1. Introduction
Increasing emissions of greenhouse gases are acknowledged by the scientific community to result in a significant increase in the global mean temperature [
1]. In some regions, this is likely to increase the frequency and severity of droughts and heatwaves [
2,
3]. Resulting from these extreme weather phenomena, long-term field studies are reporting tree growth decline and mortality [
4,
5,
6,
7]. Anomalously long or intense mortality events can have long-term impacts on a range of ecosystems and populations [
8]. Mortality can impact biodiversity functions and ecosystem services such as carbon and nutrient cycling, which could exacerbate biophysical and biochemical climate feedbacks [
8]. One potential consequence of such mortality events is the increased growth of understory vegetation, which in turn can impact successional pathways, productivity and surface hydrology [
5]. Therefore, increasing tree mortality can reduce competition among plant communities, leading to a reduction in the ability of forests to absorb CO
2 [
6].
Tree mortality can take place through different mechanisms. For instance, hydraulic failure occurs when water supply is reduced during high evaporative demand, causing xylem conduits to become air-filled. This halts the flow of water, bringing plant tissues to complete cellular death [
5,
9]. Another mechanism, carbon starvation, occurs when plant stomata close to prevent hydraulic failure, reducing carbon uptake at a time of continued metabolic demand for carbohydrates [
5]. Carbon starvation may be exacerbated during drought by photoinhibition and increased respiratory demands associated with elevated temperatures [
10]. Hydraulic failure occurs if a drought is sufficiently intense for plants to run out of water before they run out of carbon [
11]. In tropical forests, mortality is driven by a combination of hydraulic failure and carbon starvation processes: where mortality is most likely triggered by hydraulic processes, their effects can be aggravated by rapid limitations in carbon uptake [
9].
In the last decade, machine learning (ML) has become a favored tool of remote sensing studies, as a set of computational algorithms and techniques that acquire knowledge from existing data based on inference strategies [
12,
13,
14]. The principle of these algorithms is to model complex classes and accept a variety of input data without making assumptions about the underlying statistical distribution of a given dataset [
12]. In remote sensing, a number of studies have found that ML methods tend to produce higher accuracy than traditional parametric classifiers [
12,
15,
16,
17]. Among ML algorithms, other approaches such as artificial neural networks (ANN), deep learning (DL), decision trees (DT), boosting machines (BM) and support vector machines (SVM) can be found.
ANN and DL algorithms can map features to classes by associating elements in one set of data with elements in a second set, motivated by the assumption that the human brain and artificial intelligence apply similar decision criteria to classification tasks [
18]. DL is similar to ANN but uses deeper neural networks using various hierarchical representations [
19]. DT is among the most intuitive simple classifiers due to its flexibility, intuitive simplicity and computational efficiency [
13]. Random forest (RF) is a specific DT classification model that produces multiple subsets of training samples and variables that are randomly selected [
20]. In RF, the same sample can be selected several times, while others may not be selected at all [
21]. Likewise, BM models are DTs that incorporate a process known as “boosting”: this algorithm generates an ensemble of decision trees, where each successive tree is fitted with the remaining residuals from the previous trees [
12]. Finally, SVM is very popular in remote sensing because of its ability to classify highly dimensional data with a limited number of training samples [
22].
Unmanned aerial vehicles (UAVs) are well suited to addressing current issues in the remote sensing of tropical forest ecology and conservation [
23]. Compared with manned aircrafts, UAVs are more flexible and economically affordable, which enables data acquisition of plant canopy measurements at optimal weather conditions [
23]. In tropical regions, UAV-derived information has been used in a wide range of ecological applications. For instance, in neotropical dry forests, UAV derived information has been applied to conservation biology [
24], the detection of liana infested regions [
25], ecological monitoring [
26], latent heat flux [
27] and canopy temperature of liana-infested and non-liana infested areas [
28].
Given the emerging importance of quantifying tree mortality as a metric of the impacts of climate change in tropical forests, we explore the use of UAVs and ten machine learning models to detect and quantify dead woody components at a tropical dry forest (TDF) site. We conducted this study in five forest plots that cover a gradient of secondary TDFs, in order to evaluate the role of ecosystem succession on the extent of tree mortality.
2. Materials and Methods
2.1. Study Site
We conducted this study at the Santa Rosa National Park Environmental Monitoring Super Site (SR-EMSS), Guanacaste, Costa Rica (
Figure 1). The SRNP-EMSS has a mosaic of TDFs in various ecological successional stages that once suffered from intense deforestation [
15]. The mean temperature is 25 °C and the average annual rainfall is 1750 mm [
29]. The dry season lasts for a minimum of 5–6 months, and it usually extends from approximately late December to mid-May [
29].
