Next Article in Journal
The Decrease in Lake Numbers and Areas in Central Asia Investigated Using a Landsat-Derived Water Dataset
Previous Article in Journal
Sentinel-1 Polarimetry to Map Apple Orchard Damage after a Storm
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Evaluating Multi-Sensors Spectral and Spatial Resolutions for Tree Species Diversity Prediction

1
School of Agricultural, Earth and Environmental Sciences, University of KwaZulu-Natal, Westville Campus, Private Bag X54001, Durban 4000, South Africa
2
Centro de Ciencias de la Atmósfera, Ciudad Universitaria, Universidad Nacional Autónoma de México Investigación Científica s/n, C.U., Coyoacán, Mexico City 04510, Mexico
*
Author to whom correspondence should be addressed.
Remote Sens. 2021, 13(5), 1033; https://doi.org/10.3390/rs13051033
Submission received: 10 January 2021 / Revised: 15 February 2021 / Accepted: 1 March 2021 / Published: 9 March 2021
(This article belongs to the Section Forest Remote Sensing)

Abstract

:
Forests contribute significantly to terrestrial biodiversity conservation. Monitoring of tree species diversity is vital due to climate change factors. Remote sensing imagery is a means of data collection for predicting diversity of tree species. Since various sensors have different spectral and spatial resolutions, it is worth comparing them to ascertain which could influence the accuracy of prediction of tree species diversity. Hence, this study evaluated the influence of the spectral and spatial resolutions of PlanetScope, RapidEye, Sentinel 2 and Landsat 8 images in diversity prediction based on the Shannon diversity index (H′), Simpson diversity Index (D1) and Species richness (S). The Random Forest regression was applied for the prediction using the spectral bands of the sensors as variables. The Sentinel 2 was the best image, producing the highest coefficient of determination (R2) under both the Shannon Index (R2 = 0.926) and the Species richness (R2 = 0.923). Both the Sentinel and RapidEye produced comparable higher accuracy for the Simpson Index (R2 = 0.917 and R2 = 0.915, respectively). The PlanetScope was the second-accurate for the Species richness (R2 = 0.90), whiles the Landsat 8 was the least accurate for the three diversity indices. The outcomes of this study suggest that both the spectral and spatial resolutions influence prediction accuracies of satellite imagery.

Graphical Abstract

1. Introduction

Forests cover about one-third of the earth’s total landmass and contain a large amount of terrestrial biodiversity [1,2]. Forest biodiversity is an expression of the differences among the living organism present in the ecosystem and it is considered as one of the means of measuring forest health and stability [3]. The interdependence and interaction among the species influence and facilitate the provision of ecosystem goods and services [4]. These ecosystem goods and services include carbon sequestration and storage, provision of habitats for wildlife, production of non-timber forest products (NTFPs), regulation of water and biogeochemical cycles [5]. Though forest biodiversity includes trees, animal species and other life forms, trees seem to be the most essential elements as without them there will be no forest and most ecosystem goods and services provision will be hindered.
The prediction and estimation of tree species diversity provide forest managers, ecologists and conservationists information to assist forest management decisions. The spatial information obtained through the estimation of the tree species is vital for effective forest management and biodiversity conservation [6]; and it provides a better understanding of forest ecological processes such as tree growth rates, species recruitment, and net productivity [7]. In recent years, remote sensors have provided data that help predict, estimate and map forests at various levels [8,9]. This is due to its large spatial coverage, less time consumption, and cost-effectiveness as compared to traditional inventories and assessments [10,11,12]. That notwithstanding, the methodological approach that establishes the relationship between remote sensing imagery and field data is identified as a robust means of predicting tree species diversity.
The advances in remote sensors, data characteristics, and processing systems have increased the potential of satellite imagery in providing accurate and robust spatially explicit estimates of tree species diversity. Satellite imagery from sensors has been employed for tree species diversity assessments for various types of forests many areas. The output of these assessments has demonstrated the ability of remote sensing satellite imagery to predict the species diversity based on field derived measured data. Furthermore, many of these studies prescribe images and approaches that could be adopted in the modelling process.
Different types of sensors, including multispectral [9,13,14,15,16,17,18,19], hyperspectral [20,21,22,23,24,25] and active ones like the Light detection and ranging (LiDAR) have been used over the years for the prediction of tree species diversity in different forest types, and climatic zones and scales. The hyperspectral images can predict tree diversity with much accuracy due to its numerous narrow bands [26,27]. On some occasions, the hyperspectral data are fused or combined with LiDAR [28,29]. This data fusion approach helps to take advantage of the ability of the hyperspectral data to detect different vegetation community and the ability of the LIDAR data to measure the structural attributes of trees species. Furthermore, the LiDAR data can bypass cloud cover, which allows incident rays to reach the target feature and the reflected rays to get to the sensor. However, the high cost in the acquiring hyperspectral and LiDAR data has hindered the mass application in diversity prediction. [15].
Multispectral images could be said to have been used much in the prediction of tree species diversity. One of the used most is the Landsat images which have proven useful in the forest zones within which they were applied [16,30,31]. Over the years there has been an improvement in its spectral bands and how they sense vegetation, especially with the Landsat 8 [32]. Since the success of diversity prediction across different forest zones depends on the ability of the spectral bands to correlate to with tree species characteristics, it is important to adopt images that have a high sensitivity to forests. Another satellite imagery that has been used is the Advanced Spaceborne Thermal Emission and Reflection Radiometer [ASTER] [15,33]. It seems to have not had much application in the prediction of diversity across many forest types as compared to the Landsat satellite imagery. However, the studies that have used it have found its spatial and spectral product capable of producing good prediction accuracies [34]. Another remote sensing imagery that has also proven robust and informative and is also freely available is Sentinel 2 imagery. It has a medium resolution and a large number of spectral bands that enhances its accuracy outputs [9,13,35]. It is also one of the images that have been used extensively for many vegetation studies. It is the only remote sensing imagery that has three red edge bands which give it some level of advantage over the other satellite imageries [36]. This is because of the chlorophyll information it contains which contributes to the high sensitivity it has for vegetation. These reasons could be the basis for the high accuracies it produces in diversity studies.
The spectral and spatial products of images including the bands, vegetation indices and texture variables are normally used as the predicting variables. Apart from the spectral bands, one the most used predictors are the normalized difference vegetation index (NDVI) [16,19,37,38,39]. It is derived from the bands with the highest absorption (Red) and reflectance (near infrared), which makes it useful under various conditions. However, one of the drawbacks of the NDVI is data saturation in areas with high leaf area index (LAI). This could also likely affect diversity estimation under certain circumstances. Texture variables such as the Gray Level Co-occurrence Matrix (GLCM) have also contributed to the prediction of diversity [8,40,41]. The spectral bands such as the near infrared (NIR), red edge (RE) and the shortwave infrared (SWIR) have been found to be important in many diversity studies [42,43,44,45]. These are located in the regions in the electromagnetic spectrum that contains vegetation information and including any of them in predictive models could improve performance outputs. The use of any of these variables could be dependent on the forest type and the tree cover density. For instance, most forest in temperate and boreal zones may not be of high density and heterogenous as compared to tropical and subtropical forests. As such forest in the tropical and subtropical forest are likely to require robust predictors as compared to temperate and boreal forests. The capabilities of the predictors may be sensor-dependent and the advancement in their design over the years has made diversity studies much more successful.
The methods and modelling techniques are also one of the main factors that contribute to diversity prediction outputs. Most studies have resorted to the use of regression which is carried out by either parametric or non-parametric machine learning algorithms. The Random Forest (RF) which is a non-parametric algorithm is one of the main algorithms that have been extensively used in predictions [13,46,47]. As a non-parametric algorithm, it does not assume a normal distribution of data and it is optimal to be used for diversity modelling of natural forests due to these characteristics. The linear regression which is a parametric algorithm has as well been used many studies [14,34,48]. The modelling technique in the use of these algorithms is an important factor to consider as one of the things that affects accuracies.
The prediction of tree species diversity in many forests and climatic zones have become necessary with time due to factors such as increasing climate change that are negatively affecting species. The availability and advancement of different sensors are continually being tested for their suitability for diversity modelling as well as increasing knowledge in their application. However, none of these studies has been carried out for subtropical natural forests in the Republic of South Africa, which creates a gap in tree diversity management. It must be noted that subtropical natural forests are characterised by high tree species diversity and density [49,50,51]. As such, it will require informative and robust imagery to predict and map their tree species diversity. Thus, evaluating multi-sensors performance and identifying the best based on their spectral and spatial resolutions is beneficial for the application of imagery in diversity prediction and mapping. Hence, our study aimed to assess how the performance and accuracies of PlanetScope, RapidEye, Sentinel 2 and Landsat 8 images could be influenced by their spectral and spatial resolution in the prediction of tree species diversity for a subtropical natural forest in KwaZulu-Natal (KZN) province, Republic of South Africa. The Shannon Index (H′), Simpson Index (D1) and the Species richness (S) together with RF regression modelling, are utilised to identify which image has a good relationship with them and produce good accuracy. The outcomes of our study will provide information on how spectral and spatial resolution could influence image model accuracies, which can provide a guide in the decision making on the imagery to select for predicting tree species diversity of subtropical natural forests. It will also contribute to existing knowledge and approach in the modelling of diversity for forest management and conservation. Furthermore, it could assist forest managers in devising measures that can enhance the conservation and protection of forest diversity.

