Large Scale Studies of Forest Dynamics

1.2.1. Forest Permanent Plots

Forest dynamics have been extensively studied through ground surveys of permanent plots (Forest Dynamics Plot, FDP) that constitute several research networks of different focuses and geographical ranges, such as RAINFOR in tropical-equatorial

4

South America and the Smithsonian’s ForestGEO across different forest biomes (Malhi et al., 2002; Anderson-Teixeira et al., 2015). Data acquired from the FDPs have permitted numerous advancements in forest ecology either locally, or at the regional and world scales by using observations from several plots (Lewis et al., 2004; John et al., 2007; Laurance et al., 2009; Cleveland et al., 2011; Wright et al., 2011; Chisholm et al., 2014; Yu et al., 2019). Multiple surveys allow researchers to draw temporal trends. For instance, results from decades of plot censuses demonstrated the changes in tropical forest dynamics (see review of Lewis et al., 2009a; Phillips & Lewis, 2014). The FDPs are ideal cases to observe the effects of unpredictable events, and large plots are practical to study infrequent disturbances, such as forest fires (Lutz et al., 2018) or snow (Song et al., 2017), that could be missed by small plots. Several plots of the ForestGEO network are located in regions affected by cyclones (Figure 1.2), and as such, they permitted to identify effects of cyclone disturbances on forest characteristics such as tree mortality and forest resistance (Hogan et al., 2018; Johnson et al., 2018).

To upscale results derived from the FDPs, the representativeness of the FDPs need to be validated. Many scientists expressed concerns regarding location of FDPs within their landscapes (e.g., Körner, 2009; Lewis et al., 2009a; Marvin et al., 2014), as biased location may invalidate generalization of their results at broader scales as discussed by Saatchi et al. (2015) regarding biomass (see also Chave et al., 2004), and by Clark (2002) about carbon sinks. Indeed, plot size, shape, and location has often been carefully chosen based on different criteria such as soil homogeneity (Ashton, 1995;

Phillips et al., 2003; Magnusson et al., 2005), forest age (‘majestic forest’), or accessibility (Phillips et al., 2003; Ostertag et al., 2005), often in attempt to minimize possible biases.

Among networks, ForestGEO and RAINFOR plots have followed guidelines to improve their representation of surrounding forests, which include plot size to reach a good representation of tree demographics.

5

Figure 1.2. Locations of Forest Dynamics Plots from the Smithsonian ForestGEO network (Anderson-Teixeira et al., 2015) in three regions with frequent cyclones. Tree density from Crowther et al. (2015). Cyclone strength is on the Saffir-Simpson scale, data from the NOAA IBTrACS archive (Knapp et al., 2010; Knapp et al., 2018).

Different methods have been used to assess and improve plot representativeness.

First, suspect plots can be removed from the analysis (Lewis et al., 2004; Phillips et al., 2004; Lewis et al., 2009b). Second, orientation with topographical gradient allows to increase representation of the environment variability (Tuomisto et al., 2003), which may be necessary when species diversity change with the terrain (e.g., Valencia et al., 2004). Third, under-sampled sites can be remotely sensed, as areas with different reflectance likely have different vegetation types (e.g., species, Rocchini et al., 2015).

Langford et al. (2016) used several images from different phenological stages and Euclidean distances based clustering to assess the representativeness of their plots in a tundra landscape, a method which was first developed to study regional networks representativeness (Hargrove et al., 2003; Hoffman et al., 2013). Di Vittorio et al. (2014) used remote sensing to show that existing FDPs in the Amazon forest did not represent

6

tree mortality, as mortality may be spatially clustered (e.g., blowdown). Particularly, large disturbance events were not well represented by plots in the Amazon (Chambers et al., 2013), and the representation of disturbance events in this region has been thoroughly discussed (see Fisher et al., 2008; Lloyd et al., 2009; Muller-Landau et al., 2014) as failing to represent forest mortality can lead to over- or under-estimation of carbon stock.

