A rule-based soil erosion model for a hilly
catchment
J. Adinarayana a,) , K. Gopal Rao b, N. Rama Krishna a ,
P. Venkatachalam a , J.K. Suri a
a
Centre of Studies in Resources Engineering, Indian Institute of Technology — Bombay, Powai, Mumbai
400076, India
b
Department of CiÕil Engineering, Indian Institute of Technology — Bombay, Powai, Mumbai 400076, India
Abstract
Catchment resources, generated from multi-source data Žremotely-sensed, map- and groundbased systems., were input to a raster-based Geographical Information Systems ŽGIS. after
geometrically co-registered to a standard grid Žpixel.. The GIS used was the PC-based, indigenously-developed package, called GRAM ŽGeo-Referenced Area Management.. A set of knowledge-based rules, for assessing the soil erosion of this heterogeneous hilly catchment, were
formulated from the knowledge of the multi-disciplinary resource-experts and the knowledge of
the local catchment resources, in addition to the field observations. This rule-based model, which
is hopefully fast and cost-effective and without effect of the individual bias, helped in inferring the
erosion intensity units that most likely to occur at any given pixel in the system. Finally, the
catchment was grouped into four different erosion intensity units namely very severe, severe,
moderate to severe and slight to moderate. The quantitative soil loss Žt hay1 yeary1 . ranges,
estimated by USLE model by a spatial information analysis approach ŽGIS., were also computed:
Ža. Very severe Ž) 50.: Žb. Severe Ž40–50.; Žc. Moderate to severe Ž20–40.; and Žd. Slight to
moderate Ž- 20.. An integrated priority delineation survey, for taking up soilrwater conservation
measures by Sediment Yield Index model, carried out in the catchment grouped the catchment into
very high, high, medium and low priority classes. Apparently, the areal extent of high and
moderate priority classes occupy about 80% of the catchment, which is closely following the areal
extent of severe and moderate to severe erosion intensity unit categories by this rule-based model.
The outcome of the results of this study will be immensely useful to monitor the dynamic changes
in the catchment environment. The soil erosion information system thus generated, identifies the
targeted problem areas for conservation planning and sustainable development of the catchment.
Although it is an early airing of results obtained from this rule-based model, the introduction of
multi-disciplinary expert’s informed opinion may provide an extension to the traditional methods
310
of assessing the soil erosion statusrmagnitude which should enable andror evaluate different
catchment management scenarios. q 1999 Elsevier Science B.V. All rights reserved.
Keywords: Geographical Information Systems; Hilly catchment; Integrated resources unit; Multi-disciplinary
knowledge-based rules; Rule-based approach; Soil erosion
1. Introduction
Soil erosion, by water, is the resultant of interplay between catchment environmental
factors such as soil, topography, drainage, rainfall and land use pattern, data upon which
are available from different sourcesrsystems. Additional information through experience
and knowledge of the multi-disciplinary experts also may be used for avoiding individual bias in assessing the soil erosion pattern of the area under interest. Hence, it is
important to study the soil erosion by analysing these multi-source catchment resources
with multi-disciplinary expertise in an integral manner. The problem is more serious
with hilly heterogeneous catchments which are subject to heavy rainfall. The technology
of Geographical Information Systems ŽGIS. provides a means of introducing information and knowledge from different sources in the decision-making process, and help in
handling and manipulating the multi-source data. A catchment is the base unit for the
study, being a complete hydrological and topographical unit.
Many researchers have been reported in the literature on this catchment environment
aspect using GIS ŽSpanner et al., 1983; De Roo et al., 1989—ANSWERS model; Bocco
and Valenzuela, 1991; Young et al., 1994—AGNPS model.. Expert systems in the form
of decision-rules have been used as a method of integrating multi-source data ŽMoore et
al., 1983; Lee et al., 1987; Skidmore et al., 1991; Harris and Boardman, 1998..
The aim of this paper is to describe for natural resources scientists a classification
method that interactively integrates conventional techniques, remote sensing, GIS and
the interpreter’s expertise and knowledge. Particularly, our objective is to generate a
GIS-assisted soil erosion map of a rain-infested heterogeneous hilly catchment in an
integral manner avoiding any individual bias.