In this study, a gradient of the SR-EMSS was sampled following the findings of Li et al. [
25] (
Table 1). In this area, succession is divided into early, intermediate and late forests based on age since abandonment [
30]. At the SRNP-EMSS, early forests are composed of patches of woody vegetation, which include several species of shrubs, small trees and young trees with a maximum height of approximately 6–8 m. Trees at the early stages of TDFs lose nearly all their leaves during the dry season [
31]. Early forest stages are dominated by species well adapted to open habitats, such as
Cochlospermum vitifolium,
Gliricidia sepium and
Rehdera trinervis, as well as sun-loving species (heliophytes) that have anemochory and autochory dispersal syndromes [
32]. The intermediate and late successional stages show significant differences in structure and composition [
33]. These differences are generally driven by species turnover, which causes a very dynamic structure and forest species composition [
34]. The intermediate and late successional stages have two vegetation layers. The first layer encompasses fast-growing deciduous tree species that reach a maximum height of 10–15 m. The second layer is below the canopy and is composed of lianas (woody vines) and adults of more shade-tolerant evergreen species and juveniles of many species [
32,
33]. Dominant species in the early stage are
Rehdera trinervis and
Guazuma ulmifolia, whereas
Calycophyllum candidissimum and
Hymenaea courbaril are dominant in the late stage [
33]. Not all trees on the intermediate and late stage are deciduous: several evergreen species are present.
2.2. Field Acquisition
Field work was conducted in five 200 × 100 m plots, listed in
Table 1, from May to July 2017. In the field, observations were made on the following categories: (1) dead components, (2) living components and (3) understory. The dead components category comprises the following: (1.1) dead woody components, (1.2) dead stand trees, (1.3) dead fallen trees, (1.4) non-photosynthetic woody components within the tree crown and (1.5) dead woody components of lianas (woody vines). The living components category comprises the following: (2.1) healthy canopy trees and (2.2) healthy lianas within the tree crowns. The understory category includes the following: (3.1) understory vegetation (shrubs, small trees and young trees), (3.2) canopy gaps (grass-like vegetation, vines, shrubs and small trees), (3.3) exposed rocks and soils and (3.4) shadowed vegetation.
On each plot, the locations of 50 dead components, 50 living components and 50 understory components were recorded. This was achieved by systematically surveying each plot with transects every 25 m along the short side of the plot. A compass and a Trimble® GeoXT® 6000 differential GPS (average precision of 0.5 m horizontal and 0.54 m vertical; Trimble, Sunnyvale, CA, USA) with a Hurricane antenna were used to record the locations. These components are referred to as ground control points (GCPs) from herein.
A RedEdgeTM 3 (MicaSense, Seattle, WA, USA) multispectral camera onboard a Draganflyer XP-4 (DraganFly Inc., Saskatoon, Canada), operating at 120 m height from the ground, was used to collect images at all plots. The Draganflyer XP-4 is a quadcopter equipped with a three-axes electronic gimbal. The airframe was equipped with a three-axes gyrostabilizer, magnetometer and accelerometer. The RedEdgeTM 3 camera has five lenses with a focal length of 5.5 mm, lens field of view of 47.2° and 1280 × 960 pixels. Each lens provides a separate 16-bit GeoTIFF image centered on a specific wavelength and full width at half maximum (FWHM): blue at 475 nm (FWHM: 20 nm), green at 560 nm (FWHM 20 nm), red at 668 nm (FWHM: 10 nm), red edge at 717 nm (FWHM: 10 nm) and near-infrared at 840 nm (FWHM: 40 nm).
Spectral signatures and multispectral images of three reference panels (a white Spectralon panel, a grey RedEdgeTM 3 panel and flat black presentation cardboard) were collected prior to and after each flight to perform a radiometric calibration to surface reflectance. Specifically, at each site, 20 RedEdgeTM 3 images were acquired at a 1.5 m distance from the reference materials. Moreover, 20 spectra were acquired of every reference material, at a 0.75 m distance from the panel. These spectra were acquired with a UniSpec-SC Dual Channel Spectrometer (PP Systems, Amesbury, MA, USA) that has a wavelength range of 310–1100 nm, FWHM of <10 nm and a sampling of 3.3 nm. The instrument’s dark signal noise removal was performed by taking a dark scan in the beginning of the measurements and later, after every ten sampled measurements. Similarly, the integration time was adjusted with the fiber-optic exposed to a white reference panel, also done in the beginning and after every ten samples. The spectra were acquired by averaging 10 scans.
2.3. Data Preprocessing
The UAV image preprocessing workflow involved three steps: (1) radiometric calibration, (2) mosaicking, and (3) data reduction and transformation.
2.3.1. Radiometric Correction and Mosaicking
To radiometrically correct the RedEdge
TM 3 UAV images, an empirical line method was used, as suggested by Kalacska et al. [
35] and Smith and Milton [
36]. The generation of orthomosaics was performed by the Pix4Dmapper (Pix4D Pro, Lausanne, Switzerland, version v3.3.29). However, in this program, the radiometric correction step was skipped because it was previously performed with the empirical line method. Five single band mosaics were obtained, one mosaic per each RedEdge
TM band. The mosaics were then orthorectified using 7 to 10 GCPs distributed across each plot and described in
Section 2.2.
