Sensitivity of Optical Satellites to Estimate Windthrow Tree-Mortality in a Central Amazon Forest

Abstract

Windthrow (i.e., trees broken and uprooted by wind) is a major natural disturbance in Amazon forests. Images from medium-resolution optical satellites combined with extensive field data have allowed researchers to assess patterns of windthrow tree-mortality and to monitor forest recovery over decades of succession in different regions. Although satellites with high spatial-resolution have become available in the last decade, they have not yet been employed for the quantification of windthrow tree-mortality. Here, we address how increasing the spatial resolution of satellites affects plot-to-landscape estimates of windthrow tree-mortality. We combined forest inventory data with Landsat 8 (30 m pixel), Sentinel 2 (10 m), and WorldView 2 (2 m) imagery over an old-growth forest in the Central Amazon that was disturbed by a single windthrow event in November 2015. Remote sensing estimates of windthrow tree-mortality were produced from Spectral Mixture Analysis and evaluated with forest inventory data (i.e., ground true) by using Generalized Linear Models. Field measured windthrow tree-mortality (3 transects and 30 subplots) crossing the entire disturbance gradient was 26.9 ± 11.1% (mean ± 95% CI). Although the three satellites produced reliable and statistically similar estimates (from 26.5% to 30.3%, p < 0.001), Landsat 8 had the most accurate results and efficiently captured field-observed variations in windthrow tree-mortality across the entire gradient of disturbance (Sentinel 2 and WorldView 2 produced the second and third best results, respectively). As expected, mean-associated uncertainties decreased systematically with increasing spatial resolution (i.e., from Landsat 8 to Sentinel 2 and WorldView 2). However, the overall quality of model fits showed the opposite pattern. We suggest that this reflects the influence of a relatively minor disturbance, such as defoliation and crown damage, and the fast growth of natural regeneration, which were not measured in the field nor can be captured by coarser resolution imagery. Our results validate the reliability of Landsat imagery for assessing plot-to-landscape patterns of windthrow tree-mortality in dense and heterogeneous tropical forests. Satellites with high spatial resolution can improve estimates of windthrow severity by allowing the quantification of crown damage and mortality of lower canopy and understory trees. However, this requires the validation of remote sensing metrics using field data at compatible scales.

1. Introduction

Windthrows (i.e., trees snapped and uprooted by wind) are the major mechanism of tree mortality in the Amazon, and can influence forest structure, species composition, and carbon balance [1,2,3,4]. Windthrows are associated with extreme rainfall events, mostly produced by mesoscale convective systems [5]. Under climate change, the intensity and frequency of convective storms is predicted to increase [6,7]. Thus, accurately quantifying the occurrence of windthrows and associated tree mortality can contribute to understanding their ecosystem effects, and to predict how forests will respond to altered disturbance regimes [8,9,10,11].

Optical satellite imagery with medium spatial resolution (from 10 m to 39.9 m pixel size) [12], such as Landsat, combined with field data has allowed researchers to assess patterns of tree damage and mortality, and to monitor forest recovery over decades of subsequent succession in different regions [13,14,15]. These studies demonstrated that windthrows affect a large portion of the Amazon and occur with a range of intensities and sizes. The Central and Northwestern Amazon regions concentrate the highest incidence of large-scale (>30 ha) events [13,16,17].

Landsat offers broad coverage of the Earth’s surface and long-term imagery data [18,19,20], allowing for reliable estimates of windthrow tree mortality at the stand, landscape, and regional scales [13,14,17]. However, robust estimates of tree mortality depend on the compatibility between the spatial resolution of optical sensors and the grain size of targets [21,22]. For this reason, using Landsat to detect windthrows formed by clusters of less than ~6–8 dead trees is challenging [23]. Even for larger windthrows (>5 ha), Landsat assessments may underestimate tree mortality by not accurately capturing relatively lower levels of disturbance on the periphery of impacted patches or due to surviving trees and the relatively fast regrowth of the natural regeneration [13,24].

Obtaining reliable estimates of tree mortality using remote sensing is not a trivial task. The quality of such estimates is highly dependent on the target of interest (e.g., single or clustering dead-trees, [25,26]), the methods used to quantify windthrow severity [27], and the characteristics of the employed sensors (e.g., spatial, temporal, and spectral resolutions [21,28,29]). The spatial resolution of satellites has been treated as an important variable in this process as it can improve the detection capability of small and diffuse vegetation damage, and thus minimize the influence of spectral mixing [30]. Spectral mixing tends to dilute the spectral signature of tree damage and mortality with other elements within the pixels. This includes shadow, exposed soil, or surviving trees within a matrix of impacted forest [27,31]. High spatial resolution satellite data, such as QuickBird, IKONOS, GeoEye, and WorldView, are suitable for detecting small canopy gaps in dense tropical forests [32,33,34,35,36] and have been shown to be efficient, even for detecting individual tree mortality in the Amazon [25,26,34,37,38].

