Sage Journals: Discover world-class research

Abstract

The impacts of climate on forestry will influence the tree species planted for commercial use and the ability to adapt to drastic changes in the environment. These changes have promoted the use of technologies such as remote sensing, to monitor compartments, to aid management in decision-making and planning. The use of time series datasets can aid in the understanding and prediction of compartment functioning and health. This study tested the use of Landsat 8 (spectral range 435–2294 nm) level 2 surface reflectance data and climatic variables (such as solar radiation and vapour pressure deficit) to predict tree sap flow measurements over an 8-year period. Using RapidMiner studio for statistical analysis, gradient boosting trees (GBTs) proved successful in achieving an R² of 0.97 and an RMSE of 2.39 LPD. It is noted that of all the species that were tested (i.e. Pinus radiata, P. elliotii, Ocotea bullata, Ilex mitis, Podocarpus latifolius, P. elliotii, Eucalyptus grandis x E. nitens and E. dunnii), the RMSE was greatly reduced with the exclusion of the pine tree species, to a reduced error rate of 0.78. This suggests that using satellite data, for modelling tree sap flow of pine trees has more variation. Nevertheless, this study has proved successful in forecasting sap flow and will contribute to future research in forest health monitoring, as well as to refine the use of time series information to predict sap flow velocity in forest plantation trees.

Keywords

sap flow velocity time series analysis forest health prediction rapidminer landsat

1. Introduction

Climate has a significant impact on the distribution and species used for commercial forestry (Drake et al., 2016; Kirilenko and Sedjo, 2007). Commercial forestry plantations in South Africa are confined to areas of high rainfall, to sustain long-term growth and vigour (Ferreira et al., 2023; Kaptein et al., 2023; Van der Merwe et al., 2022). The various challenges of environmental influences and competition with other industries make site-species matching a priority. However, fluctuations in climatic conditions may lead to unforeseen risks such as drought, hail and pests and diseases (Baldi and La Porta, 2020; Lehmann et al., 2015; Urban et al., 2017; Van der Merwe et al., 2022). To adapt to this, companies rely on tree breeding strategies to bolster the resilience of trees to changes in climate and the introduction of pests and diseases (Matthews et al., 2018; Solh and Van Ginkel, 2014; Warburton and Schulze 2008). Several studies have analysed the impact of climate change on compartment health (Czimber and Galos, 2016; Keskitalo et al., 2016; Schoene and Bernier, 2012). The effects of this may lead to reduced yields, stunted growth and even mortality, greatly reducing the likelihood of a return on investment from commercial tree stands (Ferreira et al., 2022; Mahlangu et al., 2018).

These impacts often result in increased frequency of extreme events such as variability in rainfall, which leads to increased intensity of droughts and floods (Van der Merwe et al., 2024). Furthermore, an increased frequency of extreme environmental conditions may have an adverse effect on the productivity of forest stands, delaying reafforestation or destroying long-term investments in the form of stunted growth and severe mortality (Engelbrecht and Monteiro, 2021; Van der Merwe et al., 2024). This delay in recovery is due to a reduction in leaf water potential and stomatal conductance, caused by environmental stress (Engelbrecht and Monteiro, 2021; Skelton and Brodribb, 2018), and can be predicted by water transport properties of trees (Skelton and Brodribb, 2018).

To improve the information gathered from forest stands, applications of remote sensing technologies are implemented, to assist in decision-making (Dragavtsev and Nartov, 2015; Ferreira et al., 2022). However, the variables selected to monitor forest health have been driven by popularity and ease of access (Ferreira et al., 2023; Xulu et al., 2019). Sap flow can provide valuable information on compartments, such as the forest water status, as monitoring is almost continuous, capturing fluctuations in gas exchange in real-time (Skelton and Brodribb, 2018). For example, a study by Børja et al. (2015), succeeded in detecting a reduction of sap flux density in stressed trees using the heat field deformations method. Furthermore, they found that this reduction was present for a short period prior to the visual appearance of stress symptoms and thus could potentially add value as part of health diagnostics tool. This information of water status and gas exchange is directly influenced by environmental variables and can be used as an indicator for stress, such as identifying the presence of adverse abiotic conditions (Mapeto et al., 2018; Skelton and Brodribb, 2018). However, the equipment is expensive and localized, and difficult to replicate across an entire forest stand. Thus, a strategic placement of sap flow measuring devices across compartments may provide valuable continuous information, to aid in decision-making that benefits forest longevity (Mapeto et al., 2018; Skelton and Brodribb, 2018).

Eco-physiological studies by Mapeto et al. (2018) and Kaptein et al. (2022), which were designed to improve our understanding of site-tree interactions, resulted in the development of significant functional relationships between environmental variables and tree water-use. Mapeto et al. (2018) used soil moisture, solar radiation and wind velocity to predict sap flow in trees with a high degree of accuracy and confidence. These relationships were in line with our understanding of plant physiological relationships (Kaptein et al., 2022; Mapeto et al., 2018).

Fundamentally, monitoring compartments is crucial to ensure long-term sustainability (Baldi and La Porta, 2020; Ferreira et al., 2023; Huang and Asner, 2009; Xulu et al., 2019). However, field surveys are tedious, as depending on the tree species and stress tolerance, it may be some time before stress symptoms become visually identifiable (Das et al., 2021; Huo et al., 2021; Pullanagari et al., 2016). Nevertheless, the use of remote sensing has provided a means to capture stress at a physiological level, through use of vegetation indices and has eased monitoring of large areas of forestry compartments (Avtar et al., 2020; Carrilo et al., 2020; Dash et al., 2017; Lechner et al., 2020; Peerbhay et al., 2022). Furthermore, the robustness of risk identification is improved with the aid of machine learning algorithms, to identify correlations between variables measured (Ferreira et al., 2023; Hamdi et al., 2019). Despite these contributions to our understanding of compartment health monitoring, few studies have attempted to use remote sensing data to predict sap flow velocity to indicate forest condition and using climatic data (Hasimoto et al., 2023; Nagler et al., 2007).

