepwise regression was used to estimate AGB of different forest types based on Landsat 8 OLI seasonal images. We found that in our study area, the best month for AGB estimation was October given that the R
2 values of different forest types were higher than 0.39. Overall, this result indicates that the October image can explain more than 39% of the information regarding the estimated AGB for each forest type. The less accuracy month for AGB estimation was August given that the R
2 value for total vegetation was only 0.27. Stepwise regression is a widely used methodology of fitting regression models based on the correlation between dependent and independent variables. During this procedure, the significance of an introduced variable is tested, and the variable that is of least significant is discarded [
62]. While selection of variables depends upon the degree of linear correlation, selection of variables with low correlation is possible; this can ultimately affect the accuracy of the model. The forest characteristics of different forest types were heterogeneous. Different forest types were different in spectral characteristics caused by the heterogeneity of the stand structures and species compositions. The correlations among the spectral variables and AGB of different forest types were also different. In this case, the performances of models for different forest types were significantly different. In our study, among all forest types analyzed, we found that the MXF models achieved the best results for AGB estimation. This indicates that the image information was most strongly correlated with MXF compared with other forest types, and therefore, the image can better reflect the condition of the mixed forest. However, when the forest types were considered in AGB estimation, model accuracy was further affected by the number of plots [
59]. In this study, there were 135 CFF plots and 128 BLF plots, whereas there only 40 MXF plots. Therefore, MXF models may have been more accurate given the far fewer number of plots analyzed compared with the models for the other forest types.
In this study, Landsat 8 OLI seasonal images were used to estimate AGB. The four seasonal images utilized were associated with four seasons of the study area (January (winter), April (spring), August (summer) and October (autumn)). The results showed that utilization of the peak season (August) image resulted in inadequate AGB estimation compared with the other seasons, in accordance with results reported by Zhu and Liu (2014) [
25]. These researchers further found that the normalized difference vegetation index (NDVI)-based AGB estimates of the forest senescing period were better than those of the peak season in a temperate forest of southeastern Ohio, USA [
25]. Furthermore, in accordance with our results, previous researchers detected over- and underestimations when utilizing Landsat 8 OLI imagery to estimate AGB in a subtropical forest in western Hunan, China [
58]. In our study, these uncertainties were common among all seasonal images analyzed. The observed underestimations within the higher range of AGB values may have been a consequence of image saturation issues affecting model performance [
56,
63]. Regarding AGB values within the lower range, model performance was likely affected by mixed pixels, thus resulting in AGB overestimation [
64]. While uncertainties were detected among all time-series images, underestimation associated with the peak season (August) within the high AGB range (>75 Mg/ha) was more serious than that associated with the other seasons. Taken together, these results suggest that image saturation more strongly influenced AGB estimation results for August than it did for the other seasons, further indicating that the uncertainties were less in the other seasons. In addition, the overestimation associated with the peak season was greater than that associated with the other seasons.
5. Conclusions
In this study, seasonal Landsat 8 OLI imagery was utilized to estimate forest AGB in a subtropical forest in northern Hunan Province, China. Study plots were classified according to forest types (CFF, BLF, MXF and total vegetation) and stepwise regression was used to select appropriate variables and thus effectively model AGB based on the seasonal images. Subsequently, models of the different scenarios (different forest types in different seasons) were compared. Given the variables selected during stepwise regression, we concluded that seasonal image textural information was most significantly correlated with AGB, and that the study area is made up of forests with complex structures. The method of AGB estimation based on forest type is very useful for improving the accuracy of AGB estimation because the model performances for the different forest types (CFF, BLF and MXF) are better than those for the total vegetation, regardless of season. The time-series images, which reflect various seasons, can affect AGB estimations, with the autumn image (October) potentially yielding the most accurate AGB estimations and the peak season (August) image being of poorer quality in a subtropical forest. We also explored the accuracies of seasonal images in remote sensing-based AGB estimation. We hope to provide new insight into using Landsat images to improve the accuracy of biomass estimation.
Future research will focus on the mechanism underlying the cause of these differences when utilizing seasonal Landsat 8 OLI images in AGB estimation of different forest types.