In addition to optical satellites, other approaches such as SAR (Synthetic-Aperture Radar [39,40]), LiDAR [41,42,43] and aerial photos [44] have been tested to estimate tree-mortality associated with windthrows. SAR offers data that are less dependent on weather effects and with spatial resolutions comparable to high- and medium-resolution satellites [45,46]. However, this technology has not yet been tested for the Amazon, and a study conducted in a temperate forest showed that windthrow detection was not reliable due to shading and overlap effects that are common in mountainous areas, as well as the polarization and distortions associated with the angle of incidence during image acquisition [46]. Although aerial photos, LiDAR, and high-resolution satellites provide high-quality data, they have relatively restricted spatial and temporal coverage, high costs, and variations in light conditions and angles of view. These aspects restrict the use of such imagery for assessing patterns of forest damage and recovery [47].

Previous studies have successfully related windthrow tree-mortality recorded in forest inventories to changes in reflectance by using Spectral Mixture Analysis (SMA) [27] of Landsat images [13,14,16]. These studies show a consistent correlation between field and remote sensing data. Recent advances in computational techniques using Machine Learning have enabled the automatic detection of windthrows in forests outside the Amazon using high spatial resolution satellite imagery [48,49,50]. These techniques could enhance future research results using high spatial resolution imagery [51,52,53].

Although optical satellites with high spatial resolution (pixel size ≤10 m) have become available for the Amazon in the last decade, they have not yet been applied for mapping windthrows and quantifying associated tree mortality. Therefore, how windthrow tree mortality estimates obtained with high-resolution satellites will be related to ground truth (i.e., precision and accuracy) in dense tropical forests remains unexplored.

Here, we combined optical remote sensing with forest inventory data across a disturbance gradient created by a single windthrow event from 2015 to investigate how spatial resolution affects the accuracy and precision of tree mortality estimates in the Central Amazon. Our study provides a framework for future research that aims to assess the severity of natural disturbances through reliable estimates of tree mortality. We addressed the following questions: (i) How does spatial resolution affect estimates of windthrow tree-mortality? (ii) Which sensor (i.e., Landsat 8, Sentinel 2, and WorldView 2) produces the most reliable estimates of tree mortality across an extent gradient of windthrow severities?

2. Materials and Methods

2.1. Study Area and Sampling Design

This research was conducted in a remote old-growth forest located near Manaus, Brazil (2°53′41″S, 60°16′26″W). This forest was impacted by a convective storm that occurred in November 2015, which propagated destructive wind gusts and rain and caused widespread tree mortality across an area of ~70 hectares (Figure 1a–c).

The windthrown area is covered by a forest transitioning from terra-firme (at the higher portions of the relief) to campinarana, such as described in adjacent areas [54,55,56,57]. Terra-firme is the most common upland forest type in the Amazon basin [58,59]. The lower portions of valleys (regionally known as baixios) can be temporally flooded by relatively small streams [60]. The terra-firme forests have a continuous canopy, dense understory, and high diversity of tree species [1,61,62,63]. Campinaranas are the predominant forest type that cover sandy soil patches with low nutrient content and poor drainage during the rainy season [64]. Compared to terra-firme forests, campinaranas have lower species richness and a distinct floristic composition due to their oligotrophic conditions [65]. In comparison to terra-firme, these forests usually have a decreased tree density, basal area, and biomass; shorter canopy; and are dominated by few tree species [55]. The transition between terra-firme and campinaranas can aggregate species from both environments, and thus produce relatively high species diversity—for instance, having between 63 to 137 species in a single hectare [54]. Our study site is located in a remote area not affected by previous human or other large-scale natural disturbances during the last several decades.

The average annual temperature and precipitation in the region of Manaus (40 km from our study site) is 26.9 ± 0.17 °C and 2.231 ± 118 mm, respectively (mean ± 95% confidence interval for the period of 1970–2016, [2]). This region has an evident dry season from July to September, with monthly precipitation usually being less than 100 mm [13,16]. Soils on the plateaus have generally high clay content that transition to sandy soils in the lower portions [66,67]. In general, the soils have low fertility, low pH, high aluminum concentration, and low organic carbon content [60]. The local relief in our study region is undulating, including plateaus, slopes, and small valleys associated with perennial drainages [60,67].

To quantify windthrow tree-mortality in the field, we carried out a detailed forest inventory following a protocol built up from previous studies [1,2,13,14,23]. Following previous research conducted across different regions of the Amazon [1,2,4,23], we selected areas spanning the entire gradient of tree mortality within the affected forest, which were inspected in the field and through previous assessments that were performed by following well-established protocols [1,2]. Since this is a remote area for which no accessing infrastructure was available (e.g., roads and trails), our sampling also considered logistics limitations and the safety of our field crew. The forest inventory was performed over two campaigns conducted in December 2016 (~1 year after the windthrow) and April 2017 (~1.5 years after the windthrow). We established three 20 m × 125 m transects, subdivided into 10 subplots of 10 m × 25 m (250 m²) (total of 30 subplots, hereafter referred to as field subplots). Our transects crossed the entire disturbance gradient, ranging from areas with little or no disturbance (i.e., unimpacted old-growth forest) to severely impacted forest with few or no surviving trees. In the field subplots, we recorded, identified, and measured the diameter at breast height (DBH) of all living trees with DBH ≥ 10 cm. Due to logistic and safety issues, we did not count dead trees in severely impacted areas. With this we also aimed to reduce random errors due to possible missing trees hidden below the large amount of coarse wood debris that is typical of windthrows (Figures S1–S4) [1,14]. Instead, we estimated the number of dead trees by subtracting the number of living/surviving trees recorded in each subplot from 15, which corresponds to mean tree density in adjacent old-growth forests (i.e., ~590 trees ha⁻¹, or 15 trees in 250 m², Table S1, [1,4,54,55,56,68,69,70]). The subplots with more than 15 live trees were considered as undisturbed (i.e., no associated windthrow tree-mortality).

