TRANSPORTATION RESEARCH RECORD 1291 I ·''"'

Comparison of the Impact of Various

Unpaved Road Performance Models on
Management Decisions
P. PAIGE-GREEN AND A. T. VISSER

Over the last .5 years. maintenance management systems for of data collected from a number of countries. The Road
unpaved roads. such as the Maintennnce and Design System and Transport Investment Model. RTIM2 (4). was developed hy
Highway Design and Maintenance System. have become increas- the Transport and Road Research Laboratory (United King-
ingly avail<1ble. Detailed data requirements have. however. deterred
dom) and is based on deterioration models developed mainly
many potential users from implementing these systems. especially
for low-volume roads. Furthermore. these systems are based on in east and west Africa.
deterioration relationships obtained from a study in Brazil. and The impact of a previously presented set of models (5)
the applicability to other areas has not been confirmed. There- developed in southern Africa has been evaluated. These models
fore . local relationships with good predictability and low data predict roughness deterioration and gravel loss over time. The
requirements needed to be developed . During a recent study in input requirements of the different sets of models were com-
southern Africa . a new set of relationships was developed for the pared. The impact that replacing the Brazilian models in the
prediction of roughness deterioration and gravel loss for low-
volume unpnved roads over time. These relationships are much MDS with the new models would have on management deci-
simpler than the previous ones. thus fulfilling local needs. The sions such as budget requirements for regraveling. routine
new relationships were compared with the existing models. The grader blading. and upgrading. was determined.
effects of the different models on management decisions about
annual gravel replacement were examined by comparing the results
of each model as it evaluated a typical unpaved road network.
The blading frequency was evaluated by the blading cost and total DESCRIPTION OF NEW MODELS
cost. including vehicle operating costs. On the basis of these
findings. conclusions about the potential necessity for upgrading
The new models were developed from data collected during
can be drawn. Road authorities worldwide must consider the
applicability of models developed in one environment to other the monitoring of 110 sections of road in the Transvaal prov-
environments and the potential impact on management decisions . ince of South Africa and Namibia over a period of more than
3 years. A factorial experiment, including a wide range of
traffic, climatic. and material properties. has been described
The necessity for improved management of road maintenance fully by Paige-Green (6). The background and methods were
is becoming increasingly important as road networks become summarized at the 4th International Conference on Low-
larger and funding is reduced in real terms. Although this is Volume Roads ( 7). The development of the set of models has
a worldwide problem. it is probably more significant in devel- been discussed previously (5). but subsequent analysis indi-
oping areas. where unacceptably rough unpaved roads result cated that particle size distribution and its method of calcu-
in very high vehicle operating costs. Major drains on the lation affected the regression model. The particle size distribu-
economy often result from increased consumption of fuel. tions were standardized. assuming a maximum size of 37.5
tires , and spare parts. which are generally imported. Valuable mm (8), and the models were reevaluated. Although the sta-
foreign exchange can be saved by providing adequate roads . tistics and variables in the model did not change significantly.
In many developing countries. spare parts are often almost minor changes to the coefficients were necessary.
impossible to obtain. and failure of an essential component
that cannot readily be replaced may render an important vehi-
cle inoperable for an indefinite period.
During the past decade, a number of maintenance man- Gravel Loss Prediction
agement systems have been developed. The Maintenance and
Design System, MDS (J,2). was developed from data obtained The best model for the prediction of gravel loss developed
in Brazil, and the Highway Design and Maintenance System. during the southern African unpaved roads study (6) was as
HDM-III, (3) was developed by the World Bank on the basis follows:

GL = D[ADT(0.059 + 0.0027N - 0.0006 P26)

- 0.367(N) - 0.0014(PF) + 0.0474(P26)] (I)
P. Paige-Green. Division of Roads and Transport Technology, Coun-
cil for Scientific and Industrial Research, Box 395. Pretoria. 0001.
South Africa. A. T. Visser. Department of Civil Engineering, Uni- [r 2 = 0.84; root mean square error (RMSE) = 5.3; n = 703;
versity of Pretoria, Pretoria. 0002. South Africa. and F = 619]
138 TRAf\'SPOR TATIO:\' RESEARCH RECUR!>/:!()/

where Jn R D[ -13 .8 + 0.00021(PF) + 0.064(5 l)