2. Materials and Methods

2.1. Study Area

The Nkandla forest reserve is an Afromontane sub-tropical forest type, and it was established in 1918. It is found in the north of KwaZulu-Natal province, Republic of South Africa. It has a total area of 2217 ha and located on 28°43′50.88″S and 30°7′9.84″E (Figure 1). A peak average temperature of 27 °C is experienced between December and January, and the lowest average temperature of 2 °C in the winter months of June and July [52]. It has an undulating and steep topography with an altitude of a minimum level of 500 m and exceeding 1300 m. It has four land cover types made up of closed canopy forest (1,059.23 ha), open canopy forest (910.60 ha), grassland (226.55 ha) and bare sites [20.97 ha] [53]. It has common tree species such as Cryptocarya myrtifolia, Trichilia dregeana, Bridelia micrantha, Elaeodendron croceum, Podocarpus henkelii and Olea capensis.

2.2. Field Inventory and Diversity Indices Estimation

Tree data information was collected from the accessible parts of the Nkandla forest reserve between 24 April 2019 and 7 May 2019. This is because it was observed from a reconnaissance survey that some parts of the forest, especially the western portions were inaccessible due to the presence of high elevation and deep slopes. Therefore, the inventory was restricted to the middle and northeastern parts which have gentle slopes. Existing transects were followed and a systematic approach was used in setting up the sampling plots in the gentle slope and relatively flat terrain. Eleven 100 m × 100 m plots (1 ha) were randomly set up in areas with the gentle slopes and flat terrain. Each of the 1 ha plots was subdivided into 25 subplots of 20 m × 20 m sizes each to facilitate the data collection. Thus, the tree data was obtained from a total of 275 subplots. In each subplot, the diameter at breast height (DBH) of tree species ≥ 5 cm was measured with a diameter tape. Other information recorded for the trees were the species name (local and scientific) and the GPS coordinates of the trees. The individual number of species was summed up for each sampling plot. This approach of tree inventory did not compromise on the data collected because similar number and types of tree species were measured and recording in most of the sample plots. The tree data were further compared and confirmed for similarity with tree list the management plan of the forest obtained from the Ezemvelo KZN Wildlife.
The relative number of each tree species was used to compute the Shannon Index (H′) [54], Simpson Index (D1) [55] and Species richness (S) [56] for each species. The was done by using the mathematical functions in equations 1, 2 and 3 for the three diversity indices respectively. These indices have been well established and they allow for comparison of tree species diversity levels at different scales [57] and they as well help to account for the evenness and richness of diversity for each site. The Species richness takes into consideration the absolute number of species in a particular ecosystem, whiles the evenness takes into consideration the relative abundance of each species [47]. The Shannon Diversity Index (H′) accounts for both the species richness, species abundance [58]. The original Simpson Index (D) emphasizes on the evenness component of diversity [55]. The Shannon Index is sensitive towards species rarity and abundance, whiles the Simpson Index is sensitive towards abundance in species distribution [56]. These indices have been used widely and are confirmed to have a relationship with spectral reflectance of remote sensing sensors [22].
H = i = 1 S p i   x   ln ( p i )
D 1 =   1 i = 1 S p i
S = N
where pi is the proportionate abundance of the ith species in the sampling plot, S is the total number of all species in a sampling plot, and ln is the natural logarithm of the proportionate abundance of species in the sampling plot.

2.3. Remote Sensing Data

We used four different sensors of different spectral and spatial resolutions because the study’s focus was to compare and assess multi-sensor spectral and spatial resolution effects on accuracies in tree species diversity prediction. The satellite imageries used in the study were Landsat 8, Sentinel 2, RapidEye and PlanetScope (Table 1). All the images were cloud-free. The Landsat 8 has 10 spectral bands covering the visible to the shortwave infrared (SWIR) region of the electromagnetic (EM) spectrum with a spatial resolution of 30 m. The Sentinel 2 has a spectral resolution of 13 also ranging from the visible range to the SWIR region of the spectrum with varying spatial resolution. The blue, green, red and near infrared (NIR) spectral bands have a spatial resolution of 10 m, whiles the three edges bands, narrow near infrared (NNIR) and the two shortwave infrared bands (SWIR 1 and SWIR2) have a spatial resolution of 20 m. The coastal aerosol (Band 1), water vapour (Band 9) and cirrus bands (Band 10) have spatial resolution of 60 m. The Landsat 8 and the Sentinel 2 are both freely available imagery that has been used extensively for vegetation related studies. The Landsat 8 is provided by the United States Geological Service (USGS) whiles Sentinel 2 is provided by the European Space Agency (ESA).
The RapidEye have a spatial resolution of 5 m and five spectral bands which ranges from the visible to the NIR region of the EM spectrum. It is also among the sensors that have been used extensively for vegetation studies. On the hand, the PlanetScope is a relatively new sensor and it is yet to be much used in diversity prediction. It has four spectral bands ranging from the visible to the NIR of the EM spectrum with a spatial resolution of 3 m. Both the RapidEye and PlanetScope are commercial sensors provided by the Planet Team.
A Landsat 8 image captured on 8 May 2019 was downloaded from the Earth Explorer website (www.usgs.gov) of the USGS. The Landsat 8 image was atmospherically corrected from Top-of-Atmosphere to surface reflectance using the apparent reflection function in ArcGIS 10.6.1. The coastal aerosol band (Band 1), the panchromatic band (Band 8), Cirrus (Band 9) and thermal infrared bands (Bands 11 and12) were not included in the bands considered for the analysis. They were excluded because the band 1 contains aerosols, band 8 is panchromatic, band 9 contains cloud information, whiles bands 11 and 12 contains thermal information. The Sentinel 2 image was captured on 14 April 2019 and was similarly downloaded from the Earth Explorer website (www.usgs.gov) of the USGS. It was atmospherically corrected using the semi-automatic classification plugin (SCP) of the QGIS 3.10 software. The image radiance was transformed into spectral reflectance with the dark object subtraction (DOS1) SCP plugin of the QGIS 3.10 software. The image was further resampled to 10 m spatial resolution using the SNAP toolbox for the spectral bands to have a uniform resolution, as they are varied. This operation was done to enhance the analysis. The Bands 1, 9 and 10 were excluded because they contain aerosols, water vapour and cloud information respectively. The PlanetScope and the RapidEye images were downloaded from the Planet Explorer website (www.planet.com/www.api.planet.com). The PlanetScope was captured on 30 April 2019 while the RapidEye was captured on 18 June 2019. The two images were atmospherically corrected by the suppliers (Planet Team) and subsequently provided to be downloaded for the analysis. The characteristics of each of the four images have been detailed in Table 1.

2.4. Important Variables Selection

The Recursive Feature Elimination (RFE) algorithm was subsequently used to select important variables to be used as input variables for the Random Forest regression model for each of the four images. This process is very important as it helps to eliminate noisy variables and reduce redundancy and computational complexities [59,60]. The RFE process of elimination is carried out in a stepwise approach involving; (1) the training of the RF model, (2) computing the permutation importance measure, (3) eliminating of the less relevant variables (features) and (4) repeating the first 3 steps until no further variables remain [60]. The most informative variables are ranked in the last stage of the steps of the backward procedure and the algorithm selects a smaller size and more efficient variable subset.
The SWIR1, SWIR2, RE2, NIR and NNIR bands were selected for the Sentinel 2, whereas the Red, NIR and RE bands were selected for the RapidEye. The VNIR bands were maintained by the algorithm for the PlanetScope after the running of several iterations. Lastly, the Green, Red, NIR and SWIR1 bands were selected for the Landsat 8.