2.2. Spectral Mixture Analysis and Remote Sensing Estimates of Windthrow Tree-Mortality

We used Landsat 8 and Sentinel 2 images acquired before and after the studied windthrow event (11 September 2015 and 27 July 2016, and 25 August 2015 and 30 July 2016, respectively). For the WorldView 2, we only used an image acquired after the event (27 July 2016). Landsat 8 and Sentinel 2 were downloaded from the Google Earth Engine platform (https://earthengine.google.com/, accessed on 25 April 2022) [71], and WorldView 2 was purchased from Digital Globe. We selected images with the minimum percent cloud cover, the closest pass dates to each other, and the greatest proximity between the acquisition date and the occurrence of the studied windthrow (Table S2).

We used Spectral Mixture Analysis (SMA) [72,73] and quantified the fractions of endmembers [27] following a well-known routine that was established in previous studies [9,13,14,16,23]. The endmembers contain specific spectral signatures of multiple elements that make up the forest surface and can be used to compute fraction images for different targets of interest [74,75]. We quantified the fraction of the following endmembers: green vegetation (GV, i.e., photosynthetically active vegetation), non-photosynthetic vegetation (NPV, i.e., dead vegetation), and shade (SHD) [27]. Endmembers were extracted and quantified using the software, ENVI 5.3 [27,76,77,78,79].

We used Landsat 8 and Sentinel 2 imagery with pixel values corrected for top-of-atmosphere (TOA) reflectance. For WorldView 2, a band-by-band radiometric correction factor was used to bring pixel values to the same range as Landsat 8 and Sentinel 2 [80]. As the evaluated satellites have different numbers of bands and respective spectral ranges, we obtained the spectral signatures of the focal endmembers by using the bands common to all of them. For Landsat 8 and Sentinel 2, we used the blue, green, red, NIR, and SWIR bands. Since WorldView 2 has a smaller spectral range, we used the blue, green, red, red edge, and NIR bands. The selected bands are sensitive to the physical, chemical, and anatomical characteristics of the leaves and trunks, thereby maximizing the distinction between GV and NPV [72,74,75].

We used the Purity Pixel Index spectral toolkit ([81], Figure S8) that relied on a well-established and reliable method that has been extensively used [1,14,16,17] to select the purest pixels from all images (e.g., areas where all trees were downed versus unaffected old-growth forest) to acquire the most accurate/representative endmembers of interest [27]. Finally, we conducted a SMA to compute fraction images of GV, NPV, and SHD [73]. Windthrows have a specific spectral signature for the NPV fraction, which originated from the deposition of large amounts of dead leaves and wood debris on the forest floor [82]. To ensure that the GV fraction represented forest patches not affected by the studied windthrow, we selected endmembers within the adjacent old-growth forest. The SHD endmember was selected from rivers in the vicinity of the impacted area.

The SHD endmember quantifies the effects of shading associated with the acquisition angle of images, canopy roughness, topography, and clouds [27,83]. Shading is highly related to the biophysical characteristics of trees and a major aspect contributing to variations in radiance and reflectance of canopies [74]. In tropical forests, shade can represent 30% of the canopy pixels and this effect can be observed in the red and NIR bands of medium- and high-resolution imagery [31]. In windthrown forests, a systematic reduction in the shadow fraction can be expected; however, the shadow fraction remains relatively high in areas that were less impacted and at the edge of gaps [23]. We conducted our analyses using the normalized GV and NPV fractions, for which the sum equals to 1 using the following equation:

$${\mathrm{NPV}}_{\mathrm{norm}}=\frac{\mathrm{NPV}}{\left(\mathrm{GV}+\mathrm{NPV}\right)}$$

(1)

where NPV_norm is the normalized values of the endmembers ranging from 0 to 1 [27]. This implies that a reduction in one fraction leads to an increase in the other fraction by the same proportion [27] (Figure S10). The best images were selected according to [84], our field observations, and the spatial distribution of NPV_norm values across the windthrow patches visible in the RGB compositions. We checked the histograms of the NPV_norm images and selected those with more than 98% of the pixels with values between 0 and 1 [85]. We also used the residual error (RMS) from the SMA as a selection criterium (i.e., the lower the better) (Figure 2).