Predicting forest health is a helpful tool in decision-making, by allowing management to alter protocols and adapt to changes in forest health to improve the longevity of forest stands (McCullagh and Nieuwenhuis, 2015; Ovchynikova and Dupliak, 2021). Furthermore, this has provided a means to estimate timber yields, supplementing information for sustainable forest management (McCullagh and Nieuwenhuis, 2015), as well as acquiring valuable information on forest ecosystems. Several studies have succeeded in predicting damage occurrences in forest stands. For instance, Huang et al., 2021, provides a good framework for time series analysis, succeeding in predicting Apocheima cinerarius Erschoff infestations in Populus euphratica, using MODIS NDVI data (R² = 0.92–0.94) to identify damage severity classes and tracking outbreak times (Huang et al., 2021). Furthermore, Hasimoto et al. (2023), succeeded in estimating canopy transpiration (R² = 0.76–0.89), using satellite remote sensing and in situ sap flow observations, which aids in the development of models that estimate transpiration (Hasimoto et al., 2023). Through the advancement of technology, the quality of information has improved, providing access to predictions of increased severity of abiotic risks, to aid in planning (McCullagh and Nieuwenhuis, 2015; Wintle and Lindenmayer, 2008). Thus, the use of sap flow for predicting health may provide a framework for forest health monitoring.

There exists a need to monitor large forest landscapes in an efficient manner, by estimating the trends in health change over time. This aids in the prediction of forest health, which improves decision-making and management practices. The primary research questions are as follows: (1) To what extent does sap flow rates differ between indigenous species and exotic species? (2) How efficient is Landsat 8 data at predicting sap flow rates with added climate data? (3) What are the most important sensor and environmental variables for predicting sap flow in commercial and indigenous forest stands?

2. Material and methods

2.1. Study area

Data from two sets of sap flow velocity (SV) experiments were selected, which were spread across South Africa (Figure 1). Firstly, a plantation in George (33°56′35″S, 22°30′41″E), Western Cape and secondly, plantations in the Zululand area of KwaZulu-Natal (28°36′06″S, 32°11′22″E). The sites in the George area were aged 13 years, with species Pinus radiata, P. elliotii, Ocotea bullata, Ilex mitis and Podocarpus latifolius (Mapeto et al., 2018). The site in the KwaZulu-Natal province and was aged 8–20 years, with species of P. elliotii, Eucalyptus grandis x E. nitens and E. dunnii (Kaptein et al., 2022).

Figure 1.

Map of HPV study areas of several compartments located in the KwaZulu-Natal region and Western Cape region of South Africa, where George compartments are represented by the red diamond, the Zululand compartments by orange circle and compartment forestry.

2.2. Field data collection and climatic variables

SV sensors were installed prior to collecting data over the period of 2013–2021. The pine and eucalypt species are widely used across South Africa due to their superior wood quality and fast growth. Climatic variables collected at the sites were solar radiation, rainfall, transpiration, temperature, relative humidity (RH) and vapour pressure deficit (VPD) in conjunction with SV. The variables were collected with the aid of weather stations placed in the areas of the sites.

The weather data variables that were collected from the study sites were rainfall (through TE525-L, Texas Electronics, Dallas, TX, USA), solar radiation (through LI-200SA, LI-COR Inc., Lincoln, NE, USA), temperature and humidity (HMP50, Vaisala, Helsinki, Finland) as well as wind speed through (Model 03,001, RM Young, Traverse City, MI, USA) (Mapeto et al., 2018).

Along with these variables, soil moisture content was measured for trees hourly, at a depth of 30 cm (through CS616, Campbell Scientific) and logged (through Campbell Scientific CR1000) (Mapeto et al., 2018).

2.3. Sap flow velocity data acquisition

Sap flow data used in the study was acquired from past studies of Mapeto et al. (2018) and Kaptein et al. (2022). The objectives of the study by Mapeto et al. (2018), were to quantify and characterize the influence of environmental conditions that occurred locally, on the whole tree sap flow of individual species of Pinus and indigenous trees (Mapeto et al., 2018). The objectives from the study of Kaptein et al. (2022), were to compare the transpiration and total evaporation rates, between Eucalyptus and Pinus species, as well as to assess the impacts of both species on the plantation water yield (Kaptein et al., 2022). Data on sap flow was acquired through the heat ratio method of heat-pulse velocity (Burgess et al., 2001; Kaptein et al., 2022; Mapeto et al., 2018). Of these data sets, daily sap flow data was extracted. These days were matched to the days Landsat data images were available for the year.

2.4. Landsat images and pre-processing

Sap flow data GPS points were used to extract data from corresponding Landsat images. Landsat 8 images were acquired from United States Geological Survey in June 2023 (USGS, https://earthexplorer.usgs.gov/), where images were selected based on availability, less than 10% cloud cover and matched to sap flow timelines. Only one image tile was needed to cover the study area of the George region, and one other to cover the study area of the KwaZulu-Natal province. Landsat 8 has seven spectral bands, at a resolution of 30 m, which were incorporated in the study, namely, Coastal Aerosol (band 1, 0.43–0.45), Blue (band 2, 0.45–0.51), Green (band 3, 0.53–0.59), Red (band 4, 0.64–0.67), Near-Infrared (band 5, 0.85–0.88), SWIR 1 (band 6, 1.57–1.65) and SWIR 2 (band 7, 2.11–2.29). A surface reflectance image was created of the study area in QGIS, using Landsat 8 level 2 data, to reduce the impacts of the atmosphere on reflectance values from the compartments measured. Furthermore, spectral bands were stacked from the acquired images using QGIS version 3.36 and extracted for calculation of indices that were used in data analysis.