2.5. Random Forest Regression Modelling

Random Forest (RF) [61] regression models were used to predict the tree species diversity based on the Shannon diversity (H′) and Simpson diversity (D1) and Species richness (S) derived from the field measured data. The prediction established the relationship between the diversity indices and the spectral characteristics of the image data. The RF is a non-parametric machine learning algorithm which can undertake both classification and regression [61]. A bagging system is used to split the data by the algorithm where a part of the data is used for training and building the decision tree. The remaining set is used for estimating the out-of-bag (OOB) error for each tree. The RF algorithm has an advantage of not overfitting data because there is a convergence of the generalization error when the number of trees increases [61,62]. It is also able to deal with the problem of multicollinearity [63,64]. The RF has two main parameters of the RF that contribute to the accuracies of models. These are the ntree and the mtry and they may be tuned or left in defaults values. The ntree has a default value of 500 and it is the total number of decision trees grown in the model. The default value of the mtry is the total number of predictor variables divided by 3 (N/3) when it used in regression models. Studies that have used the default values of both parameters have obtained satisfactory results [65,66]. Aside from these characteristics, the RF enables the assessment and ranking of statistical significance of each predicting variable in the model with the use of its variable importance feature.
The four models were implemented with the “randomForest” package [67] in the R statistical software environment [68]. The spectral pixel values of each of the four images were extracted and used in the models. The 275 sample plot values of each of the Shannon (H′) and Simpson (D1) diversity indices and Species richness (S) values computed from tree species data were partitioned into 70% training data (192), and 30% independent validation data (83) in a random selection approach. We calibrated each RF regression model with the training data and then applied the bootstrapping of 500 iterations to predict the diversity.
A parameter optimization process was carried out to find the best ntree and mtry values for the RF model of each of the four satellite imageries. The “tuneRF” function in the “randomForest” package was used to find the optimal mtry value for the models. The value obtained after the process was 1 for all the models. On the other hand, the optimal ntree values obtained for the Sentinel 2, RapidEye, PlanetScope and the Landsat 8 models were 600, 500, 900 and 400 respectively. The ntree and the mtry values were then used in models for predicting the tree species diversity. The independent 83 validation set of each image was subsequently used for the validation of prediction accuracies.

2.6. Models Evaluation

The four RF regression models’ predictive abilities were compared and assessed based on two main statistical parameters. These parameters were the coefficient of determination (R2), and the root mean squared error (RMSE). The means of the 500 bootstrapped samples were used to calculate the accuracy parameters values. The RF regression model with the highest R2 and lowest RMSE values was determined as the most accurate.

2.7. Variable Importance

The variable importance feature of the RF algorithm was applied to evaluate and rank the predicting variables according to their statistical importance in contributing to the accuracy of each model. The importance of each variable is determined by the percentage increase in mean squared error (%IncMSE). The %IncMSE denotes the effect of a predicting variable in a model when it is removed from it. This was assessed to determine the spectral bands that play an important role in the prediction and correlated well with the Shannon Diversity Index (H′) and Simpson Diversity Index (D1) and Species richness (S) for the subtropical natural forest.

3. Results

3.1. Field Inventory Data Analysis

The descriptive statistics for the Shannon Index (H′) and Simpson Index (D1) and Species richness (S) that were computed for the field inventory is presented in Table 2.

3.2. Sensor Performance Evaluation

The RF model was utilized to evaluate the performance of the four sensors for the prediction of the tree species diversity for Shannon Index, Simpson Index, and the Species richness. Their performances was evaluated based on the R2 and the RMSE. The model with the highest R2 and lowest RMSE was considered as more accurate and robust.
As illustrated in Table 3, the Sentinel 2 image model was the most accurate (R2 = 0.926, RMSE = 0.148) for the prediction of tree species diversity derived using Shannon Index while the RapidEye emerged as the second accurate (R2 = 0.902, RMSE = 0.147) for the same diversity index. The PlanetScope model was the third accurate (R2 = 0.898, RMSE = 0.156) with the Landsat 8 model being the least accurate (R2 = 0.529, RMSE = 1.748).
The Sentinel 2 and the RapidEye were the most accurate with a comparable accuracy output (R2 = 0.917, RMSE = 0.043 and R2 = 0.915, RMSE = 0.044 respectively) for the tree species prediction tree with the Simpson Diversity Index [D1] (Table 3). Whereas the PlanetScope produced the second-best accuracy (R2 = 0.899, RMSE = 0.045), and Landsat 8 was the least accurate (R2 = 0.410, RMSE = 0.063).
The Sentinel 2 was once more the most accurate (R2 = 0.923, RMSE = 1.983), under the Species richness (S), whiles the PlanetScope was the second accurate (R2 = 0.900, RMSE = 1.293) (Table 3). The RapidEye was the third accurate (R2 = 0.833, RMSE = 1.287), and the Landsat 8 was the least accurate model (R2 = 0.532, RMSE = 1.746).
The statistical evaluation conducted for the prediction has been presented in Table 4. It was observed that there was a slight underestimation for the prediction under the Shannon Index and the Species richness by all the four images. On the other hand, the prediction for with the Simpson Index had the field measured values and the predicted values correlated much better as they were within ranges of each other. Scatter plots produced by each RF model of the imageries which establishes the relationship between the field measured and predicted diversity under the Shannon Index, Simpson Index and Species richness are presented in Figure 2, Figure 3 and Figure 4.

3.3. Predicting Important Variables

The Variable Importance feature of the RF was utilized to rank the importance of each remote sensing variables for the prediction of the tree species diversity. RF regression algorithm provides the percentage increase mean square error (%IncMSE), which was used to rank the variables. The variables for each of the sensors under the Shannon Index, Simpson Index and the Species richness were ranked in decreasing order of importance for each RF model.
Table 5 illustrates the ranking of the important variables of the Sentinel-2 spectral bands used for the prediction of tree species diversity under the Shannon, Simpson, and Species richness indices. The SWIR1 band was the most important variable for the tree species diversity predicted using Shannon index. The second-best to the least important variables were the SWIR 2, RE2, NNIR and the NIR respectively.
For the Simpson Index (D1) predictions, the RE2 was the most important variable, whiles the SWIR1 was the second important variable. The third and fourth positions were occupied by the SWIR2 and NNIR. The NIR was again least important in the prediction. The %IncMSE values indicated that it played a very minimal role for this diversity index. Regarding the Species richness, the SWIR1 was once more the most important variable, whiles the NNIR, SWIR2, RE2 and NIR followed as second to the least, respectively.
The important variable ranking for the RapidEye spectral bands used in the RF model has been illustrated in Table 6 for the Shannon Index, Simpson Index and the Species Richness. The most important variable under the Shannon Index was the Red band. The second was the NIR whereas the RE was last. The ranking of the most important variables for the Simpson Index and the Species richness was the same. The NIR was the most important band, whiles the RE and Red bands were the second and third respectively.
Table 7 displays the important variables for the PlanetScope image under the three diversity indices. The Green and Red bands had the same level of significance under the Shannon Index in their contribution to the accuracy of the image’s model. They shared the first position whiles the Blue band was third and the NIR was the least significant. For the Simpson Index, the Green band was the important variable for the prediction done under the Simpson Index. The NIR was the second-best contributor to the accuracy with the Blue and Red being third and last. With the Species richness, the NIR emerged as best variable and the Green band was the second best. The Red and Blue bands shared the third and fourth position respectively.
The variable importance of the Landsat 8 was not much different from that of the Sentinel 2, RapidEye and the PlanetScope as presented in Table 8. The SWIR1 which is one of the key spectral bands of the Landsat 8 was the most important variable for the Shannon Index. The NIR occupied the second position, followed by the Green and Red bands as third and last respectively. The reverse was the situation under the Simpson Index, where the NIR was the most important and the SWIR1 was the second. The Red band emerged as the third and the Green was the last. In the case of the Species richness, the SWIR1 was the best variable and the NIR following as the second-best variable. The Red and Green bands occupied the third and last position respectively.