The differences between the NPV_norm before (i.e., old-growth) and after disturbance (ΔNPV) provide a quantitative measure of windthrow tree-mortality [9,86]. The larger the ΔNPV values, the greater the windthrow tree-mortality [9,14,23]. For Landsat 8 and Sentinel 2, we calculated the ΔNPV fraction using the selected pair of images as described earlier (Table S6). In addition, we focused on images whose acquisition date was as close as possible to the windthrow date. We defined the probable date of occurrence based on the analysis of a time-series of Sentinel 2, which has a shorter revisiting time (~5 days) compared to Landsat 8 (~16 days).

This approach could not be implemented for WorldView 2, for which we had a single image after the windthrow. The NPV_norm and ΔNPV for Landsat 8 and Sentinel 2 varied with similar ranges and amplitudes (Figures S5 and S6, Tables S5 and S6). As supported by this pattern and the old-growth stage of forest surrounding the windthrown areas (i.e., no human or large-scale natural disturbances), comparison analyses were conducted for NPV_norm of the three satellites.

We calculated NPV_norm at subplot level (Table S3). For this, we converted the NPV_norm images in raster format to NPV_norm polygons in shapefile format using QGIS [87]. We further intersected the subplot shapefile polygon with the NPV_norm polygons for isolating all segments inside respective subplots. Thus, for each field subplot we obtained one or more segments of NPV_norm. We calculated the area of each NPV_norm polygon to obtain a weighted NPV value for each subplot [1,2,4] as:

$$\mathrm{NPV}=\frac{{\mathrm{NPV}}_{\mathrm{norm}}\ast \mathrm{A}}{250}$$

(2)

where NPV is the weighted value within each subplot, NPV_norm is the normalized value of each pixel, and A is the area of each pixel (m²) that is fully or partially included in the respective subplot; 250 is equivalent to the area (m²) of each subplot.

Apart from the 30 field subplots, we estimated windthrow tree mortality for 100 virtual subplots (10 m × 25 m; hereafter referred to as virtual subplots) that were randomly distributed across the disturbed forest (Figure 1d,e). These virtual subplots were used to evaluate the robustness of tree mortality estimates by the three satellites in adjacent areas containing a greater variation of windthrow severity (Figure 1e). The NPV weighted values for virtual subplots were obtained using the same method described for field subplots.

2.3. Remote Sensing Estimates of Windthrow Tree Mortality

We fitted a Generalized Linear Model (GLM) [88] relating NPV with field-measured windthrow tree-mortality (see Section 2.1). The GLMs were fitted for each satellite. We described the distribution of residuals using the binomial family, and used the logit link function, in which linear predictors must be reversed to the scale of the observations by means of an inverse function [88].

We used the field subplots only for model fitting, thus avoiding overfitting problems for the cases in which the validation was performed using the same subplots [89]. Virtual subplots were not used for validation/training; we focused on evaluating the performance of the models, and also assessed the Residual Deviance using Analysis of Variance (ANOVA). For each model we calculated the Akaike Information Criteria (AIC, [90]), Standard Error of the Estimates (S_yx), Root Mean Square Error (RMSE), Residual Standard Deviation (Sigma), and the Adjusted Coefficient of Determination (R²) computed with the Kullback-Leibler divergence formula (R²_KL, [91]). The R²_KL is suitable for exponential family models (e.g., logit), which retains the informative properties of the fit due to the inclusion of regressors [91,92]. Unlike the linear model, which is based on the Ordinary Least Squares (OLS) approach, R² in these models does not represent the proportion of variability explained, because the binomial distribution does not have the same variability structure as the normal distribution used in simple linear regression [93]. For this reason, some methods for measuring the goodness-of-fit of GLM models, in particular for log-link families (e.g., binomial, exponential, gamma), have been constructed to express one more diagnostic measure of the nonlinear models. Among the available methods, the R²_KL (Kullback’s formula) provides a measure of the divergence between the distribution of observed and estimated values [91,92]. This measure can be used analogously to the R² for quantifying how much of the variation of field-measured tree mortality was captured by the fitted models [94]. The best model was the one with the lower Residual Deviance, AIC, S_yx, RMSE, Sigma, and higher R²_KL.

We assessed the precision (i.e., minimum [Min], maximum [Max], median, and mean) and accuracy (i.e., standard deviation [SD], standard error [SE], and 95% confidence interval of the mean [hereafter referred to as 95% CI]) of remote estimates of tree mortality. For comparing our results with those reported for other Amazon regions [2,4], we also expressed windthrow tree-mortality as categories of severities: old-growth/undisturbed forest [≤4% of windthrow tree mortality]; low windthrow severity [4% < windthrow tree mortality ≤ 20%]; moderate windthrow severity [20% < windthrow tree mortality ≤ 40%]; high windthrow severity [40% < windthrow tree mortality ≤ 60%]; and extreme windthrow severity [>60% of windthrow tree mortality]).