Table 1 illustrates corresponding Landsat images available, with sap flow velocity data acquisition timelines. On days Landsat images were unavailable, sap flow velocity data was subsequently removed from analysis. Monitoring of the KwaZulu-Natal province sites started in late 2019.

Table 1.

Landsat images available in correspondence to sap flow data.

Start date	End date	Landsat image availability
24/4/2013	4/12/2013	8
21/1/2014	6/5/2014	5
24/11/2015	3/12/2015	2
20/1/2016	5/12/2016	8
22/1/2017	31/12/2017	12
2/12/2019	18/12/2019	2
3/1/2020	20/12/2020	11
5/1/2021	20/10/2021	9

A preliminary test in RapidMiner was conducted testing several indices available to Landsat 8, to determine which indices would potentially correlate well with sap flow, and thus those were calculated using Landsat bands (Table 2). These indices were selected, due to their ability to detect water stress and moisture levels in vegetation. The environmental variables of water stress and moisture content have been proven to correlate well with sap flow (Mapeto et al., 2018).

Table 2.

Vegetation indices used for data analysis.

Vegetation indices	Acronym	Equation	Reference
Advanced Vegetation Index	AVI	${[N I R * (1 - R) * (N I R - R)]}^{1 / 3}$	Gabron et al. (2000)
Normalized Difference Moisture Index	NDMI	$\frac{N I R - S W I R}{N I R + S W I R}$	Gao (1996)
Normalized Difference Vegetation Index	NDVI	$\frac{N I R - R}{N I R + R}$	Rouse et al. (1973)
Normalized Difference Water Index	NDWI	$\frac{G - N I R}{G + N I R}$	Gao (1996)
Normalized Pigment Chlorophyll Ratio Index	NPCRI	$\frac{R - B}{R + B}$	Merzlyak et al. (1999)

2.5. RapidMiner

Due to the streamlined process of selecting data and managing inputs, RapidMiner was selected for data analysis (Figure 2). Through a preliminary test of the performance of the available algorithms, gradient boosting trees were selected to help visualize the data set and perform cross validation.

Figure 2.

Model design for predicting sap flow velocity data in RapidMiner, through Process Windows operator.

The number of trees selected for GBT was 500, with a maximum depth of 20. The process of the model contains several tasks that are carried out in sequence, such as fetching data, sorting, time series and cross validation. After this process, the model performance is evaluated (Ho et al., 2019), and data can be visualized to view correlations.

2.5.1. Time series method

Time series of sap flow data was performed in RapidMiner, through the Process Windows operator. This operator takes the input time series and creates windows, and from the subprocesses for each of the windows, it loops. The dates of measurements taken were used as the indices attribute, the horizon attribute was set to sap flow, and the window size (number of values in one window) was set to a value 7, other variables were left to default. The next operator parameters were differentiated, which takes the time series attributes and calculates the differentiated values. The attributes selected were set to a subset for variables other than sap flow. The lag was set to the default of 1 with the differentiation method of subtraction. Next the differentiated values were put through the aggregate operator, selecting the mean and stand deviation as attributes. Finally, values were pushed through a GBT model for accuracy assessment of prediction in the cross-validation operator.

2.5.2. Gradient boosting trees

Through the process of training weak learners, gradient boosting trees (GBTs) are produced resulting in predictions that provide analysable data. To increase the robustness of the model, many shallow decision trees, or weaker learners, are created to improve the overall accuracy and minimize error (Callens et al., 2020; Chen and Guestrin, 2016). Boosting is considered popular, due to its ability to perform well with smaller datasets. Furthermore, the training speed of models is fast, which adds to its popularity (Chen and Guestrin, 2016; Joharestani et al., 2019; Zhang et al., 2019). The first equation of GBT follows with equation (1) (Guillen et al., 2023; Hastie et al., 2009):

f_{0} (x) = \begin{array}{c} argmin \\ γ \end{array} \sum_{j = 1}^{n} L (y_{i j}, γ)

(1)

To minimize the mean square error (MSE), the value $γ$ is included in all observations, as well as the specific expected value of loss function $L (Y_{i j}, y)$ (Guillen et al., 2023; Hastie et al., 2009). The number of interactions was calculated for $i = 1, 2, \dots, N$ , in equation (2) at (a) (Guillen et al., 2023; Hastie et al., 2009):

(a) $r_{i m} = - {[\frac{\partial L (y_{i}, f (x_{i}))}{\partial f (x_{i})}]}_{f = f m - 1}$

(b) $γ_{j m} = \begin{array}{c} argmin \\ γ \end{array} \sum_{x_{i} \in R_{j m}} L (y_{i}, f_{m} - 1 (x_{i}) + γ)$

(d) \hat{f} (x) = f_{m} (x)

(2)

The regression tree is fitted to the target $r_{i m}$ , through terminal regions $R_{j m}, j = 1, 2, \dots, J_{m}$ , at (b). Finally, the resulting (d) is connected $K$ different tree expansions, $f_{k M} (x), k = 1, 2, \dots, K$ (Guillen et al., 2023; Hastie et al., 2009).