4. Discussion

In recent years there has been the launch and availability of free and commercial remote sensors that produce imageries which are adopted for forest vegetation-related research. The spectral and spatial attributes are vital for remote sensing imagery, and these could influence their suitability, and robustness for the characterization and prediction of forest attributes such as tree species diversity [69]. The sensor type influences and contributes much to accuracy [70]. Therefore, the assessment of different sensors based on their spectral and spatial resolution in the prediction and mapping species diversity is beneficial to ecologist and remote sensing experts. It is worth noting, that each sensor does have its strength and limitation [71], as a result of their spectral and spatial resolutions. This was displayed in the accuracy produced under each of the three diversity indices. Furthermore, it indicates the relationship between the predicting spectral variables and the indices.
The Sentinel 2 imagery was the most accurate and performed better than the RapidEye, PlanetScope and Landsat 8 for the prediction using the Shannon Index and the Species richness. It was also the best image together with the RapidEye under the Simpson Index as both had a comparable high R2 and low RMSE. Several factors could account for the higher performance of the Sentinel 2 than the other images. Firstly, the five important spectral bands (RE2, NIR, NNIR, SWIR1 and SWIR 2) selected through the application of the Recursive Feature Elimination (RFE) may have been robust than that of the other three images. The availability of the red edge and the SWIR bands for the Sentinel 2 might have also contributed significantly to its accuracy. The red edge and SWIR bands, which are also positioned in Sentinel 2, have a higher sensitivity to healthy vegetation and minimum susceptibility to saturation [72,73]. These attributes of the bands make them effective for diversity prediction in high density natural forest. It is important to note that the sensitivity of the red edge and the SWIR bands enhance their correlation with vegetation [74]. This sensitivity may be attributed to the narrow bandwidth and their location in the electromagnetic spectrum. It is also likely that the spectral bands of the Sentinel 2 are more informative than that of the RapidEye, PlanetScope and the Landsat 8. This may explain the better correlation of the Sentinel 2 with the field measured Shannon Index, Simpson Index and the Species richness that led to its high accuracy. Also, the larger number of spectral bands used for the RF regression model of the Sentinel 2 could have enhanced its capability and influenced the high accuracy. Findings of Rocchini, Ricotta [75] indicated that a large number of spectral bands increased diversity prediction accuracy, thus, suggesting the preference of large number spectral bands to a small number [76]. The spatial resolution of the Sentinel 2 could have also been a key factor because the pixels of the image are likely to have more tree species falling within it. As a result, more information on vegetation might have been preserved for the image. Since the three diversity indices rely on the types and number of species, the ability of the image to have more trees falling well within its pixels is vital for higher accuracy in predictions. Our study shares a similarity with Mallinis, Chrysafis [47], who also found the Sentinel 2 performing better than the RapidEye in species diversity prediction in the Mediterranean region. Among other reasons, the study indicated that the absence of SWIR bands for the RapidEye could be a contributing factor, which have some relations to our findings. The inherent capability of the Sentinel 2 that enhances the detection and characterizing of vegetation have been confirmed in other research [13,77,78], which further validates our findings.
A knowledge of the variables that contributed most to the accuracy of models is important in modelling. It helps to select key variables that are robust, reduces redundancy and noise in the prediction and characterisation of vegetation attributes [59,79]. With regards to the Sentinel 2, the RE2, SWIR1 and the SWIR2 contributed significantly to accuracy outputs both under the Shannon index and the Species richness, mainly due to their high sensitivity to vegetation. Immitzer, Neuwirth [77] also observed that the red edge and the SWIR bands were useful and produced better accuracy for broadleaf species classification. In addition, the importance of the red edge band is emphasized by Grabska, Hostert [9], while Persson, Lindberg [35] and [80] highlights the significance of the SWIR vegetation variability classification and separation. It is worth stating that, though the NIR had a higher reflectance for healthy vegetation, it was the least contributor to the higher accuracy of the Sentinel 2. It was not robust enough for the prediction as it could not enhance the capabilities of the image. With the advancement and increase in remote sensing imagery and their application to vegetation and forests attribute characterisation and mapping, the identification of these key bands is vital.
In the prediction with the Simpson Index, the RapidEye performed better than the PlanetScope and the Landsat 8 as it produced a comparable higher accuracy together with the Sentinel 2. This could have been due to the availability of the red edge and the NIR band for the RapidEye [77,80,81], which may have significantly contributed to the higher accuracy it produced under this diversity index. Though, it has been suggested that having a larger number of variables are important [75,76], it is also possible that selecting few but robust and informative bands as inputs variables for a model could help produce noise and produce higher accuracies. That might have worked for the RapidEye under the Simpson Index. On the other hand, its finner spatial resolution could have had an effect on accuracies under the Shannon Index and the Species richness. It is indicated that higher spatial resolution of satellite imageries usually contain the structural attributes of vegetation community, but some information on the species type and the relative abundance is lost [82]. This may further account for why it placed second to the Sentinel 2 under the Shannon Index and Species richness. Taking individually, its coefficient of determination for the three diversity indices ranged between 0.83 and 0.92, accounting for its good explanation of the variance and suitability for diversity modelling. The RapidEye has been found useful in vegetation studies such as intra and inter-species biomass prediction [83], forest structural information [84], tree species classification [10] and urban vegetation classification [43]. Hence, it could further be evaluated in similar studies to ascertain its suitability for diversity prediction.
The PlanetScope is a relatively new image as compared to the RapidEye, Sentinel 2 and Landsat 8. It was the second-best image for the Simpson Index, but the third-best for the Shannon and the Species Richness in the prediction. Though four spectral bands were used for its RF regression model, its spectral bands are likely less informative and sensitive to vegetation as compared to the RapidEye and the Sentinel 2. Its bands are made up of only the visible and near infrared (VNIR) and lacks bands such as the red edge and the SWIR. This might have also accounted to the low accuracies it had as compared to the Sentinel 2 and the RapidEye. As identified by our study findings and other vegetation related studies [47], the red edge and SWIR are very useful and contributes to model accuracies. Similarly, to the RapidEye, the fine spatial resolution of the PlanetScope might have also reduced its ability to have a high number of species, thereby producing lower accuracies for the Shannon Index and the Species richness. On a positive side, it has a very good temporal resolution (revisit time) of one day, which makes it a suitable image for time series species diversity studies. It could also be accessed for vegetation phenological and seasonal variation studies because of the daily revisit time that could capture seasonal changes observed in vegetation. In the variable importance assessment, the Green and NIR bands were much accurate respectively for the Shannon Index, Simpson Index and the Species richness. Generally, the VNIR bands are common to most satellite images and are sensitive and correlate well with vegetation [32]. Among the VNIR bands, the Red, Green and the NIR have high reflectance for healthy vegetation and could be considered as part of the spectral bands employed for diversity prediction in high density natural subtropical forests.
The low performance of the Landsat may be directly related to the low spatial resolution as compared to the other images. Its accuracy for the Shannon and Simpson indices were just about half of that of the other images. Contrary to the findings of our studies, it has provided satisfactory accuracies in studies, [30,85], though it was not compared with other images. On a more general basis, it is among the images that have been used for vegetation studies including diversity [18,19,30]. Furthermore, its bands have been designed and improved for detecting and mapping vegetation [86,87], and it has proven to be useful for those vegetation studies. Similarly, the most important variables among the spectral bands used for the prediction under the three indices were the SWIR 1 and the NIR. The importance of these bands needs not to be overemphasized as their capabilities have already been indicated for the other images. On an individual basis, it may be useful for diversity prediction as has been found in vegetation related studies. Its high amount of historical data could be explored for multitemporal and time series diversity studies.
Generally, the spectral bands had a high relationship with the Shannon index, Simpson index and the Species richness with most of the accuracies for the Sentinel 2, RapidEye and PlanetScope. Successful diversity estimation with the utilisation of remote sensing data would be dependent on the spectral variables that could suitably capture the species diversity for the landscape in question [30]. Therefore, spectral bands in the VNIR up to the SWIR region could be used to further ascertain their suitability for diversity prediction and mapping in natural subtropical forests.
Concerning the diversity indices, the use of either one of them could be dependent on the objective of the study, the forest type and the image. Spectral bands respond differently to them in their application to diversity prediction. However, little attention has been given to finding out much about their sensitivity to the species distribution patterns [30], with the use of spectral variables. Since species abundance, richness and evenness are likely to change with time, it may be important to determine the indices that best correlates with spectral variables through seasonal and temporal studies.
The Random Forest regression algorithm was very beneficial in the prediction by each image model. It demonstrated the capability to handle different types of complex remote sensing image data [88]. Since it is a non-parametric machine learning algorithm, it does not assume normality [89]. This attribute is useful for natural forests since they are mostly heterogeneous and do not have a normal distribution. Furthermore, it can handle redundancy, reduce noise and deal with multicollinearity [61,63]. All these might have probably influenced the functioning of the models to produce satisfactory accuracies. It could explain why the RF is mostly adopted for most vegetation related studies including diversity prediction.
The findings of our study have shown the capability of the images and important spectral bands most especially for the Sentinel 2, RapidEye and PlanetScope that are optimal for the prediction and mapping of tree species diversity. The output of our study is important for forest managers and ecologists in the modelling and prediction of trees species diversity. This could assist forest managers and ecologist in the selection of images and spectral bands for the prediction of diversity in natural subtropical forests. Generally, it could assist in the application of remote sensing technology and modelling in the estimation of diversity.

5. Conclusions

Our study assessed how spectral and spatial resolutions influence the accuracy of remote sensing imagery models based on the Shannon index, Simpson index and Species richness for the Nkandla natural forest in the Republic of South Africa. Since various sensors perceive vegetation differently based on their spatial and spectral resolutions, finding a suitable one for the prediction of the tree species diversity in high density natural forest is important. It has been demonstrated in our studies and others that both the spectral and spatial resolutions of satellite imagery have much influence on the accuracies of images. The medium spatial resolution of Sentinel 2 and its spectral resolution makes it more capable in the prediction of the diversity. Though the RapidEye, PlanetScope and the Landsat 8 had lower performances than the Sentinel 2, it is not indicative that they may not be used for diversity prediction in natural subtropical forests. Since their abilities has been demonstrated in our study, they may be used to further ascertain the condition under which they could work better. On an individual basis, each of imageries may be applied as they produced satisfactory accuracies. Also, since there are no generic spectral and spatial resolutions for diversity prediction currently, more studies could be carried out to test different sensors in various forest types to ascertain which could work much better.