2.4. Statistical Analysis

To determine the most appropriate statistical tools for this study, we assessed the assumptions required for the use of parametric and non-parametric approaches [95]. We tested whether our measurements and estimates of windthrow tree mortality were normally and homogeneously distributed using the Shapiro-Wilk and Levene variance tests, respectively. As our data were not normally distributed and field subplots were nested in the transects, we further conducted a non-parametric approach. We further applied the Kruskal-Wallis test on medians [93] to assess possible differences between the distribution of estimated and measured values of windthrow tree mortality. Differences between field and remote sensing estimates were assessed using the Mann-Whitney post-hoc test, which allows for comparing paired data sets. Although the arithmetic mean is not a precise measure of central tendency for non-parametric data [93], we also reported weighted mean values of field-measured and remote-estimated tree mortality, as supported by the Central Limit Theorem [96]. We further analyzed differences among satellites using ANOVA, followed by the Tukey post-hoc test. Statistical analyses were conducted in the R 4.2.2 software [97] and based on a probability level of 95%.

3. Results

3.1. How Does Spatial Resolution Affect Satellite Estimates of Windthrow Tree-Mortality?

Across all field subplots (0.75 ha), we sampled 341 live and 121 dead trees. This is equivalent to 616 ± 13.3 trees ha⁻¹ (455 ± 70.2 and 161 ± 63.7 live and dead trees ha⁻¹, respectively) (mean ± 95% CI). Live trees were distributed across 142 species, 86 genera, and 36 families. The mean windthrow tree-mortality measured in the field was 26.9 ± 11.1%, with a median of 13%, and minimum and maximum values of 0% and 93%, respectively (Table S3).

Variations in windthrow tree mortality were efficiently captured by three satellites evaluated in our study. Overall, the best agreement between field measurements and satellite estimates (i.e., lower Residual Deviance, AIC, S_yx, RMSE, Sigma, and higher R²_KL) was achieved with Landsat 8, followed by Sentinel 2 and WorldView 2. This result indicates that increasing the spatial resolution (i.e., from Landsat 8 to Sentinel 2 and WorldView 2) implied a systematic loss of quality of fit of the models. Moreover, increasing the spatial resolution did not consistently modify the intercept and slope values of models. Surprisingly, all satellites showed minimum values of tree mortality ranging from ~9–11% (intercept) and a slope close to one (

Table 1. Fitting summary of GLM models used to relate field with remote sensing estimates of windthrow tree-mortality in a Central Amazon forest, Brazil. AIC = Akaike Information Criteria; S_yx = standard error of the estimates; RMSE = root mean square error; Sigma = residual standard deviation; R²_KL = Kullback-Leibler coefficient of determination.

Model	Residual Deviance	AIC	Syx	RMSE	Sigma	R2KL	Coefficients
							a (Intercept)	b (Slope)
Landsat 8	125.33	183.37	0.2096	0.194	2.116	0.4342	9.08	0.9837
Sentinel 2	136.51	194.55	0.2211	0.209	2.208	0.3837	11.21	0.9719
WorldView 2	150.01	208.05	0.2234	0.219	2.315	0.3237	10.61	0.9977

).

We found a strong and positive correlation between NPV_norm and field measurements of windthrow tree-mortality (p < 0.001, all, Figure 3). The mean and median of remote estimates were statistically similar for both field (F = 0.003 and p = 0.99 [mean test]; X²_{Kruskal-Wallis} = 3.23 and p-value = 0.36 [median test]) and virtual subplots (F = 0.70 and p = 0.49 [mean test]; X²_{Kruskal-Wallis} = 1.58 and p-value = 0.45 [median test])). Increasing the spatial resolution of images led to an overall decrease of model uncertainties (lower standard deviation, standard error, confidence interval, and interquartile interval). There was an increase in the range of estimated values in a narrower range of distribution across the disturbed landscape, which also resulted in systematic reductions of the interquartile ranges (

Table 2. Descriptive statistics of windthrow tree-mortality (%) measured in the field and estimated with remote sensing for a Central Amazon forest, Brazil. Min and Max-minimum and maximum values; Q1 and Q3-first and third median quartiles; Iqr-Interquartile range; SD-Standard Deviation; SE-Standard Error; CI-95% Confidence Interval. * Mean windthrow tree mortality similar among satellites.

Subplot Type	Measure	Min	Max	Median	Q1	Q3	Iqr	Mean	SD	SE	CI
Field	Field	0.0	93.0	13.0	0.0	47.0	47.0	26.9	29.7	5.4	11.1
	Landsat 8	10.3	80.1	16.1	12.8	29.8	17.0	26.5 *	22.5	4.1	8.4
	Sentinel 2	11.4	75.0	16.3	14.0	22.8	8.8	26.5 *	21.2	3.9	7.9
	WorldView 2	12.6	80.8	18.0	15.4	22.3	6.9	26.5 *	19.7	3.6	7.4
Virtual	Landsat 8	11.1	81.1	17.6	15.0	43.7	28.7	30.2	22.1	2.2	4.4
	Sentinel 2	11.7	80.2	19.5	16.2	40.9	24.7	30.3	19.5	1.9	3.9
	WorldView 2	13.1	92.0	21.2	16.8	32.2	15.4	27.4	16.1	1.6	3.2

).