GL = gravel thickness loss (mm): + 0.137(P26) + 0.0003(JV)(ADT)
D ~ time period under consideration in hundreds of
days (days/100); + GM(6.42 - 0.063 P26)] (4)
ADT = average daily traffic in both directions:
N = Weinert N-value (9), which ranges from 1 in wet (re 0.22; /1 = 7005: RMSE = 0. 15: F = 288)
areas to more than 1() in arid areas and incorpo-
rates annual rainfall:
PF = plastic limit x percentage passing 0.075-mm sieve : where
and In R natural logarithm of change of roughness with time.
P26 = percentage passing 26.5-mm sieve . D = number of days since last blading in hundreds (days '
The t values were 3.72 . 6.64 , -3 .63 . -9.72 . -4 .50. and 17.84 100),
for the respective coefficients. The Brazil study (1,2) resulted PF = plastic limit x percentage passing 0.075-mm sieve.
in a number of gravel-loss prediction models, the following Sl = season dummy variable ( l for dry season. 0 for wet
model being the most useful: season),
P26 = percentage passing 26 .5-mm sieve.
GL = D[l.58 + 0.366(G) + 0.083(SV) - 0.210(PI) N = Weinert N-value.
ADT = average daily traffic . and
+ 0.0132(NC) + 0.0081(NT) + 420.45/R] (2) GM = grading modulus (sum of percentages retained on
2.0-, 0.425-, and 0.075-mm sieves divided by 100).
where
G = absolute value of grade (percent) , The f values for the respective coefficients were - 14 .52 . 6.86.
PI = plasticity index (percent) , 6.75 , 14.22. 13.66, 13.72. and -13 .29.
SV = percentage of the surfacing material passing the 0.075- If a value of 100 counts/km is predicted for the change in
mm sieve, roughness, the actual value will , with 95 percent confidence.
NC = average daily car and pickup traffic in both direc- lie between 74 and 135 counts/km .
tions. The Brazil study (1 ,2) involved a program of extensive
NT = average daily truck traffic in both directions. and roughness measurements that resulted in the following model
R = radius of horizontal curvature (m). for the change in the natural logarithmic value of roughness
(LDQ) in terms of the QI value in counts/km:
HDM-III uses models developed from the data collected in
Brazil to obtain the following gravel-loss prediction equation:
LDQ = D{0.4314 - 0.1705(T2) + 0.001159(NC)
MLA = 3.65(3.46 + 0.246(MMP)(RF)
+ 0.000895(NT) - 0.000227(NT)(G)
+ (KT)(ADT)] (3)
+ S[-0.1442 - 0.0198(G) + 0.0062l(SV)
where - U.Ul42(Pl) - 0.00U617(NC)]} (5)
MLApredicted annual material loss (mm/year),
=
MMPmean monthly precipitation (m),
= where
RFroad rise plus fall (m/km).
=
KTmax{O; [0.022 + 0.969/57.300(KCRV)
= T2 = surfacing type dummy variable ( 1 for clay. 0 for
+ 0.00342(MMP)(P75) - 0.0092(MMP)(PI) other),
- O.lOl(MMP)]}, and NC = average daily car and pickup traffic in both direc-
KCRV = curvature (degrees/km) . tions,
NT = average daily bus and truck traffic in both directions.
One of the main advantages of the new model is its sim- G = absolute value of vertical grade (percent) ,
plicity compared to the existing models. Aspects such as the S = season dummy variable (0 for dry season. I for wet
vertical grade and horizontal curvature, which need to be season),
averaged for a road link, are excluded from the new model. SY = percentage of surfacing material passing the 0.075-
No estimate of the rainfall is necessary, and the necessary mm sieve, and
laboratory testing is minimal. All the parameters required can PI = plasticity index of surfacing material (percent).
be easily obtained by relatively unskilled staff in unsophisti-
cated laboratories. The model used in HDM-III was developed to constrain
the tendency of these models to overestimate roughness at
high levels under infrequent maintenance . The rate of rough-
Roughness Deterioration ness progression is decreased as roughness tends towards the
maximum for a particular material. The result of this exercise
The best model determined for southern African conditions was the following, somewhat complex, model :
(6) to predict the change in roughness in quartercar index
(QI) counts per kilometer was as follows: (6)
 Paig~-Green and Visser

where LRA = 1.4035 - 0.0239( W) - 0.0048(SV)