Author Contributions

E.G.-A., M.G. and A.M.-P. conceptualized the study. E.G.-A. and M.G. directed and coordinated the research as well as the analysis. E.G.-A. led the writing, with contributions from M.G. and A.M.-P. All authors have read and agreed to the published version of the manuscript.

Funding

This research and study received a block grant from the National Research Foundation (NRF) through the South African System Analysis Centre (SASAC) of the Republic of South Africa, with grant number 118770. The funders had no role in the study design, analysis, preparation of the manuscript and decision to publish it.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data for this article has been placed on Mendeley data. It could be accessed with the link; http://dx.doi.org/10.17632/jwph4yf788.1.

Acknowledgments

The authors are grateful to the Planet Explorer Team for providing freely RapidEye and Planet scope under their Education and Research programme. The authors appreciate the support pro-vided by Sharon Louw (Ecologist) of Ezemvelo KwaZulu-Natal Wildlife (EKZNW) in secur-ing a permit to conduct the study in the Nkandla forest reserve. We thank Elliackim Zungu (Conservation Manager), Simon Makhaye, Sandebulawa Biyela and other supporting staff of EKZNW for assisting in field data collection. We also thank Charles Zungunde (University of KwaZulu-Natal) for driving the field team across the forest as well as assisting in data collection. Finally, we thank the journal editors and reviewers for their efforts, time and in-sights.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Gamfeldt, L.; Snäll, T.; Bagchi, R.; Jonsson, M.; Gustafsson, L.; Kjellander, P.; Ruiz-Jaen, M.C.; Fröberg, M.; Stendahl, J.; Philipson, C.D. Higher levels of multiple ecosystem services are found in forests with more tree species. Nat. Commun. 2013, 4, 1340. [Google Scholar] [CrossRef]
  2. Aerts, R.; Honnay, O. Forest restoration, biodiversity and ecosystem functioning. J. BMC Ecol. 2011, 11, 29. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  3. Wang, R.; Gamon, J.A. Remote sensing of terrestrial plant biodiversity. Remote Sens. Environ. 2019, 231, 111218. [Google Scholar] [CrossRef]
  4. Iverson, L.R.; McKenzie, D. Tree-species range shifts in a changing climate: Detecting, modeling, assisting. Landsc. Ecol. 2013, 28, 879–889. [Google Scholar] [CrossRef]
  5. Millennium Ecosystem Assessment. Ecosystems and Human Well-Being; Island Press: Washington, DC, USA, 2005. [Google Scholar]
  6. Turner, W.; Spector, S.; Gardiner, N.; Fladeland, M.; Sterling, E.; Steininger, M. Remote sensing for biodiversity science and conservation. Trends Ecol. Evol. 2003, 18, 306–314. [Google Scholar] [CrossRef]
  7. Luo, W.; Liang, J.; Gatti, R.C.; Zhao, X.; Zhan, C. Parameterization of biodiversity–productivity relationship and its scale dependency using georeferenced tree-level data. J. Ecol. 2019, 107, 1106–1119. [Google Scholar] [CrossRef]
  8. Fundisi, E.; Musakwa, W.; Ahmed, F.B.; Tesfamichael, S.G. Estimation of woody plant species diversity during a dry season in a savanna environment using the spectral and textural information derived from WorldView-2 imagery. PLoS ONE 2020, 15, e0234158. [Google Scholar] [CrossRef]
  9. Grabska, E.; Hostert, P.; Pflugmacher, D.; Ostapowicz, K. Forest stand species mapping using the Sentinel-2 time series. J. Remote Sens. 2019, 11, 1197. [Google Scholar] [CrossRef] [Green Version]
  10. Adelabu, S.; Mutanga, O.; Adam, E.E.; Cho, M.A. Exploiting machine learning algorithms for tree species classification in a semiarid woodland using RapidEye image. J. Appl. Remote Sens. 2013, 7, e073480. [Google Scholar] [CrossRef]
  11. Ustuner, M.; Sanli, F.B.; Abdikan, S. Balanced Vs Imbalanced Training Data: Classifying Rapideye Data with Support Vector Machines. ISPRS Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2016, 41, 379–384. [Google Scholar] [CrossRef]
  12. Kavzoglu, T.; Mather, P. The use of backpropagating artificial neural networks in land cover classification. Int. J. Remote Sens. 2003, 24, 4907–4938. [Google Scholar] [CrossRef]
  13. Chrysafis, I.; Korakis, G.; Kyriazopoulos, A.P.; Mallinis, G. Predicting Tree Species Diversity Using Geodiversity and Sentinel-2 Multi-Seasonal Spectral Information. Sustainability 2020, 12, 9250. [Google Scholar] [CrossRef]
  14. Madonsela, S.; Cho, M.A.; Ramoelo, A.; Mutanga, O.; Naidoo, L. Estimating tree species diversity in the savannah using NDVI and woody canopy cover. Int. J. Appl. Earth Obs. Geoinf. 2018, 66, 106–115. [Google Scholar] [CrossRef] [Green Version]
  15. Mutowo, G.; Murwira, A. Relationship between remotely sensed variables and tree species diversity in savanna woodlands of Southern Africa. Int. J. Remote Sens. 2012, 33, 6378–6402. [Google Scholar] [CrossRef]
  16. Arekhi, M.; Yilmaz, O.Y.; Yilmaz, H.; Akyuz, Y.F. Can tree species diversity be assessed with Landsat data in a temperate forest? Environ. Monit. Assess. 2017, 189, 586. [Google Scholar] [CrossRef] [PubMed]
  17. Foody, G.M.; Cutler, M.E.J. Mapping the species richness and composition of tropical forests from remotely sensed data with neural networks. Ecol. Model. 2006, 195, 37–42. [Google Scholar] [CrossRef] [Green Version]
  18. Gillespie, T.W.; de Geode, J.; Aguilar, L.; Jenerette, G.D.; Fricker, G.A.; Avolio, M.L.; Pincetl, S.; Johnston, T.; Clarke, L.W.; Pataki, D.E. Predicting tree species richness in urban forests. Urban Ecosyst. 2016, 20, 839–849. [Google Scholar] [CrossRef]
  19. Gould, W. Remote Sensing of Vegetation, Plant Species Richness, and Regional Biodiversity Hotspots. Ecol. Appl. 2000, 10, 1861–1870. [Google Scholar] [CrossRef]
  20. Carlson, K.M.; Asner, G.P.; Hughes, R.F.; Ostertag, R.; Martin, R.E. Hyperspectral Remote Sensing of Canopy Biodiversity in Hawaiian Lowland Rainforests. Ecosystems 2007, 10, 536–549. [Google Scholar] [CrossRef]
  21. Colgan, M.S.; Baldeck, C.A.; Feret, J.B.; Asner, G.P. Mapping Savanna Tree Species at Ecosystem Scales Using Support Vector Machine Classification and BRDF Correction on Airborne Hyperspectral and LiDAR Data. Remote Sens. 2012, 4, 3462–3480. [Google Scholar] [CrossRef] [Green Version]
  22. Oldeland, J.; Wesuls, D.; Rocchini, D.; Schmidt, M.; Jürgens, N. Does using species abundance data improve estimates of species diversity from remotely sensed spectral heterogeneity? Ecol. Indic. 2010, 10, 390–396. [Google Scholar] [CrossRef]
  23. Kalacska, M.; Sanchez-Azofeifa, G.A.; Rivard, B.; Caelli, T.; White, H.P.; Calvo-Alvarado, J.C. Ecological fingerprinting of ecosystem succession: Estimating secondary tropical dry forest structure and diversity using imaging spectroscopy. Remote Sens. Environ. 2007, 108, 82–96. [Google Scholar] [CrossRef]
  24. Laurin, V.G.; Chan, J.C.-W.; Chen, Q.; Lindsell, J.A.; Coomes, D.A.; Guerriero, L.; Frate, F.D.; Miglietta, F.; Valentini, R. Biodiversity mapping in a tropical West African forest with airborne hyperspectral data. PLoS ONE 2014, 9, e97910. [Google Scholar]
  25. Schäfer, E.; Heiskanen, J.; Heikinheimo, V.; Pellikka, P. Mapping tree species diversity of a tropical montane forest by unsupervised clustering of airborne imaging spectroscopy data. Ecol. Indic. 2016, 64, 49–58. [Google Scholar] [CrossRef]
  26. Nagendra, H.; Lucas, R.; Honrado, J.P.; Jongman, R.H.G.; Tarantino, C.; Adamo, M.; Mairota, P. Remote sensing for conservation monitoring: Assessing protected areas, habitat extent, habitat condition, species diversity, and threats. Ecol. Indic. 2013, 33, 45–59. [Google Scholar] [CrossRef]