2.3.2. Data Reduction and Transformation
Despite multispectral sensors having advantages over Red Green Blue (RGB) technology due to their larger number of bands and thus larger amount of information, their use implies a substantial increase in data volume. Image transformation of high-resolution remote sensing data has been proven to be useful for the quantification of forest structure and biomass [
37] and in the estimation of the extension and succession of tropical dry forests [
15], among other applications. In this study, three transformation methods were applied to reduce redundancies in individual bands in order to bring out the total information captured by a combination of bands: principal components analysis (PCA), tasseled cap (TC) and texture analysis (TA) (
Figure 2). All transformations were performed using the five multispectral mosaics. In the case of the PCA transformation, the first principal components were retained, but in the case of the TC transformation, only the third component was retained. The TA transformation was conducted using a Gabor filter, restricted to a maximum of five scales and ten directions, to search for elements in a localized region of an image with specific frequency content in particular directions [
38].
2.4. Classification Models
The “No Free Lunch” theorem of computing sciences states that, without having substantive information about the modeling problem, there is no single model that is better than other models [
39]. In this context, Kuhn and Johnson [
17] suggested trying a wide variety of classification models to determine which model performs better. Consequently, this study used ten ML classification algorithms, shown in
Table 2: support vector machines with linear kernel (SVML), support vector machines with polynomial kernel (SVMP), support vector machines with radial kernel (SVMR), random forest (RF), conditional inference tree (CIT), C4.5-like trees (C45), gradient boosting machines (GBM), neural network (NNT), averaged neural network (ANN) and deep neural network (DNET). This approach covers most of the available classification models of support vector machines, decision trees, boosting machines and artificial neural networks.
2.5. Creation of Training and Validation Datasets
The total number of 2250 pixels of the multispectral mosaics (450 pixels per mosaic, five mosaics in total) was divided into two datasets: training and validation. The training dataset was used for the model development and the validation dataset was used to estimate the performance of the model. This was achieved in three steps. In the first step, a 3 × 3-pixel area was extracted from the multispectral mosaics around each of the 750 ground truth observations (dead components, living components, understory). This step was conducted using the raster extraction function of the QGIS software package (QGIS Development Team, 2009, QGIS Geographic Information System, Open Source Geospatial Foundation, v3.8). In the case of leafless crowns of the dead component’s category, care was taken to select pixels from the central regions of crowns to ensure that dead woody components, as opposed to understory, would dominate the collected spectra. In the third step, this dataset was randomly split into a training dataset (70% of the pixels) and a validation dataset (30% of the pixels) using the “createDataPartition” function of the R program.
2.6. Implementation of Classification Models
The optimal number of training samples was estimated using the bootstrap “632 method” of the “Caret” package of the R program. This was done to avoid overfitting the classification models: the “632 method” creates a performance estimate that is a combination of a simple bootstrap estimate and an estimate from re-predicting the training set [
40]. Bootstrap error rates tend to have less uncertainty than other methods such as k-fold cross-validation, especially if the training set size is small [
17]. Likewise, the optimal tune-up parameters of the classification models were estimated with the “expand.grid:caret” and “tuneGrid:caret” functions of the R program using several combinations of values and parameters, shown in
Table 2. Two kinds of classification models were generated to quantify the extent of mortality across the five TDF successional plots: five plot-specific models and a general model. The plot-specific models were trained only with data from one plot. The general model was trained and validated with data from the five plots. This is in accordance with the findings of Miltiadou et al. [
41], who found that a model that is based on all possible training samples and patterns can perform better than a model generated from a single sampled plot with a lesser number of samples.
2.7. Model Validation
The models were validated by calculating the accuracy and kappa statistics. These statistics were calculated on a per-pixel basis rather than per crown or trunk because the classification was performed using pixels and not objects. Specifically, these metrics were estimated using the validation dataset, with the function “confusionMatrix:caret” of the R program.
2.8. Differences in the Spatial Coverage of the Dead Woody Components between Plots
The spatial coverage of each of the three categories (dead components, living components and understory) of the study was investigated by transforming the multispectral (raster) image mosaics into a vector format. This work was carried out using the conversion function of the QGIS software package. Finally, potential statistical differences between the means of these three categories were studied using the one-way analysis of variance (ANOVA) and Tukey’s post hoc tests.
5. Conclusions
This study demonstrates that it is feasible to detect and quantify dead woody components such as dead stands and fallen trees using multispectral UAV imagery and ML techniques. Of the ML algorithms used in the study, the relatively high accuracy values and low processing times of RF and SVMP made them superior to the other models. Likewise, this study illustrates, on one hand, how the tuning parameters of the ML algorithms affect the accuracy of the classification results, and, on the other hand, how a maximum number of training samples can increase the accuracy of ML classification models.
This study found differences in the coverage of dead woody components across the successional stages of a secondary tropical dry forest. The early successional stages showed the highest coverage of dead woody components, followed by the intermediate stage. Although we found differences between plots, there were no differences in dead woody components between the early and intermediate successional stages.
Finally, further research related to this study could include discrimination of each dead woody component—for instance, the identification of individual dead trees, which could be used for carbon and nutrient cycle modeling.