QI(TG 1) = roughness at time TG 1 (QI counts/km): + 0.01694(PI) + 0.6307(LRB) + 0.1499(Tl)
=
roughness at time TG 2 (QI counts/km):
Ql(T-G 2 )
= time elapsed since last grading (days):
TG 1 , TG 2 + 0.3096(T2) + 0.00020(NT) + 0.2056(BS)
= exp[c(TG 2 - TG 1)], 0 < b < l;
b - 0.01183(Pl)(BS) (8)
= - 0.001 [0.461 + 0.174 AOL + 0.0114 AOH
c
- 0.0287(AOT)(MMP)]: and where
QIMAX1 = max[279 - 421(0.05 - MGO;) + 0.22C
- 9.93(RF)(MMP); 150] (for Section j) LRA = natural logarithm of roughness after blading.
LRB = natural logarithm of roughness before blading.
where W = road width (m),
Tl = surfacing type dummy variable ( 1 if surfacing type
AOL = average daily light vehicle traffic ( <3500 kg) in
is quartzite, 0 if other). and
both directions;
BS = season during which blading occurred (0 if dry sea-
AOH = average daily heavy vehicle traffic (2:3500 kg) in
son, 1 if wet season).
both directions;
AOT = average daily vehicular traffic in both directions: HDM-III makes use of the following model:
MMP = mean monthly precipitation (m/month):
RF = average rise and fall of the road (m/km); QI 1•ftcr> = QIMIN1 + a(QI< 0010 ,c) - QIMIN,) (9)
C = degree of horizontal curvature (degrees/km);
MGOi = material gradation dust ratio defined as MGOi = where
1 if P4251 = 0, and MGOi = P075/P425i if P425i
> O; QI<•fler> = roughness after blading (QI counts/km).
P0751 = percent passing 0.075-mm sieve for Section j; and QI 1bdmel = roughness before blading (QI counts/km).
P425i = percent passing 0.425-mm sieve for Section j. a = 0.533 + 0.230 MGD;.
QIMINi = max{lO; min(lOO; 4.69 D95i(J - 2.78 MG,)]}
In order to constrain the prediction model at high roughness MG; = min(MGMi. 1 - MGM,. 0.36). and
values. the maximum value of the roughness is artificially MGM; = (MG075i + MG425; + MG02,)/3
limited to a QI of 150 counts/km. This value appears to be
too low for the average unpaved road in rural and developing where
areas in southern Africa, although it may be defined as an MG075i In (P075/95)/ln (0.075/D95,) if 095, > 0.4.
acceptable limit before blading is necessary. otherwise 0.3;
MG425; = In (P425/95)/ln (0.425/D95,) if 095, > 1.0.
Roughness After Blading otherwise 0.3; and
MG02i = In (P02/95)/ln (2.0/D95 1) if D95, > 4.0. other-
The change in roughness needs to be related to a datum for wise MG425 1.
use in a maintenance management system. For this purpose, The Brazilian and HDM-III models require significantly
a model to predict the roughness after blading (the starting more input, and the latter is especially cumbersome for devel-
point of roughness progression after each maintenance oper- oping areas. However, any of these models in association with
ation) was developed. The following model was the best the roughness progression models can be used to determine
obtained for southern African conditions (6): (a) the blading frequency necessary to retain the road rough-
ness between upper and lower limits as required. or (b) eco-
LRA = 1.07 + 0.699(LRB) + 0.0004(ADT) nomic efficiency by comparing the cost of maintenance with
- 0.13(DR) + 0.0019(LMS) (7) the road-user savings. However, the lower roughness limit is
strongly dictated by the particle size distribution and plasticity
(r 2 = 0.62; n = 1601; RMSE = 0.28; F = 650) of the wearing course material. The required roughness after
blading may not always be achievable on roads with exces-
where sively oversized material or inadequate plasticity (i.e., those
highly susceptible to the formation of rhythmic corrugations)
LRA = natural logarithm of roughness after blading (QI through normal maintenance procedures.
counts/km),
LRB = natural logarithm of roughness before blading (QI
Discussion of Results
counts/km),
ADT = average daily traffic in both directions,
An analysis of the applicability of the deterioration models
DR = dust ratio (ratio of percentage passing 0.075- and
for unpaved roads used in the different management systems
0.425-mm sieves), and
indicated that the models developed outside southern Africa
LMS = laboratory-determined maximum size (mm), not
were not always applicable to southern African conditions
greater than 75 mm.
(6). The average annual gravel losses for the sections moni-
The t values for the respective coefficients were 14.98, 45.99, tored during the experiment were predicted as 13.0 and 21.8
4.51, -2.67, and 3.23. The Brazilian model (J,2) for the mm by the Brazilian and HDM-III models, respectively; the
roughness after blading is as follows: actual measured value was 13.9 mm. The prediction of the
140 TRA :\'SPORTA TIO ,\ ' RESEARCH RECORD 129/