  27. Ghosh, A.; Fassnacht, F.E.; Joshi, P.K.; Koch, B. A framework for mapping tree species combining hyperspectral and LiDAR data: Role of selected classifiers and sensor across three spatial scales. Int. J. Appl. Earth Obs. Geoinf. 2014, 26, 49–63. [Google Scholar] [CrossRef]
  28. Zhao, Y.; Zeng, Y.; Zheng, Z.; Dong, W.; Zhao, D.; Wu, B.; Zhao, Q. Forest species diversity mapping using airborne LiDAR and hyperspectral data in a subtropical forest in China. Remote Sens. Environ. 2018, 213, 104–114. [Google Scholar] [CrossRef]
  29. Sun, Y.; Huang, J.; Ao, Z.; Lao, D.; Xin, Q. Deep Learning Approaches for the Mapping of Tree Species Diversity in a Tropical Wetland Using Airborne LiDAR and High-Spatial-Resolution Remote Sensing Images. Forests 2019, 10, 1047. [Google Scholar] [CrossRef] [Green Version]
  30. Madonsela, S.; Cho, M.A.; Ramoelo, A.; Mutanga, O. Remote sensing of species diversity using Landsat 8 spectral variables. ISPRS J. Photogramm. Remote Sens. 2017, 133, 116–127. [Google Scholar] [CrossRef] [Green Version]
  31. Mohammadi, J.; Shataee, S. Possibility investigation of tree diversity mapping using Landsat ETM+ data in the Hyrcanian forests of Iran. Remote Sens. Environ. 2010, 114, 1504–1512. [Google Scholar] [CrossRef]
  32. Dube, T.; Mutanga, O. Evaluating the utility of the medium-spatial resolution Landsat 8 multispectral sensor in quantifying aboveground biomass in uMgeni catchment, South Africa. ISPRS J. Photogramm. Remote Sens. 2015, 101, 36–46. [Google Scholar] [CrossRef]
  33. Feilhauer, H.; Schmidtlein, S. Mapping continuous fields of forest alpha and beta diversity. Appl. Veg. Sci. 2009, 12, 429–439. [Google Scholar] [CrossRef]
  34. Mutowo, G.; Murwira, A. The spatial prediction of tree species diversity in savanna woodlands of Southern Africa. Geocarto Int. 2012, 27, 627–645. [Google Scholar] [CrossRef]
  35. Persson, M.; Lindberg, E.; Reese, H. Tree species classification with multi-temporal Sentinel-2 data. J. Remote Sens. 2018, 10, 1794. [Google Scholar] [CrossRef] [Green Version]
  36. Mutowo, G.; Mutanga, O.; Masocha, M. Evaluating the Applications of the Near-Infrared Region in Mapping Foliar N in the Miombo Woodlands. Remote Sens. 2018, 10, 505. [Google Scholar] [CrossRef] [Green Version]
  37. Pau, S.; Gillespie, T.W.; Wolkovich, E.M. Dissecting NDVI–species richness relationships in Hawaiian dry forests. J. Biogeogr. 2012, 39, 1678–1686. [Google Scholar] [CrossRef]
  38. Oindo, B.O.; Skidmore, A.K. Interannual variability of NDVI and species richness in Kenya. Int. J. Remote Sens. 2010, 23, 285–298. [Google Scholar] [CrossRef] [Green Version]
  39. Krishnaswamy, J.; Bawa, K.S.; Ganeshaiah, K.N.; Kiran, M.C. Quantifying and mapping biodiversity and ecosystem services: Utility of a multi-season NDVI based Mahalanobis distance surrogate. Remote Sens. Environ. 2009, 113, 857–867. [Google Scholar] [CrossRef]
  40. George-Chacon, S.P.; Dupuy, J.M.; Peduzzi, A.; Hernandez-Stefanoni, L. Combining high resolution satellite imagery and lidar data to model woody species diversity of tropical dry forests. Ecol. Indic. 2019, 101, 975–984. [Google Scholar] [CrossRef]
  41. Ozdemir, I.; Norton, D.A.; Ozkan, U.Y.; Mert, A.; Senturk, O. Estimation of tree size diversity using object oriented texture analysis and aster imagery. Sensors 2008, 8, 4709–4724. [Google Scholar] [CrossRef] [Green Version]
  42. Otunga, C.; Odindi, J.; Mutanga, O.; Adjorlolo, C. Evaluating the potential of the red edge channel for C3 (Festuca spp.) grass discrimination using Sentinel-2 and Rapid Eye satellite image data. Geocarto Int. 2018, 34, 1123–1143. [Google Scholar] [CrossRef]
  43. Tigges, J.; Lakes, T.; Hostert, P. Urban vegetation classification: Benefits of multitemporal RapidEye satellite data. Remote Sens. Environ. 2013, 136, 66–75. [Google Scholar] [CrossRef]
  44. van Deventer, H.; Cho, M.A.; Mutanga, O. Improving the classification of six evergreen subtropical tree species with multi-season data from leaf spectra simulated to WorldView-2 and RapidEye. Int. J. Remote Sens. 2017, 38, 4804–4830. [Google Scholar] [CrossRef]
  45. Tuominen, S.; Näsi, R.; Honkavaara, E.; Balazs, A.; Hakala, T.; Viljanen, N.; Pölönen, I.; Saari, H.; Ojanen, H. Assessment of classifiers and remote sensing features of hyperspectral imagery and stereo-photogrammetric point clouds for recognition of tree species in a forest area of high species diversity. Remote Sens. 2018, 10, 714. [Google Scholar] [CrossRef] [Green Version]
  46. Pedro, M.S.; Rammer, W.; Seidl, R. Tree species diversity mitigates disturbance impacts on the forest carbon cycle. Oecologia 2015, 177, 619–630. [Google Scholar] [CrossRef] [PubMed]
  47. Mallinis, G.; Chrysafis, I.; Korakis, G.; Pana, E.; Kyriazopoulos, A.P. A Random Forest Modelling Procedure for a Multi-Sensor Assessment of Tree Species Diversity. Remote Sens. 2020, 12, 1210. [Google Scholar] [CrossRef] [Green Version]
  48. John, R.; Chen, J.; Lu, N.; Guo, K.; Liang, C.; Wei, Y.; Noormets, A.; Ma, K.; Han, X. Predicting plant diversity based on remote sensing products in the semi-arid region of Inner Mongolia. Remote Sens. Environ. 2008, 112, 2018–2032. [Google Scholar] [CrossRef]
  49. Ouyang, S.; Xiang, W.; Wang, X.; Zeng, Y.; Lei, P.; Deng, X.; Peng, C. Significant effects of biodiversity on forest biomass during the succession of subtropical forest in south China. For. Ecol. Manag. 2016, 372, 291–302. [Google Scholar] [CrossRef]
  50. Huang, Y.; Ma, Y.; Zhao, K.; Niklaus, P.A.; Schmid, B.; He, J.-S. Positive effects of tree species diversity on litterfall quantity and quality along a secondary successional chronosequence in a subtropical forest. J. Plant Ecol. 2017, 10, 28–35. [Google Scholar] [CrossRef] [Green Version]
  51. Sun, Z.; Liu, X.; Schmid, B.; Bruelheide, H.; Bu, W.; Ma, K. Positive effects of tree species richness on fine-root production in a subtropical forest in SE-China. J. Plant Ecol. 2017, 10, 146–157. [Google Scholar] [CrossRef] [Green Version]
  52. Ezemvelo KwaZulu-Natal Wildlife. Nkandla Forest Complex: Protected Area Management Plan. 2015, pp. 1–156. Available online: www.kznwildlife.com/Documents/Nkandla%20Forest%20Complex%20Management%20Plan%20Final%2012102016%20compressed.pdf (accessed on 1 January 2021).
  53. Gyamfi-Ampadu, E.; Gebreslasie, M.; Mendoza-Ponce, A. Mapping natural forest cover using satellite imagery of Nkandla forest reserve, KwaZulu-Natal, South Africa. Remote Sens. Appl. Soc. Environ. 2020, 18, 100302. [Google Scholar] [CrossRef]
  54. Shannon, C. A mathematical theory of communication. J. Bell Syst. Tech. J. 1948, 27, 379–423. [Google Scholar] [CrossRef] [Green Version]
  55. Simpson, E. Measurement of diversity. J. Nat. 1949, 163, 688. [Google Scholar] [CrossRef]
  56. Morris, E.K.; Caruso, T.; Buscot, F.; Fischer, M.; Hancock, C.; Maier, T.S.; Meiners, T.; Müller, C.; Obermaier, E.; Prati, D. Choosing and using diversity indices: Insights for ecological applications from the German Biodiversity Exploratories. Ecol. Evol. 2014, 4, 3514–3524. [Google Scholar] [CrossRef] [Green Version]
  57. Daly, A.J.; Baetens, J.M.; de Baets, B.J.M. Ecological diversity: Measuring the unmeasurable. Mathematics 2018, 6, 119. [Google Scholar] [CrossRef] [Green Version]