The sensitivity of satellites also varied across the windthrow tree-mortality gradient. Landsat 8 and Sentinel 2 had a similar trend curve that differed from that of WorldView 2 (Figure 3a). Overall, increasing the spatial resolution systematically decreased the quality of tree mortality (Figure 3b). Windthrow tree-mortality obtained from Landsat 8 and Sentinel 2 had a more uniform residual distribution than that of WorldView 2, which indicates that the sensitivity of the former satellites was less susceptible to bias at the extremes of the severity gradient (Figure 3c). Increasing spatial resolution led to higher sensitivity for capturing diffuse tree mortality, possible reflecting crown, and minor damage. However, it also led to underestimated estimates of tree mortality in severely damaged areas (Figure 4).

3.2. Which Sensor Produces the Most Reliable Estimates of Windthrow Tree-Mortality across an Extent Gradient of Windthrow Severity?

All satellites produced consistent estimates of windthrow tree-mortality from moderate to high disturbance severity (i.e., 25% to 42% of windthrow tree-mortality). Remote estimates overestimated field measurements by 5–10% from old-growth forest to moderate windthrow severity. From moderate to extreme windthrow severity, we found the opposite pattern—i.e., 10–15% underestimation (Figure 5a). Remote estimates of tree mortality for our virtual subplots covering a wider disturbance gradient were similar from low to moderate windthrow tree-mortality. From moderate to extreme windthrow tree-mortality, Landsat 8 and Sentinel 2 yielded comparable values, which were also different from those obtained from WorldView 2 (Figure 5b). Although with greater associated uncertainties at specific ranges of tree mortality (low to moderate and high to extreme severity), Landsat 8 and Sentinel 2 captured a relatively higher proportion of the existing variation of windthrow tree mortality over a wider disturbance gradient.

4. Discussion

4.1. Relating Satellite Data and Field Data

Systematic increase of spatial resolution implied changes on the sensitivity of different satellites to capture windthrow tree-mortality at the extremes of the disturbance gradient (i.e., low and severe mortality) (Table 2). In windthrown areas there is an abrupt reduction of GV followed by a consequent increase of NPV. Nonetheless, the NPV signal can be lost in less than a year due to the fast growth of pioneer species and survivor saplings and seedlings originally suppressed at the understory of the forest [1,14,23]. As supported by our data, the loss of the NPV signal might be critical for medium to high spatial resolution imagery with relatively small pixel sizes, which are more likely to be influenced by the fast emergence of photosynthetic/green vegetation (GV) [27] (Figure S10).

The convective storm that impacted our study region occurred about eight months (November/2015) prior to the acquisition of all satellite imagery (July/2016). Therefore, our NPV-fraction images may already contain some GV sign of the regenerating forest. This was evidenced by the relatively high reflectance values in the NIR and red edge bands (wavelength ~0.83 µm), which indicate a relatively high content of GV (i.e., photosynthetically active vegetation) [27,75] mixed with NPV (i.e., dead vegetation, wood debris) (Figure S9). Systematic reductions of spectral mixing with increasing spatial resolution (i.e., from Landsat 8 to Sentinel 2 and further WorldView 2) may be related to specific admixture levels of GV. Although understory trees were not recorded in this study, our data support previous research indicating that high NPV values related to the dead vegetation tend to be rapidly diluted due to the fast natural regeneration (Figure S10).

The detection of tree damage and mortality using satellite data can be affected by several factors, such as canopy heterogeneity, understory density, angle-of-view/illumination, and shadow effects [74,78,98]. Overall, gains in spatial resolution improve the detection accuracy of detecting treefall gaps [33,34,37,42,99] as the spectral information tend to be highly influenced by geometric features of relatively small targets [32]. However, the number of pixels needed to capture the physical properties of the surfaces associated with large-scale windthrows is highly determined by the spectral resolution of imagery [100]. Since the mean reflectance of a given area is strongly influenced by the material with the highest reflectance values [74], the lower the spatial resolution of the sensor (i.e., the larger pixel size), the greater the mixture of signals produced by surfaces with relatively high reflectance. Smaller pixels can reduce these effects [78], which is supported by the higher sensitivity of WorldView 2 to capture diffuse tree mortality in less-disturbed areas and at edges of highly disturbed forest patches. WorldView 2 also produced higher tree mortality estimates than Landsat 8 and Sentinel 2 at low and moderate windthrow severity (4% < windthrow tree-mortality ≤ 40%). Given the relatively large diameter and area (mean of ~35 m², [101,102]) of canopy trees in our study region, this result suggests that WorldView 2 (4 m² pixel area) can potentially capture vegetation damage smaller than that measured in the field (i.e., individual tree-mortality). The intercepts for tree mortality varied from 9–11%, indicating that all satellites overestimated field values (Table 1). The variation of intercepts (Figure 3a)may also be related to scaling mismatches between field and imagery data and fine-scale variations of environmental attributes beyond the scope of this study. Furthermore, our models indicate that uncertainties associated with remote estimates of windthrow tree-mortality are larger at the extremes of the disturbance gradient (Figure 3b). R² values indicate that our models captured between 43% and 56% of the variation in field-measured mortality. We believe these values are robust given the variation of disturbance intensity typical of windthrown forests [1,2,4] and the time gap between the studied event and the acquisition date of employed images. Estimated tree-mortality was accurate (i.e., similar to the ground truth) across an extent range of windthrow severity (25–42%, moderate to high severity). As previously reported, these severities accounted for decadal changes in floristic composition [1,62,103], biomass stock and balance [2,4], organic soil carbon stocks [104], and insect diversity [105,106] in other Central and Northwestern Amazon forests affected by windthrows. Future investigations are needed to improve the remote detection of smaller canopy disturbances across local-to-regional variations in topography, climate, soils, and forest attributes [42,43,44,107].