average roughness was even further out. being 96.1 and 66 .2 be selected. One of the criteria for this network was that all
QI counts/km for the Brazil and HDM-III models. respec- the information required for the old and the new performance
tively . Tpe actual average measured value was 80.0 counts/ models had to be available . In many instances. when perfor-
km . The new model predicted values of 13.1 mm and 77 .4 mance models are applied to a network of roads. estimates
QI counts/km for the average annual gravel loss and average of properties that are not very sensitive to the outcomes are
roughness. The Brazilian model predicts the annual gravel made. Consequently. no readily available network of roads
loss accurately but differs from the roughness prediction con- contained all the required information. A fictitious network
siderably. because the maximum size is not taken into ac- was developed. This network consisted of all the test sections
count (10). used to develop the new southern African performance models.
The Brazilian and HDM-III models are fairly complicated. A random number generator assigned a length to each link .
necessitating the determination of a number of geotechnical In this way. a network of 2662 .3 km comprising 77 links was
properties. identification of the material types . and estimation developed . Traffic ranged from 20 to 333 vpd. The material.
of the average vertical grade (rise and fall) and average degree traffic . and environmental characteristics used in the anal ysis
of road curvature. over the total length of the link. Because were those pertaining to the actual roads studied during the
these models were developed mainly for use in developing development of the models .
countries. where computing facilities are often rudimentary
(even at regional or head offices) and the skill levels of the
road personnel may be low. the usefulness of the models. EVALUATION OF THE MODELS' EFFECTS ON
especially in remote areas. is questionable. MANAGEMENT DECISIONS
A significant aspect of the new models is their simplicity.
The data required can be obtained quickly and cheaply using The effects of the old and new performance models on man-
relatively unskilled labor. The models developed in this study agement decisions were compared . Gravel loss . maintenance
eliminated the necessity to identify the material type and to budget, equipment requirements , overall network condition.
estimate the average grade and curvature for the road link . and upgrading requirements were studied . Analyses were car-
Simple indicator tests requiring minimal equipment and only ried out on the original MDS using the Brazilian models (old
basic operator training are required for the input parameters MDS) and the new southern African models (new MDS) .
for the model. The predictive capability, however, has not
been diminished through this process. In fact. it has generally
been improved for local conditions. Gravel Loss