  58. Ifo, S.A.; Moutsambote, J.-M.; Koubouana, F.; Yoka, J.; Ndzai, S.F.; Bouetou-Kadilamio, L.N.O.; Mampouya, H.; Jourdain, C.; Bocko, Y.; Mantota, A.B. Tree species diversity, richness, and similarity in intact and degraded forest in the tropical rainforest of the Congo Basin: Case of the forest of Likouala in the Republic of Congo. Int. J. For. Res. 2016, 2016, 1–12. [Google Scholar] [CrossRef] [Green Version]
  59. Ghosh, A.; Joshi, P.K. A comparison of selected classification algorithms for mapping bamboo patches in lower Gangetic plains using very high resolution WorldView 2 imagery. Int. J. Appl. Earth Obs. Geoinf. 2014, 26, 298–311. [Google Scholar] [CrossRef]
  60. Wang, D.; Wan, B.; Qiu, P.; Su, Y.; Guo, Q.; Wang, R.; Sun, F.; Wu, X. Evaluating the Performance of Sentinel-2, Landsat 8 and Pléiades-1 in Mapping Mangrove Extent and Species. Remote Sens. 2018, 10, 1468. [Google Scholar] [CrossRef] [Green Version]
  61. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef] [Green Version]
  62. Rodríguez-Galiano, V.F.; Abarca-Hernández, F.; Ghimire, B.; Chica-Olmo, M.; Atkinson, P.M.; Jeganathan, C. Incorporating Spatial Variability Measures in Land-cover Classification using Random Forest. Procedia Environ. Sci. 2011, 3, 44–49. [Google Scholar] [CrossRef] [Green Version]
  63. Abdel-Rahman, E.M.; Ahmed, F.B.; Ismail, R. Random forest regression and spectral band selection for estimating sugarcane leaf nitrogen concentration using EO-1 Hyperion hyperspectral data. Int. J. Remote 2013, 34, 712–728. [Google Scholar] [CrossRef]
  64. Ramoelo, A.; Cho, M.A.; Mathieu, R.; Madonsela, S.; Van De Kerchove, R.; Kaszta, Z.; Wolff, E. Monitoring grass nutrients and biomass as indicators of rangeland quality and quantity using random forest modelling and WorldView-2 data. Int. J. Appl. Earth Obs. Geoinf. 2015, 43, 43–54. [Google Scholar] [CrossRef]
  65. Duro, D.C.; Franklin, S.E.; Dubé, M.G. A comparison of pixel-based and object-based image analysis with selected machine learning algorithms for the classification of agricultural landscapes using SPOT-5 HRG imagery. Remote Sens. Environ. 2012, 118, 259–272. [Google Scholar] [CrossRef]
  66. Nitze, I.; Barrett, B.; Cawkwell, F. Temporal optimisation of image acquisition for land cover classification with Random Forest and MODIS time-series. Int. J. Appl. Earth Obs. Geoinf. 2015, 34, 136–146. [Google Scholar] [CrossRef] [Green Version]
  67. Liaw, A.; Wiener, M. Classification and regression by randomForest. R News 2002, 2, 18–22. [Google Scholar]
  68. R Core Team. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing: Vienna, Austria, 2013. [Google Scholar]
  69. Ganivet, E.; Bloomberg, M. Towards rapid assessments of tree species diversity and structure in fragmented tropical forests: A review of perspectives offered by remotely-sensed and field-based data. Forest Ecol. Manag. 2019, 432, 40–53. [Google Scholar] [CrossRef]
  70. Fassnacht, F.E.; Hartig, F.; Latifi, H.; Berger, C.; Hernández, J.; Corvalán, P.; Koch, B. Importance of sample size, data type and prediction method for remote sensing-based estimations of aboveground forest biomass. Remote Sens. Environ. 2014, 154, 102–114. [Google Scholar] [CrossRef]
  71. Lu, D. The potential and challenge of remote sensing-based biomass estimation. Int. J. Remote Sens. 2006, 27, 1297–1328. [Google Scholar] [CrossRef]
  72. Todd, S.W.; Hoffer, R.M.; Milchunas, D.G. Biomass estimation on grazed and ungrazed rangelands using spectral indices. Int. J. Remote Sens. 1998, 19, 427–438. [Google Scholar] [CrossRef]
  73. Mutanga, O.; Skidmore, A.K. Narrow band vegetation indices overcome the saturation problem in biomass estimation. Int. J. Remote Sens. 2004, 25, 3999–4014. [Google Scholar] [CrossRef]
  74. Rajah, P.; Odindi, J.; Mutanga, O.; Kiala, Z. The utility of Sentinel-2 Vegetation Indices (VIs) and Sentinel-1 Synthetic Aperture Radar (SAR) for invasive alien species detection and mapping. Nat. Conserv. 2019, 35, 41–61. [Google Scholar] [CrossRef]
  75. Rocchini, D.; Ricotta, C.; Chiarucci, A. Using satellite imagery to assess plant species richness: The role of multispectral systems. J. Appl. Veg. Sci. 2007, 10, 325–331. [Google Scholar] [CrossRef]
  76. Wang, R.; Gamon, J.A.; Schweiger, A.K.; Cavender-Bares, J.; Townsend, P.A.; Zygielbaum, A.I.; Kothari, S. Influence of species richness, evenness, and composition on optical diversity: A simulation study. J. Remote Sens. Environ. 2018, 211, 218–228. [Google Scholar] [CrossRef]
  77. Immitzer, M.; Neuwirth, M.; Böck, S.; Brenner, H.; Vuolo, F.; Atzberger, C. Optimal Input Features for Tree Species Classification in Central Europe Based on Multi-Temporal Sentinel-2 Data. J. Remote Sens. 2019, 11, 2599. [Google Scholar] [CrossRef] [Green Version]
  78. Martin-Gallego, P.; Aplin, P.; Marston, C.; Altamirano, A.; Pauchard, A. Detecting and modelling alien tree presence using Sentinel-2 satellite imagery in Chile’s temperate forests. For. Ecol. Manag. 2020, 474, 118353. [Google Scholar] [CrossRef]
  79. Millard, K.; Richardson, M. On the Importance of Training Data Sample Selection in Random Forest Image Classification: A Case Study in Peatland Ecosystem Mapping. Remote Sens. 2015, 7, 8489–8515. [Google Scholar] [CrossRef] [Green Version]
  80. Bolyn, C.; Michez, A.; Gaucher, P.; Lejeune, P.; Bonnet, S. Forest mapping and species composition using supervised per pixel classification of Sentinel-2 imagery. Biotechnol. Agron. Société Environ. 2018, 22, 16. [Google Scholar]
  81. Parson, E.A. Climate Engineering in Global Climate Governance: Implications for Participation and Linkage. Transnatl. Environ. Law 2013, 3, 89–110. [Google Scholar] [CrossRef] [Green Version]
  82. Nagendra, H. Using remote sensing to assess biodiversity. Int. J. Remote Sens. 2001, 22, 2377–2400. [Google Scholar] [CrossRef]
  83. Dube, T.; Mutanga, O.; Elhadi, A.; Ismail, R. Intra-and-inter species biomass prediction in a plantation forest: Testing the utility of high spatial resolution spaceborne multispectral RapidEye sensor and advanced machine learning algorithms. Sensors 2014, 14, 15348–15370. [Google Scholar] [CrossRef] [Green Version]
  84. Wallner, A.; Elatawneh, A.; Schneider, T.; Knoke, T. Estimation of forest structural information using RapidEye satellite data. Forestry 2014, 88, 96–107. [Google Scholar] [CrossRef] [Green Version]
  85. Maeda, E.E.; Heiskanen, J.; Thijs, K.W.; Pellikka, P.K.E. Season-dependence of remote sensing indicators of tree species diversity. Remote Sens. Lett. 2014, 5, 404–412. [Google Scholar] [CrossRef]
  86. El-Askary, H.; Abd El-Mawla, S.; Li, J.; El-Hattab, M.; El-Raey, M. Change detection of coral reef habitat using Landsat-5 TM, Landsat 7 ETM+ and Landsat 8 OLI data in the Red Sea (Hurghada, Egypt). Int. J. Remote Sens. 2014, 35, 2327–2346. [Google Scholar] [CrossRef]
  87. Pahlevan, N.; Lee, Z.; Wei, J.; Schaaf, C.B.; Schott, J.R.; Berk, A. On-orbit radiometric characterization of OLI (Landsat-8) for applications in aquatic remote sensing. Remote Sens. Environ. 2014, 154, 272–284. [Google Scholar] [CrossRef]
  88. Belgiu, M.; Drăguţ, L. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  89. Fassnacht, F.E.; Latifi, H.; Stereńczak, K.; Modzelewska, A.; Lefsky, M.; Waser, L.T.; Straub, C.; Ghosh, A. Review of studies on tree species classification from remotely sensed data. Remote Sens. Environ. 2016, 186, 64–87. [Google Scholar] [CrossRef]