3. Accuracy assessment

Prior to the use of RapidMiner, a sap flow data sheet containing all environmental variables, sap flow, dates and species was compiled in excel, in preparation for data analysis. Through use of the cross-validation operator in RapidMiner, data was divided into 10 subsets, from which 9 of those subsets were used to build a model (training), using 1 as a test subset, to measure the performance of each iteration. There were 100 iterations made from these datasets. Thereafter the performance of the model was produced to compare accuracy of R square and mean error squared in sap flow units, establishing the range in which sap flow can be predicted. This was also done to evaluate the performance per species per site as well as an overall statistical analysis.

4. Results

4.1. Descriptive statistics of all species and the corresponding variables measured at yearly intervals

The descriptive statistics for pine species, Pinus radiata and P. elliotii, and their corresponding variables are presented in Table 3. The highest mean sap flow of 53.81 LPD occurred in the years 2013–2014. The mean sap flow for pine species has reduced over the years, reaching its lowest value of 2.69 LPD in 2020–2021. Solar radiation mean readings were associated with sap flow undulate over the years, starting at 19.65 MJ in the years 2013–2014 and reaching 19.32 MJ when sap flow was at its lowest point in 2020–2021. VPD and RH also undulates, with a high mean of 2.15 kPa in 2013–2014, and a low mean of 1.22 kPa in 2020–2021 for VPD, and RH reaching a mean of 60.38% in 2013–2014 and 62.07% in 2020–2021.

Table 3.

Descriptive statistics of Pinus species and corresponding environmental variables.

Date	Parameter	Mean	Minimum	Maximum	Standard deviation
2013–2014	Sap flow (L/day)	53.81	21.94	81.34	20.91
	Solrad (MJ/day)	19.65	11.96	29.48	5.97
	VPD (kPa/day)	2.15	1.22	4.19	0.88
	RH (%/ day)	60.38	36.06	84.69	12.68
2015–2016	Sap flow (L/day)	24.88	7.45	62.79	13.65
	Solrad (MJ/day)	22.44	9.85	30.94	8.67
	VPD (kPa/day)	1.84	1.14	3.09	0.67
	RH (%/ day)	67.29	43.63	90.94	7.85
2017–2019	Sap flow (L/day)	13.78	4.12	32.55	6.72
	Solrad (MJ/day)	20.51	10.45	31.70	7.54
	VPD (kPa/day)	2.61	1.16	5.18	1.14
	RH (%/ day)	58.93	31.93	85.91	10.50
2020–2021	Sap flow (L/day)	2.69	0.46	4.60	1.30
	Solrad (MJ/day)	19.32	12.60	26.80	4.81
	VPD	1.22	0.50	2.43	0.45
	RH	62.07	30.70	93.45	11.08

Sap flow – sap flow velocity; Solrad – solar radiation; VPD – vapour pressure deficit; RH – relative humidity.

The descriptive statistics for the indigenous species, Ocotea bullata, Ilex mitis and Podocarpus latifolius, are presented in Table 4. Sap flow follows the same trend as with pine species, decreasing over the years, achieving a mean of 13.14 LPD in 2013–2014 and 6.27 LPD in 2017–2018. Environmental variables undulate in indigenous species as well, with solar radiation achieving a mean of 19.64 MJ in 2013–2014 and 20.15 MJ in 2017–2018. VPD achieved a mean of 2.15 kPa in 2013–2014 and 2.64 kPa in 2017–2018, where RH achieved a mean of 60.37% in 2013–2014 and 58.64% in 2017–2018.

Table 4.

Descriptive statistics of indigenous species and corresponding environmental variables.

Date	Parameter	Mean	Minimum	Maximum	Standard deviation
2013–2014	Sap flow (L/day)	13.14	2.15	24.54	5.77
	Solrad (MJ/day)	19.64	11.95	29.47	5.81
	VPD (kPa/day)	2.15	1.22	4.19	0.86
	RH (%/ day)	60.37	36.05	84.69	12.34
2015–2016	Sap flow (L/day)	8.05	2.85	14.08	2.93
	Solrad (MJ/day)	22.43	9.84	30.94	8.60
	VPD (kPa/day)	1.83	1.13	3.09	0.66
	RH (%/ day)	67.28	43.63	90.94	7.78
2017–2018	Sap flow (L/day)	6.27	0.98	12.97	3.28
	Solrad (MJ/day)	20.15	10.44	31.70	7.69
	VPD (kPa/day)	2.64	1.16	5.17	1.16
	RH (%/ day)	58.64	32.34	84.95	10.79

Sap flow – sap flow velocity; Solrad – solar radiation; VPD – vapour pressure deficit; RH – relative humidity.

The descriptive statistics for Eucalyptus species, Eucalyptus grandis x E. nitens and E. dunnii, are presented in Table 5. Sap flow reaches a mean of 2.89 LPD in the years 2020–2021. The other environmental variable means for 2020–2021 were as follows: The Eucalyptus achieved the lowest solar radiation of all species achieving a mean of 19.23 MJ, VPD achieved a mean of 1.23 kPa and RH achieved a mean of 61.92%.

Table 5.

Descriptive statistics of Eucalyptus species and corresponding environmental variables.

Date	Parameter	Mean	Minimum	Maximum	Standard deviation
2020–2021	Sap flow (L/day)	2.89	0.89	7.48	1.65
	Solrad (MJ/day)	19.23	12.60	26.80	4.75
	VPD (kPa/day)	1.23	0.50	2.43	0.43
	RH (%/ day)	61.92	30.56	93.28	11.03

Sap flow – sap flow velocity; Solrad – solar radiation; VPD – vapour pressure deficit; RH – relative humidity.