MDS The thickness of gravel lost. calculated according to the old

MDS , ranged from 5.0 mm (the minimum permitted) to 27.6
The MDS was originally developed by Visser (J) using data mm; the gravel loss calculated by the new MDS ranged from
collected during the World Bank study in Brazil. The system 5.0 to 22.1 mm . The total volume of regraveling maten al
has subsequently been improved and adapted for use on per- required was 306 000 m 3 according to the old MDS and 259 000
sonal computers. An overview of its operation and use as a m3 by the new MDS. The new MDS thus predicted that 20
management aid has been fully described (1,10). percent less material needed to be replaced annually com-
Management personnel generally desire certain informa- pared to the prediction of the old MDS. Dy using the new
tion for the routine maintenance and upgrading of an unpaved MDS in southern Africa, budget requirements for gravel loss
road network. This information includes the following: would be reduced by 20 percent. At a gravel cost of RIO per
m3, the savings would be about R0.5 million. This outcome ,
• How much money should be budgeted for regraveling. which is different from that found by calculating the average
and what volume of material will be required annually? gravel loss, is the result of the different length and traffic
• What routine blading budget is required? characteristics of the network.
• What are the consequences if the required budget cannot
be provided?
• How many motor graders are required to perform the Maintenance Budget Requirements
maintenance for the selected budget?
• How often should every link be bladed to ensure optimal Total cost comprises the road-user cost and the road main-
economic allocation of the maintenance funds? tenance cost. The point where the total cost curve reaches a
• Which roads are economically justified to upgrade to bitu- minimum is the optimum maintenance position. This means
minous standard, and what funds are required for this? that one unit of saving in road-user cost is balanced by one
unit of cost to maintain the road . At this point, the marginal
All of these questions can be answered accurately and effi- benefit-cost (B/C) ratio is 1. By means of the MDS package.
ciently using the MDS. the minimum position was calculated for both models. For
the old MDS, the optimum maintenance cost (OMC) was
R5 .62 million and the road-user cost was R74.3 million. For
GENERATION OF ROAD NETWORK the new MDS , the OMC was R4 .76 million and the road-user
cost was R70.9 million.
To evaluate the effect of the new unpaved road performance These results corroborate the previous statements about
models on management decisions, a network of roads had to roughness progression. The rate of roughness progression is
Paige-(j reen and Visser

lower for the new model. which means that the network is in weighted average network roughness for the new MDS was
a better condition with less maintenance . Hence. the OMC 63 .2 QI counts/km. whereas the prediction for the old MDS
and road-user cost are lower for the new model. was 81.5 QI counts/km . Besides the lower predicted budget
The aim of maintenance scheduling is to ensure that main- and equipment requirements and better road conditio n of the
tenance is applied in such a manner that the economic advan- new MDS. significantly lower road-user costs would be incurred
tage is the same on all links . For a few links this can be done by the traveling public .
manually. but as soon as a network of roads is considered.
mathematical optimization techniques have to be applied . In
the MOS . dynamic programming techniques were used . The Upgrading Requirements
economic advantage. expressed as the marginal BfC ratio.
was computed for a range of maintenance budgets. The plots The MDS has the capability of determining when it is eco-
for the old and new models are shown in Figure 1. At the nomically justifiable . in terms of total costs. to upgrade an
OMC. the marginal BfC ratio is l. unpaved road to paved standard . For a given road under
Public authorities do not normally obtain sufficient funds consideration, the output is the year when paving is justified.
for maintenance to be applied at economic optimality. Public for a range of construction costs . This permits evaluation of
expenditure on maintenance generally lies between marginal the sensitivity of the construction cost.
BfC rati os of 3 and 5. For the purpose of this comparison. a The same 16 roads were selected for paving by both sets
marginal BfC ratio of 4.0 was selected. The required budgets of models during the following IO-year period for the con-
for blading were R3.521 million for the old MOS and R2 .898 struction costs used . Because of the slower deterioration of
million for the new MOS (i.e .. a 20 percent savings). In both the road predicted by the new models. however. the time of
instances. the inability to supply this budget would result in paving was generally set a few years later than that of the old
additional road-user costs of at least R4 for every Rl in main- models for the higher construction-cost scenarios. This time
tenance funds not provided. Again, the new MDS resulted schedule again represented savings to the road authority.
in a 20-percent reduction of the predicted budget require-
ments over those of the old MDS.
Discussion of Results