Figure 1. Map of the study area. Note: A is the Nkandla forest reserve, B is the map of South Africa indicating the location of KZN province and the forest and C is the map of Africa indicating the location of South Africa.
Figure 1. Map of the study area. Note: A is the Nkandla forest reserve, B is the map of South Africa indicating the location of KZN province and the forest and C is the map of Africa indicating the location of South Africa.
Remotesensing 13 01033 g001
Figure 2. Scatter plot for the Shannon Index prediction. (A) is for Sentinel 2, (B) is for RapidEye, (C) is for PlanetScope, and (D) is for Landsat 8. The blue line is the line of best fit and the dashed line is the 1:1 line as shown on the individual plots.
Figure 2. Scatter plot for the Shannon Index prediction. (A) is for Sentinel 2, (B) is for RapidEye, (C) is for PlanetScope, and (D) is for Landsat 8. The blue line is the line of best fit and the dashed line is the 1:1 line as shown on the individual plots.
Remotesensing 13 01033 g002
Figure 3. Scatter plot for the Simpson Index prediction. (A) is for Sentinel 2, (B) is for RapidEye, (C) is for PlanetScope, and (D) is for Landsat 8. The blue line is the line of best fit and the dashed line is the 1:1 line as shown on the individual plots.
Figure 3. Scatter plot for the Simpson Index prediction. (A) is for Sentinel 2, (B) is for RapidEye, (C) is for PlanetScope, and (D) is for Landsat 8. The blue line is the line of best fit and the dashed line is the 1:1 line as shown on the individual plots.
Remotesensing 13 01033 g003
Figure 4. Scatter plot for the Species richness predictions. (A) is for Sentinel 2, (B) is for RapidEye, (C) is for PlanetScope, and (D) is for Landsat 8. The blue line is the line of best fit and the dashed line is the 1:1 line as shown on the individual plots.
Figure 4. Scatter plot for the Species richness predictions. (A) is for Sentinel 2, (B) is for RapidEye, (C) is for PlanetScope, and (D) is for Landsat 8. The blue line is the line of best fit and the dashed line is the 1:1 line as shown on the individual plots.
Remotesensing 13 01033 g004
Table 1. Details of the spectral and spatial resolution of satellite imageries.
Table 1. Details of the spectral and spatial resolution of satellite imageries.
Sentinel 2Landsat 8RapidEyePlanetScope
BandsBandwidth
(nm)
Spatial Resolution (m)BandsBandwidthSpatial Resolution (m)BandsBandwidth
(nm)
Spatial Resolution (m)BandsBandwidth
(nm)
Spatial Resolution (m)
Blue458-52310Blue452-51230Blue440-5105Blue455-5153
Green543-57310Green533-59030Green520-5905Green500-5903
Red650-68010Red636-67330Red630-6855Red590-6703
RE1698-71320NIR851-87930RE690-7305NIR780-8603
RE2733-74820SWIR11566-165130NIR760-8505
RE3773-79320SWIR22107-229430
NIR785-89910
NNIR855-87520
SWIR21565-165520
SWIR22100-228020
Note: RE: Red edge; NIR: Near infrared; NNIR: Narrow Near infrared; SWIR: Shortwave infrared.
Table 2. Descriptive statistics of Shannon Index (H′), Simpson Index (D1) and Species richness (S) produced from the field inventory data.
Table 2. Descriptive statistics of Shannon Index (H′), Simpson Index (D1) and Species richness (S) produced from the field inventory data.
ParameterShannon IndexSimpson IndexSpecies Richness
Mean2.0550.8919
Minimum0.9490.1554
Maximum2.7180.99315
Standard Deviation0.2900.0682.47
Table 3. RF sensor model accuracies for the Sentinel 2, RapidEye, PlanetScope and the Landsat 8 for Shannon Index, Simpson Index and the Species Richness.
Table 3. RF sensor model accuracies for the Sentinel 2, RapidEye, PlanetScope and the Landsat 8 for Shannon Index, Simpson Index and the Species Richness.
ImageShannon
Index
Simpson
Index
Species
Richness
R2RMSEp ValueR2RMSEp ValueR2RMSEp Value
Sentinel 20.9260.148<2.2 × 10 16 0.9170.043<2.2 × 10 16 0.9231.183<2.2 × 10 16
RapidEye0.9020.147<2.2 × 10 16 0.9150.044<2.2 × 10 16 0.8331.287<2.2 × 10 16
PlanetScope0.8980.156<2.2 × 10 16 0.8990.045<2.2 × 10 16 0.9001.293<2.2 × 10 16
Landsat 80.5291.748<2.2 × 10 16 0.4100.063<2.2 × 10 16 0.5321.746<2.2 × 10 16
Table 4. The statistical analysis of the prediction made with the RF for each of the images under the Shannon Index, Simpson Index and the Species richness.
Table 4. The statistical analysis of the prediction made with the RF for each of the images under the Shannon Index, Simpson Index and the Species richness.
Satellite
Image
ParameterShannon
Index
Simpson
Index
Species
Richness
Sentinel 2Mean2.050.899.24
Minimum1.510.556.58
Maximum2.340.9512.7
Standard deviation 0.150.041.33
RapidEyeMean2.050.899.22
Minimum1.400.535.73
Maximum2.440.9413.48
Standard deviation 0.180.041.41
PlanetScopeMean2.050.899.26
Minimum1.610.566.25
Maximum2.440.9512.67
Standard deviation 0.150.041.30
Landsat 8Mean2.060.899.04
Minimum1.730.736.13
Maximum2.420.9513.37
Standard deviation 0.150.041.41
Table 5. Variable importance ranking for the Sentinel 2 image model under the Shannon Index, Simpson Index and the Species richness.
Table 5. Variable importance ranking for the Sentinel 2 image model under the Shannon Index, Simpson Index and the Species richness.
Shannon IndexSimpson IndexSpecies Richness
Band%IncMSEBand%IncMSEBand%IncMSE
SWIR117.75RE26.71SWIR118.57
SWIR214.11SWIR16.05NNIR16.19
RE213.52SWIR25.43SWIR213.27
NNIR8.97NNIR4.50RE210.88
NIR4.89NIR−0.03NIR6.99
Table 6. Variable importance ranking for the RapidEye image model under the Shannon Index, Simpson Index and the Species richness.
Table 6. Variable importance ranking for the RapidEye image model under the Shannon Index, Simpson Index and the Species richness.
Shannon IndexSimpson IndexSpecies Richness
Band%IncMSEBand%IncMSEBand%IncMSE
Red19.64NIR6.28NIR12.58
NIR14.24RE4.79RE9.25
RE14.01Red1.91Red8.39
Table 7. Variable importance ranking for the PlanetScope image model under the Shannon Index, Simpson Index and the Species richness.
Table 7. Variable importance ranking for the PlanetScope image model under the Shannon Index, Simpson Index and the Species richness.
Shannon IndexSimpson IndexSpecies Richness
Band%IncMSEBand%IncMSEBand%IncMSE
Green9.93Green6.03NIR14.54
Red9.91NIR5.83Green13.46
Blue7.78Blue5.44Red12.78
NIR7.04Red2.01Blue10.11
Table 8. Variable importance ranking for the Landsat 8 image model under the Shannon Index, Simpson Index and the Species Richness.
Table 8. Variable importance ranking for the Landsat 8 image model under the Shannon Index, Simpson Index and the Species Richness.
Shannon IndexSimpson IndexSpecies Richness
Band%IncMSEBand%IncMSEBand%IncMSE
SWIR118.01NIR8.36SWIR118.96
NIR15.38SWIR17.87NIR16.71
Green14.16Red5.11Red13.91
Red11.96Green3.77Green13.80
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Gyamfi-Ampadu, E.; Gebreslasie, M.; Mendoza-Ponce, A. Evaluating Multi-Sensors Spectral and Spatial Resolutions for Tree Species Diversity Prediction. Remote Sens. 2021, 13, 1033. https://doi.org/10.3390/rs13051033

AMA Style

Gyamfi-Ampadu E, Gebreslasie M, Mendoza-Ponce A. Evaluating Multi-Sensors Spectral and Spatial Resolutions for Tree Species Diversity Prediction. Remote Sensing. 2021; 13(5):1033. https://doi.org/10.3390/rs13051033

Chicago/Turabian Style

Gyamfi-Ampadu, Enoch, Michael Gebreslasie, and Alma Mendoza-Ponce. 2021. "Evaluating Multi-Sensors Spectral and Spatial Resolutions for Tree Species Diversity Prediction" Remote Sensing 13, no. 5: 1033. https://doi.org/10.3390/rs13051033

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop