The destruction of forest roads has significant adverse impacts on forest accessibility, resulting in heightened maintenance and environmental expenses, as well as posing potential threats to human life. Enhancing the sustainable and secure accessibility to forest regions necessitates a more comprehensive knowledge of the deterioration of forest pathways across temporal and spatial dimensions. The objective of this study is to utilize destruction prediction models to identify pertinent variables related to the state of the Hyrcanian forest road network in Golestan province. Additionally, the study aims to determine the most significant explanatory variables. To achieve the research objective, a set of 50 points were established along forest roads at a uniform distance of 200 meters from each other. The response variable, which pertains to the visibility of the forest road infrastructure, was recorded in binary format. Additionally, the explanatory variables were documented along a strip transect with a width of 10 meters perpendicular to the road. The study employed two conventional logistic regression models and a novel random forest model to forecast the deterioration of forest roads. Three indicators consist of TWI, natural ground gradient and cover density of cut slope, can predict road damage, as demonstrated by the implementation of two models. The random forest model exhibited superior accuracy to logistic regression, as evidenced by its success rate of 0.73 and prediction rate of 0.68. The results indicate that machine learning models offer significant insights into predicting road conditions and ensuring access to the Hyrcanian forests.
Research Article
Using logistic regression and random forest techniques to predict the degradation of forest roads
https://doi.org/10.21203/rs.3.rs-3247554/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Preventive measures
cut slope
sustainable access
Loveh
restoration
Forest roads are necessary for forestry activities and harvesting, and it is necessary to maintain their quality and perform routine maintenance to assure access to forest resources (Couler et al. 2006). Despite the fact that logging operations are the primary reason for the construction of forest roads around the globe, access for recreational use and forest fire suppression are also of great importance. Although there are relatively few forest roads in comparison to other regions, more than a thousand kilometers of forest roads in northern Iran need annual maintenance and recovery. Construction of new roads and bridges, reconstruction and maintenance of existing roads based on road type grading, drainage improvement, roadside vegetation clearance, and resurfacing of roads with a high level of deterioration are among the tasks performed annually. Prior to the prohibition of harvesting in the forests of the north of Iran, the Natural Resources and Watershed Management Organization, which is responsible for more than 12,000 kilometers of forest roads in the north of Iran, spent more than 80 billion Rials annually on road construction and maintenance. However, it was the obligation of the contractors working on the forestry project to complete these duties. After the implementation of the plan to stop the exploitation of the northern forests and the depletion of financial resources, the annual construction and maintenance expenses of forest roads have been completely neglected.
Forest road deterioration is primarily affected by two variables: water flow and traffic density (Kastridis et al. 2020). Wet roads are most damaging when there is a steady flow of vehicles. Therefore, numerous buildings and technical structures such as waterways, piping, sediment ponds, side streams, etc. are utilized in the construction of forest roads so that the road bed is less affected by water flows. Meanwhile, the terrain, forest stand type, and adherence to technical requirements in building forest roads, as well as maintaining and managing them, are crucially important. Easy monitoring of forest roads at risk of destruction is possible with even the most basic information on the topography and coverage state of the roadside stands, as well as the health or destruction of the road and road-side bed. Knowing which parts of roadways are most vulnerable to damage is, thus, crucial. This will extend the life of the roads and, if necessary, make it easier to identify, maintain, and restore areas susceptible to damage.
In forest roads, early detection of severe damage such as subgrade emergence is crucial, as this type of damage frequently and potentially occurs in the lower and middle layers. In addition, by generating a missing layer beneath the surface, the road surface is damaged, the surface track is obliterated within days, the bottom layer is exposed, leading to significant safety issues. On the other hand, unanticipated road damage necessitates immediate maintenance or repair, resulting in traffic delays, increased fuel consumption, and pollution (White et al. 2013). This process generates a substantial amount of carbon, which, in addition to its high cost and duration, also depletes natural resources (Kiss et al., 2015). In such a case, the identification of the factors contributing to the destruction of roads enables the adoption of preventative measures or prompt action to avert future occurrences of similar destruction (Waga et al. 2020; Shtayat et al. 2022).
In the form of a high-performance prediction model, machine learning techniques are a valuable method for identifying the factors influencing road deterioration (Guo and Hao 2021). By simulating situations in which severe damage has occurred at the roadway's upper levels, these techniques help to identify and examine sections of the road that have not yet exhibited a distinct reaction on the surface. In this manner, a significant challenge can be reduced to a minor issue, and the road's value can be preserved. In this regard, the current study aims to identify the most significant factors influencing the destruction of forest pathways in the form of a prediction model by employing the Logistic Regression (LR) and Random Forest (RF) models.