4.2. Sap flow prediction results for all species tested

Prediction results of sap flow using the respective satellite vegetation indices and climatic data variables when combining all species into one dataset are presented in Figure 3. The model achieved an R² of 0.97 overall, with an RMSE of 2.39 LPD in sap flow.

Figure 3.

Prediction results of all species, displaying variation between predicted sap flow and observed sap flow.

As mentioned, the Pine species that were included in the measurement had the highest error rate in sap flow units. When Pine species were removed from the model, accuracy resulted in an R² of 0.97 and an RMSE of 0.78 LPD in sap flow. This suggested that sap flow is more difficult to predict in pine than in the indigenous and Eucalyptus species, and that there may be more outliers in model scoring. Overall, the model was successful in predicting sap flow, with predicted and actual sap flow closely overlapping in instances of measurements, providing a framework to conduct future experiments.

4.3. Correlation analysis for individual species using a combined dataset consisting of satellite, climatic and sap flow data

Accuracy results of all species of sap flow predictions tests are presented in Table 6. The results were favourable, with the lowest being of the indigenous trees Podocarpus latifolius, with an R² value of 0.77. Both the indigenous and Eucalyptus species had a low RMSE, with the former achieving the lowest error rate of 0.41 LPD in sap flow. The Pine species had the highest error rate, with Pinus elliotii reaching an RMSE of 2.32 LPD and Pinus radiata reaching an error rate of 5.58 across all species.

Table 6.

Correlation accuracies for all species, along with error rates in units of sap flow.

Species	Coefficient of determination (R²)	RMSE in sap flow (LPD)
Ilex mitis	0.9000	1.1039
Ocotea bullata	0.8708	1.5022
Podocarpus latifolius	0.7742	1.7358
Eucalyptus grandis x E. nitens and E. dunnii	0.9088	0.4055
Pinus elliotii	0.9392	2.3196
Pinus radiata	0.9362	5.5814

4.4. Relationship between observed and predicted sap flow for individual species

The accuracy of observed sap flow and predicted sap flow of Ilex mitis is presented in Figure 4. These figures give a slightly better indication of sap flow prediction accuracy and outlie in the prediction. The prediction achieved a favourable R² value of 0.74.

Figure 4.

Observed versus predicted sap flow for Ilex mitis presented as a time sequence and scatterplot.

Figure 5 illustrates the observed and predicted sap flow accuracy of Ocotea bullata, which achieved a favourable R² value of 0.73.

Figure 5.

Observed versus predicted sap flow for Ocotea bullata presented as a time sequence and scatterplot.

The observed and predicted sap flow accuracy of Podocarpus latifolius is illustrated in Figure 6, similarly achieving a favourable accuracy of R² = 0.76. This accuracy was slightly higher than the other species in the indigenous trees tested. However, all species achieved favourable results.

Figure 6.

Observed versus predicted sap flow for Podocarpus latifolius presented as a time sequence and scatterplot.

Figure 7 illustrates the observed sap flow and predicted sap flow results for Eucalyptus grandis x E. nitens and E. dunnii. Eucalyptus species were grouped for the purpose of single specie prediction, as RapidMiner was unable to predict sap flow for individual species due to data size. Eucalyptus species achieved a favourable R² value of 0.75.

Figure 7.

Observed versus predicted sap flow for Eucalyptus grandis x E. nitens and E. dunnii presented as a time sequence and scatterplot.

The observed and predicted sap flow accuracy of Pinus elliotii is illustrated in Figure 8, achieving an R² value of 0.77.

Figure 8.

Observed versus predicted sap flow for Pinus elliotii presented as a time sequence and scatterplot.

Lastly, the observed and predicted sap flow accuracy for Pinus radiata is illustrated in Figure 9. Pinus radiata also achieved a favourable accuracy of R² = 0.76. Overall, all species have achieved similar results and have favourable predictions of sap flow.

Figure 9.

Observed versus predicted sap flow for Pinus radiata presented as a time sequence and scatterplot.

4.5. Variable importance of all environmental variables and sensors tested

Table 7 presents the most influential variables, which contributed to model scoring. Of all spectral bands measured, band 6 (shortwave infrared 1) was the most influential in scoring, with a weight of 0.001391. Furthermore, of the indices used, NDMI (0.013935), NPCRI (0.004067) and NDWI (0.001358) were the most significant in scoring. Species was the most influential with a weight of 0.016051. Of the environmental variables, vapour pressure deficit was most significant with a weight of 0.015513, followed by relative humidity with 0.012576. The remaining variables did not contribute to positive scoring importance in model scoring.

Table 7.

Scoring weights of the most influential environmental and remote sensing variables in GBT.

Variable	Scoring importance
Species	0.016051
Vapour pressure deficit	0.015513
NDMI	0.013935
Relative humidity	0.012576
NPCRI	0.004067
Band 6	0.001391
NDWI	0.001358

5. Discussion

This study proved successful in predicting sap flow through measurements using satellite and climatic information. The accuracies of all species were favourable, however, there was a clear higher rate of error amongst the Pinus species in comparison to the indigenous and Eucalyptus species. The advantages of acquiring this information through Landsat 8 images, through predicting forest health in the present, allow for the prediction of possible damage in the future which would aid in management decisions (Drake et al., 2016; Ferreira et al., 2023; Kirilenko and Sedjo 2007; Van der Merwe et al., 2022). However, the lack of image availability due to cloud cover does increase the difficulty of predicting data over the years. Furthermore, sap flow data obtained was measured hourly, scaled up to daily data to match daily images availability of Landsat 8. If remote sensing data can be acquired at least daily for future studies, this could result in a highly detailed analysis of predictive capabilities over time.