1.2.2. Remote Detection of Vegetation and of its Properties

Common remote sensing techniques used to study vegetation are either based on passive optical sensing (light reflectance, spectral) of the vegetation surface, or active such as the LiDAR that can go through the forest cover, and these techniques have been extensively used to improve the understanding of forest structure (e.g., Asner et al., 2008; Vitousek et al., 2009). Besides its use in local studies, optical remote sensing also permits detection of vegetation cover at the global scale (Crowther et al., 2015), and its application to follow vegetation change through time (Hansen et al., 2013; Song et al., 2018) has provided insights on world forests dynamics (Baccini et al., 2017).

Remote sensing of plant properties lies in their varying reflectance in different parts of the light spectrum (Figure 1.3). Because of chlorophyll and leaf structure, healthy mature leaves have a high absorbance of blue and red wavelengths, whereas their reflectance is higher in the green wavelengths and dramatically increases past the red wavelengths, in the red-edge and the near infrareds (NIR, see Figure 1.3). Besides photosynthetic pigments, leaves moisture can also be monitored as water has a great absorbance in the short-wave infrareds (SWIR).

7

Figure 1.3. Spectrum of leaf reflectance over the light spectrum from blue to medium infrareds. Reflectance decreases correspondingly to the presence of chlorophyll a and b in the blue and red wavelengths, water absorbs light at different infrared wavelengths. Corresponding bands from sensors of Landsat 5, 7, and 8 are indicated.

Figure adapted from Pu (2017).

The change in reflectance over different wavelengths have been used to detect vegetation and as proxies for leaf and vegetation canopy structural and physiological properties using vegetation indices (VIs) that involve multiple wavelengths and coefficients, such as ratio of red-NIR (chlorophyll) and NIR-SWIR (water). Sensitivity to vegetation characteristics vary greatly among VIs (Huete et al., 1997a, review by Ollinger, 2011), and some have been largely used, such as the Enhanced Vegetation Index (EVI), to study large areas. For example, Huete et al. (2006) used EVI to confirm plots observations in the Amazon, and demonstrated that the amount of sunlight was more determinant for rainforest dynamics than rainfall during the dry season. Eddy flux measurements further confirmed these observations and suggested that VIs could be proxies for the estimation of gross primary productivity. In the Amazon and the Andes, Asner et al. (2015) demonstrated that leaf chemical traits could be monitored

8

using remote sensing (LiDAR and spectroscopy), so as to avoid the lack of landscape representativeness of plots in that region (see also Marvin et al., 2014).

1.2.3. Studies on Forest Landscapes Disturbances through Remote Sensing

Remote sensing techniques are appropriate to monitor forest disturbances (reviewed by Frolking et al., 2009), and among them, vegetation indices, used to compare pre- and post-disturbances images, have been the basis of numerous studies.

VIs have been used to follow several disturbance types such as herbivory (Townsend et al., 2012), deforestation (Schultz et al., 2016), and fire (Kane et al., 2014). Forest change caused by cyclones has also been the focus of several studies using either VIs or other derives of aerial images, such as the identification of non-photosynthetic vegetation (NPV) or the disturbance index based on tasselled cap transformation (Ayala-Silva &

Twumasi, 2004; Wang et al., 2010; Rossi et al., 2013; Negrón-Juárez et al., 2014; Hu &

Smith, 2018; de Beurs et al., 2019; Hall et al., 2020 among others). In combination with ground surveys, studies have permitted to identify VIs that are sensible to vegetation change such as the EVI (Rossi et al., 2013), and NDII (Normalized Difference Infrared Index, Wang et al., 2010). However, analyses of forests dynamics have also been done using aerial images alone, such as the assessment of the effects of two major hurricanes in Puerto Rico showing the effect of distance to cyclone eye and the surface type on hurricane effects (Hu & Smith, 2018). In Taiwan, Normalized Difference Vegetation Index (NDVI) has been used to assess typhoon damages in Lienhuachih and Fushan forests, and their relationships with topography (Kang et al., 2005; Lee et al., 2008).