Study area
The study territory, Loveh forests, is located in the Iranian province of Golestan at 55° 41' to 55° 46' longitude and 37° 17' to 37° 19' latitude. The minimum and maximum height above sea level in the study area are 1250 meters and 1820 meters, respectively, and the greatest canopy cover is found on slopes with an incline greater than 30 degrees along the northern and northwestern slopes. The primary roads of Loveh forests (series 1), named Haj abdulah and Ghanbar, with a combined length of 11,3 kilometers have been constructed in the region (Fig. 1). According to the most recent revision plan booklet from 2012, 2,094,950 million Rials were considered for the rehabilitation and maintenance of existing major roads. Due to the implementation of the cessation of forest utilization plan, the dismissal of forestry project managers, and the absence of budget allocation, no construction work has been conducted in this field.
Methods
To determine the most influential factors affecting subgrade emergence (Fig. 2), 50 points were established on forest roads in October 2021, 200 meters apart, using a random starting point. In addition to recording the subgrade emergence as a binary response variable, the characteristics of the explanatory variables were recorded along a 10-meter-wide transect perpendicular to the road (Table 1) (Girardin et al. 2021). Using the digital elevation model within the Saga 2.2.7 software, the Topographic Position Index (TPI) was also calculated for each recorded position. The technical specifications and road maintenance along the examined roadways are shown in Table 2.
|
Predictor variables |
Minimum |
Maximum |
Average |
|---|---|---|---|
|
Road gradient (%) |
1 |
7 |
3.278 |
|
Cross slope (%) |
1 |
4 |
2.304 |
|
Road surface template |
Centerline crown = 0%, In slope = 27.27, Out slope = 72.73% |
||
|
TPI |
Ridges = 57.83, Slope = 26.09, Canyons = 26.08 |
||
|
Natural ground gradient (%) |
25 |
90 |
55.562 |
|
Cut slope (%) |
0 |
90 |
64.304 |
|
Cut length (m) |
0 |
7 |
2.708 |
|
Cut slope stability |
Stable = 61%, Unstable = 39% |
||
|
Cover density of cut slope (%) |
0 |
90 |
37.130 |
|
Fill slope (%) |
0 |
90 |
39.455 |
|
Fill length (m) |
0 |
6 |
1.398 |
|
Fill slope stability |
Stable = 87.77%, Unstable = 12.12% |
||
|
Cover density of fill slope (%) |
0 |
95 |
37.606 |
|
TWI |
1.8 |
12.5 |
6.813 |
|
Ditch blockage |
No blockage = 48.48, Blockage = 51.52 |
||
|
Road technical components |
Average recorded in the field |
Standard |
|---|---|---|
|
Travel way width (m) |
2.323 |
3.5 |
|
Shoulder width (Out slope) (m) |
0.783 |
0.5 |
|
Shoulder width (In slope) (m) |
0.670 |
0.5 |
|
Ditch width up (m) |
0.467 |
1 |
|
Ditch width down (m) |
0.261 |
0.3 |
|
Ditch depth (m) |
0.153 |
0.35 |
LR and RF techniques were used to determine the most influential subgrade emergence factors in the form of a prediction model. The LR model is typically defined as follows (Eq. 1):
Where Z is the multiple linear equation derived from the logit function or the response variable in the preceding equation.
The objective of this form of model in the analysis is to produce a final model that includes all significant variables. Consequently, the stepwise regression model and AIC criterion were employed. In each stage, a non-significant variable was eliminated from the model, and if the model improved by dropping the variable, the resulting model was confirmed and used as the basis for the next step. This procedure was repeated until an optimal model was achieved. The predictive variables in the optimal model have been identified as the most crucial variables (Mozaffaree Pour and Oja 2022).
In order to implement the RF model, an n-observation random sample of 66% of the primary data was selected using the bootstrap method. Then, using recursive algorithms, a classification tree was grown for the designated bootstrap sample. After a large number of trees have been generated, each tree votes for the most popular class. By combining the votes of various trees, a class prediction is made for each sample. In the subsequent step, the recursive algorithm proceeds until the tree reaches its maximum size (one final node for each observation), without pruning. The initial three stages are then repeated T times to create an RF. The T is typically between 500 and 1,000 trees, and m is log (p), where p is the number of explanatory variables (Genuer 2008). A RF is so large that deciphering it is an arduous undertaking, so its information must be summarized using quantitative indicators. The significance of the variable of one of these indicators. On the basis of this index's value, the most influential variables were selected. The randomForest package was used to adapt the samples to the random forest model, and the Mean Decrease Gini and Mean Decrease Accuracy indices were used to compare and select the most suitable model (Genuer 2008).