5.1. Potential of remote sensing data for determining sap flow velocity

The sap flow data used in this study was acquired from past research that exceeded the archives of image availability of Sentinel-2. Thus, Landsat 8 was the only available remote sensing data that could be acquired, as its image availability reaches back the farthest. The resolution of Landsat data is moderate (30 m), and thus tree level analysis of forest stands is not possible (Brewer et al., 2022; Ferreira et al., 2023; Forkuor et al., 2017). However, as sap flow data acquisition is expensive and difficult to reproduce over an entire forest compartment and sample sites are scattered across forest stands to represent forest health, the relationship between sap flow, Landsat 8 and environmental variables does show a positive outcome for monitoring forest condition and health (Ferreira et al., 2023; Huang and Asner 2009). Nevertheless, the resolution of Landsat did not have an adverse effect on the ability to perform a time series of data, and the prediction of sap flow is still successful.

The results of this study show the potential of using remote sensing data for sap flow prediction in commercial (R² > 0.90) and ingenious trees (R² > 0.77) with high accuracy. While the respective studies used other variables to determine sap flow, only a few studies looked at remote sensing data, such as Indices or satellite bands (Hasimoto et al., 2023; Nagler et al., 2007) to improve on sap flow detection. The results obtained using satellite vegetation indices are in line with the above literature proving and alternative methods for further testing of future work.

5.2. Climatic variables for determining sap flow velocity

Of the environmental variables collected from local site weather stations, RH and solar radiation had the highest impact on model scores and sap flow predictions, as the correlation between sap flow changes and solar radiation (which has a significant influence on transpiration rates) is high (Kaptein et al., 2023; Mapeto et al., 2018). Changes in climatic conditions on site may have an adverse impact on water availability, which will impact RH and transpiration (Czimber and Galos 2016; Keskitalo et al., 2016; Mapeto et al., 2018). Thus, measuring these variables for future sap flow predictions may add value to the accuracy of the time series. Finally, GBT performed well within RapidMiner, with an accuracy of R² = 0.96, RMSE = 6.13 and R² = 0.97, RMSE = 0.78 for sap flow (when Pinus was removed from time series). Similarly, a study by Aminuddin et al. (2024) also found that GBT to be the better performing algorithm, through testing several machine learning algorithms in RapidMiner and through forecasting robust wind speeds (Aminuddin et al., 2024).

The results of this study show that combining climate variables with remote sensing data can provide a valuable methodology to determine sap flow in tree compartments. The important variables shown in this study have been confirmed to have a strong relationship with sap flow in other studies (Kaptein et al., 2022; Mapeto et al., 2018). However, with the utility of satellite technologies, spatial datasets can now be used over time to enhance ground measurement station data and to predict in areas where there is restricted data available, such as resource-poor countries.

5.3. Monitoring tree health in future studies

This study demonstrated that sap flow can be predicted using tree species, weather data and variables extracted from Landsat imagery. It was demonstrated that the software package RapidMiner was a convenient way to visualize and analyse data. Furthermore, although sap flow is expensive information to gather, when sample stations are spread across forest compartments, ground-based data collected from sap flow stations can continue to be used as reference data to improve the predictive models. Therefore, management and planning can be made easier, when paired with infield data collection. Future research may focus on a comparative analysis of RapidMiner, versus traditional data analysis methods such as R or Python, as well as comparisons across tree species on other sites. This research can benefit the forestry sector, by providing a framework that aids in forest management through monitoring. Predicting canopy moisture status can provide insight into potential risks and provide more information to intervene and apply efficient management tactics. This research can also aid in the analysis of catchment dynamics and landscape ecology, by providing information pertaining to water-use efficiency. Finally, it aids in differentiating between water-use by different species across variable sites.

6. Conclusion

The ability of Landsat 8, paired with RapidMiner and GBT, to predict sap flow over time has made a valuable contribution to tree health monitoring. The most influential variables in predictions were species, and amongst the environmental variables were VPD and RH. As seen with the results, pine species significantly influenced the overall performance of predictions when they were included. The research in this chapter has demonstrated the use of RapidMiner as a tool to provide a quick and efficient analysis of data, as well as to visualize the changes in sap flow over the course of monitoring. Results also suggest that the pine species were the least performing and had the highest variation in sap flow prediction results. The results achieved were favourable; however, more research is required in monitoring and predicting the sap flow of different species through remote sensing.

Nevertheless, the results indicate the usefulness of Landsat 8 in predicting sap flow, which can be valuable for a range of applications in management and decision-making frameworks used in forestry landscapes. This research will contribute to future studies, help refine prediction models, potentially assist with water budgeting and allocation procedures and aid in the health monitoring of commercial forest landscapes.

Footnotes

Acknowledgements

The authors would like to thank Sappi Forests-SA for granting access to the study sites at Shaw Research Centre and excellent working conditions.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by the National Research Foundation of South Africa (grant number 114898).

ORCID iD

Regardt Ferreira

References

Aminuddin

Hesty

Konitat

, et al. (2024) Promoting wind energy by robust wind speed forecasting using machine learning algorithms optimization. Evergreen 11: 354–370.

Avtar

Komolafe

Kouser

, et al. (2020) Assessing sustainable development prospects through remote sensing: a review. Remote Sensing Applications: Society and Environment 20: 100402.

Baldi

La Porta

(2020) Molecular approaches for low-cost point-of-care pathogen detection in agriculture and forestry. Frontiers in Plant Science 11: 570862.

Børja

Svetlik

Nadezhdin

, et al. (2015) Sap flux - a real time assessment of health status in Norway spruce. Scandinavian Journal of Forest Research 31: 450–457.