Equipment Requirements and Network Condition The economic advantages of maintenance management have
been shown in a number of countries. However. the economic
Because of the lower budget requirements of the new model. advantages of using the new models in the MOS for so uthern
fewer motor graders were required . Using the old MOS . 15 African roads are clear. On the network under discussion .
graders were necessary to maintain the network optimally: annual savings of Rl. l million. made up of R0 .5 million for
only 12 we re necessary on the basis of the prediction of the regraveling and R0.6 million for blading. are possible.
new MDS. Even though fewer graders were required . the Because this savings is on a network of only 2660 km. the
overall condition predicted by the new MDS was better. The potential savings on the numbered unpaved road network of

40

OLD MOS

30
Q
I-
<[
a:
I-
(fl NEW MOS
0
u
..... 20
!:::
...,
u..
...,z
CD

_J
<[
z 10
Ci
a:
<[
::!:

0
400 800 1200 1600 2000 2400 2800 3200 3600 4000 4400 4800 5200 5600 6000
MAINTENANCE BUDGET Rx 1000

FIGURE I Graph for determining the appropriate maintenance budget.

142 TRANS l'ORTA TI O /\' R ES EA RCH RECORD 1291

140 000 km are enormous. assuming the network used in this 2. A . T . Visser and W. R. Hudson . Pcrform;mce. Design and M;iin-
study was a representative sample . The factorial experimental tenance Relationships for Unpaved Low-Volum~ Roads. Jn
Transportation Research Record li98. TRB. National Research
design u~ed to select the original roads should ensure this. Council. Washington. D.C .. 1983. pp . 164-17-t.
3. T . Watanatada. C. G . Harral. W. D . 0 . Paterson. A. M. Dhar-
eshwar. A. Bhandari. and K. Tsunokawa. The Higlnl'lly Design
CONCLUSIONS and Maintenance Standards Model IHDM-IJJJ, Vol. 1. Descrip-
tion of the HDM-111 Model. Johns Hopkins Univcrsitv Press.
Baltimore. Md .. 1987.
The new performance-related models developed from south- 4. L. L. Parslev and R. Robinson . The TRRL Road /111·est111e11t
ern African roads permit more confidence in the predictions Model for Del'elopi11g Countries f R"rtM1 J. Lahoratory Report
of gravel loss and routine maintenance requirements for 1057 . Transport and Road Research Lahorator\'. Crowthurm: .
unpaved low-volume roads in southern Africa. The predic- U.K., 1982. .
5. P. Paige-Green . Some Surface Roughness. Loss and Slipperiness
tions also result in significantly lower budget requirements
Characteri tics of Unpaved Ro<lds. Proc.. /st /111ematio11a/ Sur-
compared to those based on the models developed in Brazil. face Characteristics Symposium. State College. Pa .. 1988.
On a small network, savings of 20 percent, or Rl. l million. 6. P. Paige-Green. The Influence of Geotec/111ical Properties 0 11 the
can be achieved. This amount alone is equivalent to the cost Performance of Gravel W earing Course Materials. Ph .D. disser-
of the research project. The results of this work also show tation, University of Pretoria. South Africa. 1989.
7. P. Paige-Green and F. Netterberg . Requirements and Properties
the danger of using models developed in a different envi- of Wearing Course Materials for Unpaved Roads in Relation to
ronment. Their Performance. In Tra11sportatio11 Research Record 1106.
TRB, National Research Council. Washington. D.C.. 1987. pp.
208-214.
8. P . Paige-Green. A Revised Meth od for /he Sie1 •e Analysis of Wear-
ACKNOWLEDGMENTS
ing Course Gravels for Unpaved Roads. Research Report DPVT8~ .
Division of Roads and Transport Technology. Council for Sci-
This paper is based on research carried out at the Division of entific and Industrial Research . Pretoria . South Africa. 1989.
Roads and Transport Technology and is published with the 9. H . H. Weinert. The Natural Road Construction Materials of
permission of the Director. Sowhem Africa . Hand R Academica. Cape Town. South Africa.
1980.
10. A. T. Visser and P. C. Curtayne. The Maintenance and Design
System: A Management Aid for Unpaved Road Networks. In
REFERENCES Transportation Research Record 1106, Vol. 1. TRB. National
Research Council. Washington. D.C.. 1987.
1. A . T . Visser. An Evaluation of Unpaved Road Performance and
Mai111enance. Ph.D. dissertation. University of Texas. Austin.
1981.