High-resolution remote sensing models (HR-models) assume that elements are larger than image pixels; low-resolution models (LR-models) assume the opposite [108]. This means that the spatial arrangement of varying elements within an image can be detected directly in HR-models because individual elements can be individualized. This is usually not achieved with LR-models [21,108]. As predicted from our best model (Landsat 8), areas of forest that experienced high severity windthrow lost ≥ 50% of live trees. While a similar estimate of tree mortality can be attributed to two independent areas, they can have contrasting disturbance gradients given by the spatial distribution of damaged/dead trees (i.e., contiguous or diffuse) across impacted areas (Figure 3a).

In windthrows, the process of detecting and estimating tree mortality will transit between HR- and LR-models, depending on the level of disturbance severity and the size of pixels. For example, in a highly impacted area (e.g., assuming 50% dead trees and 50% live trees) inside one Landsat 30 m × 30 m pixel, the spatial autocorrelation among the trees is not detected because tree crowns and trunks are, in general, smaller than the pixel size. In this case, an unique reflectance value (i.e., single pixel) shall represent the different targets (live and dead trees). The resulting measurement is highly influenced by the objects with high reflectance [75,100]. However, in HR-models, the same area is represented by several pixels (e.g., 2 m × 2 m pixel of WorldView 2). In this case, the spatial arrangement of the vegetation becomes relevant because trees are larger than the pixels, which tend to resemble their neighbors because they are likely to capture the signature of the same homogeneous target [27]. In HR-models, the varying information captured by several pixels will represent the entire set of live and dead trees. [21]. Depending on the size and severity of the windthrow in relation to the pixel size, HR- and LR-models will have different relationships between the number of dead trees in the field and the spectral/spatial information extracted at pixel level. Our results support that the more pixels within the subplots, the more detail on the spectral variability across the disturbed forest. However, tree damage (e.g., defoliation, broken branches, and partial loss of crown) should also be quantified in future forest inventories to maximize the potential of existing sensors with higher spatial-resolution than those we employed here (e.g., Planet Scope [109], Ikonos [110] and QuickBird [111]).

A precise match between field and remote sensing data also improves landscape estimates of tree mortality. In our study, pixels outside/inside observational subplots may have been incorrectly accounted for. Our plots were georeferenced with a navigation GPS (Garmin 64 CSx), which has a nominal spatial resolution usually >5 m. This effect was likely more relevant for WorldView 2. Geolocation errors for Landsat 8 are at the order of 30 m within the United States, and 50 m globally [112]. For Sentinel 2 and WorldView 2, global errors are at the order of 12.5 m [113,114] and 3.5 m [115], respectively. However, previous studies using the same plot size indicated a high correspondence between ground truth and the NPV signal obtained with Landsat [1,2].

As previously demonstrated, ∆NPV is highly correlated with windthrow tree mortality [1,2,23]. As supported by our results (Figure S5, Table S5), using NPV_norm for quantifying windthrow tree-mortality also produced robust and unbiased estimates. Nonetheless, the nature of such a relationship can be affected by variations in forest deciduousness and disturbance history (e.g., fragmented forests with larger proportion of edges or those submitted to multiple disturbances). Ideally, imagery covering the target forest should be available for pre- and post-disturbance conditions, and with the shortest possible time-interval between scenes.

4.2. Trade-Off between Precision and Accuracy of Satellite Estimates

By increasing spatial resolution and, consequently, the number of pixels representing a given target, we observed a systematic reduction of uncertainties (SD, SE, and CI) associated with our estimates of windthrow tree-mortality (Table 2). However, reductions in uncertainty were followed by systematic reductions in the quality of model fits, mostly due to large errors at the extremes of the disturbance gradient (i.e., low and extreme windthrow severity). This pattern suggests a trade-off between precision and accuracy [116]. In this case, precision indicates the reduction of uncertainty around the mean estimates, and accuracy indicates the closeness of the estimates in relation to the ground truth. In our study, all satellites delivered tree mortality estimates closer to the ground truth over a wide range of disturbance (from moderate to high windthrow severity). Future studies may extend and replicate our methods in regions with different forest attributes and environmental conditions.