Brewer

Clulow

Sibanda

, et al. (2022) Estimation of maize foliar temperature and stomatal conductance as indicators of water stress based on optical and thermal imagery acquired using an unmanned aerial vehicle (UAV) platform. Drones 6: 169.

Burgess

SSO

Adams

Turner

, et al. (2001) An improved heat pulse method to measure low and reverse rates of sap flow in woody plants. Tree Physiology 21: 589–598.

Callens

Morichon

Abadie

, et al. (2020) Using random forest and gradient boosting trees to improve wave forecast at A specific location. Applied Ocean Research 104: 102339.

Carrilo

Patocka

Dobrovolny

, et al. (2020) Monitoring bark beetle forest damage in central europe. A remote sensing approach validated with field data. Remote Sensing 12: 3634.

Chen

Guestrin

. (2016). XGBoost: A Scalable Tree Boosting System. KDD’16, San Francisco, CA, August 13–17.

10.

Czimber

Galos

(2016) A new decision support system to analyse the impacts of climate change on the Hungarian forestry and agricultural sectors. Scandinavian Journal of Forest Research 31: 664–673.

11.

Das

Christopher

Apan

, et al. (2021) Evaluation of water status of wheat genotypes to aid prediction of yield on sodic soils using UAV-thermal imaging and machine learning. Agricultural and Forest Meteorology 307: 108477.

12.

Dash

Watt

Pearse

, et al. (2017) Assessing very high-resolution UAV imagery for monitoring forest health during a simulated disease outbreak. ISPRS Journal of Photogrammetry and Remote Sensing 131: 1–14.

13.

Dragavtsev

Nartov

(2015) Application of thermal imaging in agriculture and forestry. European Agrophysical Journal 2(1): 15.

14.

Drake

Hanson

Lowrey

, et al. (2016) The carbon fertilization effect over a century of anthropogenic CO2 emissions: higher intracellular CO2 and more drought resistance among invasive and native grass species contrasts with increased water use efficiency for woody plants in the US southwest. Global Change Biology 23: 782–792.

15.

Engelbrecht

Monteiro

(2021) The IPCC assessment report six working group 1 report and southern Africa: reasons to take action. South African Journal of Science 117: 1–7.

16.

Ferreira

Peerbhay

Louw

, et al. (2022) A spatio-temporal analysis of baboon damage using sentinel-2 imagery and extreme gradient boosting. Geocarto International 37: 10–2943.

17.

Ferreira

Peerbhay

Louw

, et al. (2023) A tree-level analysis of baboon damage in commercial forest stands using deep learning techniques. Southern Forests: A Journal of Forest Science 85: 65–73.

18.

Forkuor

Dimobe

Serme

, et al. (2017) Landsat-8 Vs. Sentinel-2: examining the added value of sentinel-2’s red-edge bands to land-use and land-cover mapping in Burkina Faso. GIScience and Remote Sensing 55: 331–354.

19.

Gabron

Pinty

Verstraete

, et al. (2000) Advanced vegetation indices optimized for up-coming sensors: design, performance, and applications. IEEE Transactions on Geoscience and Remote Sensing 28: 6–2505.

20.

Gao

(1996) NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment 58: 257–266.

21.

Guillen

Aparicio

Esteve

(2023) Gradient tree boosting and the estimation of production frontiers. Expert Systems with Applications 214: 119134.

22.

Hamdi

Brandmeier

Straub

(2019) Forest damage assessment using deep learning on high resolution remote sensing data. Remote Sensing 11: 1976.

23.

Hasimoto

Chiu

Onda

, et al. (2023) Satellite remote sensing model for estimating canopy transpiration in cypress plantation using in situ sap flow observations and forest inventory. ISPRS Journal of Photogrammetry and Remote Sensing 206: 258–272.

24.

Hastie

Tibshirani

Friedman

(2009) The Elements of Statistical Learning: Data Mining, Inference and Prediction. Berlin: Springer Science and Business Media.

25.

Afan

El-Shafie

, et al. (2019) Towards A time and cost-effective approach to water quality index class prediction. Journal of Hydrology 575: 148–165.

26.

Huang

Asner

(2009) Applications of remote sensing to alien invasive plant studies. Sensors 9: 4869–4889.

27.

Huang

Ding

Zhu

, et al. (2021) Assessment of poplar looper (Apocheima cinerarius Erschoff) infestation on euphrates (Populus euphratica) using time-series MODIS NDVI data based on the wavelet transform and discriminant analysis. Remote Sensing 13: 2345.

28.

Huo

Persson

Lindberg

(2021) Early detection of forest stress from European spruce bark beetle attack and A new vegetation index: normalized distance red & SWIR (NDRS). Remote Sensing of Environment 255: 112240.

29.

Joharestani

Cao

, et al. (2019) PM2.5 prediction based on random forest, xgboost, and deep learning using multisource remote sensing data. Atmosphere 10: 373.

30.

Kaptein

Clulow

Toucher

, et al. (2022) Comparative water-use by fast-growing E. Grandis X E. Nitens clonal hybrid and Pinus elliottii near the two streams research catchment. South Africa. EGUsphere 27: 4467–4484.

31.

Kaptein

Clulow

Toucher

, et al. (2023) Plantation water productivity (pwpwood) and not water-use efficiency (wue) as the measure of commercial plantation yield improvement: a review. Southern Forests: A Journal of Forest Science 85(1): 01–10.

32.

Keskitalo

ECH

Bergh

Felton

, et al. (2016) Adaption of climate change in Swedish forestry. Forests 7: 28.

33.