The Area Under the Curve (AUC) index was utilized to assess the accuracy and efficacy of the aforementioned predictive models. This criterion is part of the evaluation of prediction and classification models. The model's accuracy is the proportion of the set of samples correctly classified by the model. In other terms, the performance of the model in identifying distinct classes is evaluated using this index. In this study, the available data were arbitrarily divided into two parts: training (70%) and testing (30%). The accuracy of the models was evaluated on the test set to ensure their reliability. In this research, all statistical analysis was conducted using R 4.0.2 software.
Using the LR, 39% of variance in the appearance of infrastructure were explained by explanatory variables, according to the findings of this study. The influential variables were chosen utilizing the sequential selection method and Akaike's information criterion, and the results are shown in Table 3. In the final model (Eq. 2), the natural ground gradient (%) and cover density in cut slope (%) were the only effective variables.
| \({\rm Y}=\frac{{e}^{\text{2.328 +0.080 X1 - 0.068 X2}}}{1+{e}^{\text{2.328 +0.080 X1 - 0.068 X2}}}\) | Equation (2) |
| Variables | DF | Deviance | AIC | Pr (> Chi) |
|---|---|---|---|---|
| Natural ground gradient (%) | 1 | 27.722 | 31.722 | 0.003693 |
| Cover density of cut slope (%) | 1 | 26.257 | 30.257 | 0.008315 |
As previously stated, the first stage of the RF is to determine the optimal number of trees and predictor variables for splitting each node. Therefore, the RF was implemented with all variables, and only those variables whose value changes caused an increase in the mean squared error were identified as effective variables on subgrade emergence (Oliveira et al., 2012). Five influential variables were incorporated into the final model. The following step's number of trees and random variables were found to be 1000 and 3, respectively, based on the error graph of the number of trees and the error graph of the number of variables for the final model. The RF explains 79.37% of the subgrade emergence variations using the selected variables. Table 4 presents the relative importance of the explanatory variables. Figure 3 depicts the partial dependence diagram of the most influential variables on the emergence of road infrastructure. With the increase of the topographic wetness index, the natural ground gradient, and the cover density of cut slope, steeply sloping graphs have demonstrated the emergence of infrastructure subgrade. The steepest incline of the TWI graph is visible between 5 and 10, after which the slope becomes more gradual. In addition, the two variables of natural ground gradient (percent) and cover density of cut slope (percent) had an increasing impact on the subgrade emergence of the road in slopes of 60% and 55%, respectively.
| Variables | Variable importance |
|---|---|
| TPI | 1.855 |
| Road gradient (%) | 4.428 |
| Cross slope (%) | 1.443 |
| Road Surface Template | 1.009 |
| Natural ground gradient (%) | 9.682 |
| Cut Slope (%) | 5.230 |
| Cut Length (m) | 1.230 |
| Cut slope stability | 1.868 |
| Cover density of cut slope (%) | 6.095 |
| TWI | 29.595 |
| Ditch blockage | 1.199 |
As previously stated, the area under the curve (AUC) metric was used to evaluate the efficacy of the prediction model. Based on the results of the validation (Table 5), the RF model was more precise than the LR model.
| Models | AUC (Success rate) | AUC (Prediction rate) |
|---|---|---|
| LR | 0.684 ± 0.031 | 0.631 ± 0.024 |
| RF | 0.883 ± 0.004 | 0.852 ± 0.019 |
Obtaining information on the quantitative and qualitative condition of forest roads allows forest managers and executives to allocate resources for road maintenance in the most efficient manner. By accumulating information regarding subgrade emergence, this study aimed to facilitate the identification of future points vulnerable to destruction. According to the results of the two models used, the TWI, the natural ground gradient, and the cover density of cut slope were the most significant environmental and technical factors influencing road destruction. With the construction of a network of forest roads, the hydrological status and the dynamics of the ecosystem are drastically altered. Compliance with technical factors during road construction and protection, as well as initial environmental conditions, influence the magnitude of these changes (De Wit et al. 2020). Given that the direction of water flows has an effect on the condition of forest roads, it seems likely that the investigated variables that have a greater effect on the performance of water flows on the surface and sides of the roads will be given a greater weight in the final model. TWI is a critical aspect of the construction, maintenance, and restoration of forest roads. Therefore, it is used to classify the character of forest roads in many forest regions of the world (Waga et al. 2020). Cochrane et al (2007) and Kiss et al (2015) found that the use of this index in evaluations led to the early detection of subgrade emergence and prevented the imposition of hefty financial costs. Kiss et al (2016) observe in their study that the preparation of the TWI, even when derived from digital elevation models with an accuracy of less than 25 meters, and its use increases the likelihood of identifying vulnerable locations on forest roads.