Although we equalized TOA values prior to comparing images, each satellite has a different setup of imaging angles (Nadir) and spectral resolution. The Nadir angle may affect the detection of targets with irregular surfaces [74,98], such as in windthrown forests with high amounts of woody debris, survivor and new trees [1]. Large angles can imply obscuration of images and consequent distortion of geometric and spectral features of targets [111,117,118,119]. Landsat 8 images are acquired under Nadir ~0°, while Sentinel 2 and WorldView 2 have off-Nadir angles (~10.4° and ~20°, respectively). These differences may have reduced the quality of the NPV signal from Sentinel 2, and especially WorldView 2, in relation to that obtained with Landsat 8. This may be critical in relatively small gaps or at the edge of highly impacted areas, where insolation is generally lower and the crown of given trees can promote strong shading effects on their neighbors [31,120,121]. To mitigate problems related to this issue, we normalized the fractions for GV and NPV without shade. This procedure ensures that the shadow fraction is filled by equal proportions of GV and NPV within the pixel [27]. Distortions due to the technical attributes of the different satellites were not the objective of the study, but we believe that these effects would not change the overall patterns we report. Apart from being usually acquired from off-Nadir angles due to acquisition costs or mission priorities, commercial high-resolution imagery, such as WorldView 2, can be costly and restricted to small areas [111]. These aspects limit the application of these for mapping windthrows and monitoring forest recovery across large regions and over time.

Indeed, our results show that spatial resolution alone does not provide sufficient evidence for the mismatch between field and remote data. Downgrading the spatial resolution of Sentinel 2 and WorldView 2 to that of Landsat 8 changes the relative weight of the NPV and GV fractions (i.e., reduction and increase of NPV and GV values, respectively) (Figure S11). This pattern indicates that spatial resolution has a direct effect on the sensitivity to detect windthrow tree-mortality, as well as to distinguish rapid post-disturbance regeneration. This has a direct effect on estimated values of windthrow severity at subplot level. Although higher spatial resolution has the potential to increase the sensitivity to detect crown damage and branch fall [44], it does not seem to have the potential to strongly improve stand-to-landscape estimates of windthrow tree-mortality. The accurate and precise estimates we achieved with Landsat 8 may also be partially related to its large pixel size (900 m²) as in comparison to subplots (250 m²).

Landsat 8 and Sentinel 2 have broad spectral resolutions and spectral bands that span from the visible to the shortwave infrared (0.43–2.2 µm) [122,123]. Meanwhile, WorldView 2 has a spectrum restricted to the visible and near infrared (0.4–1.04 µm) [115]. Different terrain materials mimicking the spectrum of others are common in dense forests, and the use of more channels (i.e., bands) is a general recommendation for extracting more suitable endmembers [27]. Thus, the lower spectral resolution of WorldView 2 may affect the process of selecting endmembers, and thus the overall quality of the Spectral Mixture Analysis.

The characteristics of the satellites we evaluated make them appropriate for specific conditions. While Landsat 8 and Sentinel 2 are freely available and cover the entire planet, WorldView 2 has high costs and covers only specific or by-demand regions. WorldView 2 has a 1.1-day revisiting time [115]; Sentinel 2 has a 5-day time [124], and Landsat 8 has a 16-day time [122]. Shorter revisiting time increases the chances of acquiring scenes shortly before and after the occurrence of windthrows. This aspect is essential to mitigate the contamination of NPV images due to fast natural regeneration. The high cloud cover in the Amazon [125] also limits the availability of images, and shorter revisit intervals can help overcome this limitation.

Landsat has a longer-term collection of images than Sentinel 2 or WorldView 2 [126], which makes it suitable for landscape-scale studies aiming at mapping windthrows and monitoring forest recovery over time [2,4]. However, Landsat may be inaccurate for quantifying and describing disturbances created by clusters of less than 6–8 fallen trees [23]. While restricted for the last 8 years, Sentinel 2 is a promising alternative for stand and landscape assessments of the extent and severity of forest disturbances at the individual-tree level. As Sentinel 2 produces estimates with overall low uncertainties, it may also perform better for quantifying diffuse tree mortality. Due to its continuity over the years [124], Sentinel 2 may soon allow for concealing longer-term analyses with high precision.

WorldView 2 estimates had lower associated uncertainties and were more accurate in areas with <20% tree mortality. However, the relatively small collection of images reduces the potential of using WorldView 2 for monitoring disturbance dynamics and forest recovery. The recently launched Planet NICFI initiative [109,127] conceals high resolution (3–5 m pixel) multispectral imagery with a revisiting interval of less than four days for most of the Amazon basin [128]. This offers a new possibility for quantifying and upscaling major mechanisms of tree damage and mortality regulating forest dynamics and functioning.

5. Conclusions

Increasing the spatial resolution of satellite images reduced the uncertainty of estimates of windthrow tree-mortality but did not improve their accuracy. Compared to WorldView 2 and Sentinel 2, Landsat 8 provided more reliable estimates that reflected our field measurements at the individual tree-level. These results highlight the feasibility of Landsat 8 for mapping windthrows and monitoring forest recovery in the Amazon. To fully benefit from high-resolution images, field surveys used for model calibration should account for individual tree-damage (e.g., defoliation, crown breakage, and branch fall). However, high spatial resolution satellite imagery are expensive and restricted to small areas and time periods, which limits their application for large-scale and long-term assessments of mortality and recovery. Future studies applying high spatial resolution imagery could focus on overcoming the limitations of coarser sensors in detecting small clusters of dead trees and quantifying individual tree-damage. This will improve current knowledge on mechanisms and rates of tree mortality, as well as on associated processes regulating forest dynamics and carbon balance in Amazon forests.