Kirilenko

Sedjo

(2007) Climate change impacts on forestry. Proceedings of the National Academy of Sciences of the United States of America 104(50): 19697–19702.

34.

Lechner

Foody

Boyd

(2020) Applications in remote sensing to forest ecology and management. One Earth 2: 405–412.

35.

Lehmann

JRK

Nieberding

Prinz

, et al. (2015) Analysis of unmanned aerial system-based cir images in forestry - a new perspective to monitor pest infestation levels. Forests 6: 594–612.

36.

Mahlangu

Mathieu

Wessels

, et al. (2018) Indirect estimation of structural parameters in South African forests using misr-hr and lidar remote sensing data. Remote Sensing 10: 1537.

37.

Mapeto

Louw

Gush

, et al. (2018) Whole-tree sap flow responses to soil water and weather variables for Pinus radiata and three indigenous species in A southern afrotemperate forest region. Southern Forests: A Journal of Forest Science 80: 329–339.

38.

Matthews

Netherer

Katzensteiner

, et al. (2018) Transpiration deficits increase host susceptibility to bark beetle attack: experimental observations and practical outcomes for ips typographus hazard assessment. Agricultural and Forest Meteorology 263: 69–89.

39.

McCullagh

Nieuwenhuis

(2015) Review of tools for growth forecasting and productivity assessment in forestry in Ireland. Irish Forestry 72: 2015.

40.

Merzlyak

Gitelson

Chivkunova

, et al. (1999) Non-destructive optical detection of pigment changes during leaf senescence and fruit ripening. Physiologia Plantarum 106: 135–141.

41.

Nagler

Jetton

Fleming

, et al. (2007) Evapotranspiration in a cottonwood (Populus fremontii) restoration plantation estimated by sap flow and remote sensing methods. Agricultural and Forest Meteorology 144: 95–110.

42.

Ovchynikova

Dupliak

(2021) Analysis, forecasting, and management of indicators of the forest industry of the region. SHS Web of Conferences 107: 100005.

43.

Peerbhay

Adelabu

Lottering

, et al. (2022) Mapping carbon content in A mountainous grassland using SPOT 5 multispectral imagery and semi-automated machine learning ensemble methods. Scientific African 17: 01344.

44.

Pullanagari

Kereszturi

Yule

(2016) Mapping of macro and micro nutrients of mixed pastures using airborne AisaFENIX hyperspectral imagery. ISPRS Journal of Photogrammetry and Remote Sensing 117: 1–10.

45.

Rouse

Haas

Schell

, et al. (1973) Monitoring vegetation systems in the great plains with ERTS. In: Third ERTS Symposium. NASA SP-351. Washington, DC: Goddard Space Flight Center.

46.

Schoene

DHF

Bernier

(2012) Adapting forestry and forests to climate change: a challenge to change the paradigm. Forest Policy and Economics 24: 12–19.

47.

Skelton

Brodribb

(2018) Lag in gas exchange recovery following natural drought associated with embolism formation. Acta Horticulturae: 49–58.

48.

Solh

Van Ginkel

(2014) Drought preparedness and drought mitigation in the developing world’s drylands. Weather and Climate Extremes 3: 62–66.

49.

Urban

Ingwers

McGuire

, et al. (2017) Stomatal conductance increases with rising temperature. Plant Signaling & Behavior 12(8): e1356534.

50.

Van der Merwe

Germishuizen

Clarke

, et al. (2022) The impact of soil, altitude, and climate on tree form and wood poperties of plantation grown Pinus patula in mpumalanga, South Africa. Holzforschung 77(1): 1–15.

51.

Van der Merwe

Lindner

Ernst

, et al. (2024) The impact of climate on the dominant height and climate thresholds for P. elliottii, P. taeda, P. patula, and P. patula x P. tecunumanii plantation forests. Canadian Journal of Forest Research 54: 895–906.

52.

Warburton

Schulze

(2008) Potential impacts of climate change on the climatically suitable growth areas of Pinus and Eucalyptus: results from A sensitivity study in South Africa. Southern Forests: A Journal of Forest Science 70(1): 27–36.

53.

Wintle

Lindenmayer

(2008) Adaptive risk management for certifiably sustainable forestry. Forest Ecology and Management 256(6): 1311–1319.

54.

Xulu

Peerbhay

Gebreslasie

, et al. (2019) Unsupervised clustering of forest response to drought stress in Zululand region, South Africa. Forests 10: 531.

55.

Zhang

Eziz

Xiao

, et al. (2019) High-resolution vegetation mapping using extreme gradient boosting based on extensive features. Remote Sensing 11: 1505.

Understanding forest sap flow velocity through satellite based remote sensing platforms: A time series analysis

Abstract

Keywords

1. Introduction

2. Material and methods

2.1. Study area

2.2. Field data collection and climatic variables

2.3. Sap flow velocity data acquisition

2.4. Landsat images and pre-processing

2.5. RapidMiner

2.5.1. Time series method

2.5.2. Gradient boosting trees

3. Accuracy assessment

4. Results

4.1. Descriptive statistics of all species and the corresponding variables measured at yearly intervals

4.2. Sap flow prediction results for all species tested

4.3. Correlation analysis for individual species using a combined dataset consisting of satellite, climatic and sap flow data

4.4. Relationship between observed and predicted sap flow for individual species

4.5. Variable importance of all environmental variables and sensors tested

5. Discussion

5.1. Potential of remote sensing data for determining sap flow velocity

5.2. Climatic variables for determining sap flow velocity

5.3. Monitoring tree health in future studies

6. Conclusion

Footnotes

Acknowledgements

Declaration of conflicting interests

Funding

ORCID iD

References