Effective predictor variables were identified as the natural ground gradient in both models and the cover density in cut slope in the final random forest model. In general, road construction and maintenance in sloping locations is fraught with numerous obstacles. In addition to increasing the cost of road construction and the need for specialized equipment, since the volume and intensity of runoff is greater in areas with a steep slope, potholes, surface erosion, and complete pavement devastation are more likely to occur there (Sun et al. 2021). In the study related to the quantitative and qualitative evaluation of the road, it is crucial to pay close attention to and promptly address these areas (Liu et al. 2019). Slope performance on modest slope gradients was shown to not be particularly sensitive to road degradation in this investigation. In support of the current findings, De Ona et al (2011) demonstrated that the effect of slope on road quality is evident at slopes greater than 50 percent. They estimated that the quantity of runoff and erosion generated by steep slopes increases exponentially, thereby increasing the likelihood of forest road pavement washing. This will be exacerbated if construction and maintenance of the road disregard technical considerations.
The cover density of cut slope was also identified as a significant factor in the deterioration of the road surface. This factor's impact on road erosion has been examined in previous research (Teixeira Coelho et al. 2001; Seutloali et al. 2015). Liu et al (2019) demonstrated in their study that vegetation on a cut slope reduces runoff and transported sediments due to the absence of direct raindrop impact and the presence of roots. Vegetation increases the soil's permeability and aids in the consolidation of excavation trenches. In this regard, Arnone et al (2016) found that plant roots in sloping areas enhance the shear strength of the soil by forming a network of connections within the soil layer and by anchoring the surface soil to deeper layers. This prevents the trench from collapsing and obstructing the forest roads' side trench. It should be highlighted that roadside trench obstruction can cause washing away of the pavement and substructure of forest roads, even if the effect was not considerable in the final model. Therefore, it is firmly recommended in numerous studies to artificially cover the trenches along the road.
The comparison of the employed methods revealed that the random forest technique was significantly more accurate. Due to its high capability for analyzing non-linear variables, random forest provides highly accurate fitting and modeling. This is despite the fact that the likelihood of overfitting and error is much lower in this model than in the logistic regression model. However, it should be kept in mind that generalized linear models continue to have a large user base due to their simplicity and transferability to managers. This technique provides a solid understanding of the accuracy of new models, such as decision tree techniques, when used as a benchmark model.
In the present investigation, a combination of environmental and technical variables was used to determine the most influential subgrade emergence factors. The results of the present study indicated that field observations coupled with data mining techniques can provide us with the necessary temporal and spatial information to identify the factors influencing the devastation of forest roads. Given that the necessary data are readily collected in the field or can be extracted for free from basic maps, this method can be implemented rapidly and with high precision on other forest roads. Undoubtedly, the incorporation of other factors will improve our understanding of road degradation in forested environments. With a fresh perspective on forest road network management, this study makes it possible to preserve the environmental, economic, and social benefits of forest roads while minimizing the obstacles that inhibit sustainable development in these three areas.
Funding
This work was supported by the Research Institute of Forests and Rangelands (RIFR) and Agricultural Research, Education and Extension Organization (AREEO) in Iran with code: 0-60-09-070-991261
Data Availability
Some or all data, models, or code that support the findings of this study are available from the corresponding author upon reasonable request.
Author Contributions
SSH & M.M: Methodology, Writing – original draft, Writing – review & editing. AA: Review & editing. HF: Collected the primary data and completed all investigations and formal analysis. All authors reviewed the manuscript.
Compliance with Ethical Standards
Conflict of Interest The authors declare no competing interests.
- Arnone E, Caracciolo D, Noto LV, Preti F, Bras RL (2016) Modeling the hydrological and mechanical effect of roots on shallow landslides. Water Resour Res 52: 8590–8621. DOI: https://doi.org/10.1002/2015WR018227
- Cochrane TA, Egli M, Phillips C, Acharya G (2007) Development of q forest road erosion calculation GIS tool for forest road planning and design. 1273–1279. DOI: https://www.researchgate.net/publication/29487398_Development_of_a_forest_road_erosion_calculation_GIS_tool_for_forest_road_planning_and_design
- Coulter ED, Sessions J, Wing MG (2006) Scheduling Forest Road maintenance using the analytic hierarchy process and heuristics. Silva Fenn 40: 143–160. DOI: https://doi10.14214/sf.357
- De Oña, J, Ferrer A, Osorio F (2011) Erosion and vegetation cover in road slopes hydroseeded with sewage sludge. Transp. Res D Transp Environ 16: 465–468. DOI: https://doi.org/10.1016/j.trd.2011.04.002
- De Witt A, Boston K, Leshchinsky B (2020) Predicting Aggregate Degradation in Forest Roads in Northwest Oregon. Forests. 11 (729): 1–11. DOI: https://doi.org/10.3390/f11070729
- Genuer R (2008) Random Forest: some methodological insights. pp 61. DOI: https://doi.org/10.48550/arXiv.0811.3619
- Guo X, Hao P (2021) Using a Random Forest Model to Predict the Location of Potential Damage on Asphalt Pavement. Appl Sci 11 (21): 1–19. DOI: https://doi.org/10.3390/app112110396
- Kastridis A (2020) Impact of Forest Roads on Hydrological Processes. Forests 11 (11): 1–12. DOI: https://doi.org/10.3390/f11111201
- Kiss K, Malinen J, Tokola T (2015) Forest road quality control using ALS data. Can J For Res 45: 1636–1642. DOI: https://doi.org/10.1139/cjfr-2015-006
- Kiss K, Malinen J, Tokola T (2016) Comparison of high and low density airborne lidar data for forest road quality assessment. ISPRS Ann. Photogramm Remote Sens Spat Inf Sci 3: 167–172. DOI: https://doi:10.5194/isprsannals-III-8-167-2016
- Liu G, Hu F, Zheng F, Zhang Q (2019) Effects and mechanisms of erosion control techniques on stairstep cut-slopes. Sci Total Environ 656: 307–315. DOI: https://doi.org/10.1016/j.scitotenv.2018.11.385
- Mozaffaree Pour N, Oja T (2022) Prediction Power of Logistic Regression (LR) and Multi-Layer Perceptron (MLP) Models in Exploring Driving Forces of Urban Expansion to Be Sustainable in Estonia. Sustain 14 (1): 1–22. DOI: https://doi.org/10.3390/su14010160
- Oliveira S, Oehler F, San-Miguel-Ayanz J, Camia A, Pereira J (2012) Modeling spatial patterns of fire occurrence in Mediterranean Europe using Multiple Regression and Random Forest. For Ecol Manage 275: 117–129. DOI: https://doi.org/10.1016/j.foreco.2012.03.003
- Seutloali KE, Beckedahl HR (2015) A review of road-related soil erosion: an assessment of causes, evaluation techniques and available control measures. Earth Sci Res J 19: 73–80. DOI: https://10.15446/esrj.v19n1.43841
- Shtayat A, Moridpour S, Best B, Rumi S (2022) An Overview of Pavement Degradation Prediction Models. J Adv Transp 7783588: 1–15. DOI: https://doi.org/10.1155/2022/7783588
- Sun C, Hou H, Chen W (2021) Effects of vegetation cover and slope on soil erosion in the Eastern Chinese Loess Plateau under different rainfall regimes. PeerJ 9: e11226, 1–14. DOI: https://doi.10.7717/peerj.11226
- Teixeira Coelho A, Cássia de Brito Galvão T, Rodrigues Pereira A (2001) The effects of vegetative cover in the erosion prevention of a road slope. Environ Manage Health 12 (1): 78–87. DOI: https://doi.org/10.1108/09566160110381968
- Waga K, Tompalski P, Coops NC, White JC, Wulder MA, Malinen J, Tokola T (2020) Forest Road Status Assessment Using Airborne Laser Scanning. For Sci 66: 501–508. DOI: https://doi.org/10.1093/forsci/fxz053
- White JC, Wulder MA, Varhola A, Vastaranta M, Coops NC, Cook BD, Pitt D, Woods M (2013) A best practice guide for generating forest inventory attributes from airborne laser scanning data using an area-based approach. For Chron 89: 722–723. DOI: https://doi.10.5558/tfc2013-132
No competing interests reported.