1.1. Background
The approach was tested in a drainage basin of about 100 km2 in the difficult
semi-arid, mountainous terrain of the Western Ghats Ž73845X –73855X E. 19816X –19823X N..
The catchment, which has a basaltic landscape, is located in the Western Plateau and
Hill Agro-climatic Region of the Indian Peninsular.
Sediments are deposited in the drainage ditches, and ultimately in Yedgaon reservoir,
a terminus of drainage for the catchment. The siltation is largely due to poor soil
conservation practices, scantily distributed vegetation cover, steep slopes and overall dry
climate with an erratic monsoon rainfall, resulting in a high susceptibility to erosion.
Cultivation Žcropping systems based on upland rice. is being introduced in the canyons
extending into the steeper mountainous area with inadequate soil conservation practices.
If erosion is not stopped, further increases in runoff, and sheet and gully erosion on
311
slopping ground will ultimately destroy the productive value of the land. Hence, a
rule-based approach for assessing the soil erosion of this heterogeneous catchment is
required for conservation planning.
2. Data sources
Keeping in view the integrated approach to be adopted, the following data were used:
Remotely-sensed data: Landsat-TM and IRS-1A LISS II-digital ŽComputer Compatible Tapes-CCTs. and visual ŽFalse Colour Composites-FCC. products of monsoon
Žkharif. and post-monsoon Žrabi. agriculture seasons.
U
Survey of India ŽSOI. topographical maps at 1:250 000 and 1:50 000 scales.
U
Available literature and maps on various themes.
U
Meteorological data Žrainfall, temperature, etc...
U
Ground survey information on soil and other terrain characteristics of the sample
sites of the watershed.
U
3. Database construction
The GIS used was the PC-based, indigenously developed, user-friendly package
called Geo-Referenced Area Management ŽGRAM.. This GIS package is a spin-off from
the Indian Department of Science and Technology sponsored project ‘Natural Resources
Data Management Systems’, which aimed to generate computer-compatible, spatially
oriented databases of natural resources and socio-economic parameters to facilitate
area-specific micro-level planning. The GRAM package operates on a small and
relatively low-budget computer configuration, and can be used as a viable tool for
natural resources management, rural and urban planning, image analysis of remotely
sensed data, watershed management, impact assessment studies, etc. GRAM has been
built with seven major modules: input module, attribute link, analysis module, terrain
module, image processing module, statistics module and print module. The software is
arranged in a hierarchical structure with sub-menus at different levels. The capability of
the system is being tested for regional applications, and work is in progress to add on
further analysis functions and to interface well-known mathematical models for resources applications ŽSengupta et al., 1996..
The data sources were integrated into the GIS in a grid cell format. GRAM allows
rapid integration of data from image and map sources. Digitisation, georeferencing and
rasterization were done in the input module. The output of this module was used as the
base for the analysis module to integrate and process on logical combinations and
knowledge-based rules. The rules were constructed from the knowledge of the local
terrain parameters and field checks, in addition to the author’s expertise in various
fields. Integration of these rules were carried out with graphic Žspatial. database
Žloose-coupled way of integration. in the analysis module of GRAM. Using the query
support, the spatial database was analysed and the final soil erosion map was produced.
312
The erosion intensity map and the results were the output of the print module. The
general procedure followed during the analysis is shown schematically in Fig. 1.
3.1. Coordinate system
A defined cell Žpixel. was chosen, which is the limit of the resolution of the system.
Each cell has an exact location in space, related to the grid orientation and cell size and
a list of assigned attributes. The information database were geometrically corrected,
using a SOI topographic map base and resampled to a 30 m resolution Žof Landsat-TM..
A grid size of 515 = 492 pixels covered the whole catchment.
3.2. Catchment resources
Accurate information on the existing landuse pattern and its spatial distribution is a
prerequisite for any catchment management programme. With the advent of remote
sensing, it has been possible to prepare this dynamic resource map at various levels of
confidence. However, the applications of multi-spectral classification techniques in
extracting vegetation classes has been inconsistent because of similarities in spectral
responses ŽBocco and Valenzuela, 1991; Adinarayana et al., 1994.. The problem is more
serious with the digital techniques under heterogeneous hilly terrain. The other problem
in the collection of landuse data is the presence of clouds and their shadows during the
Fig. 1. Generalised processing flow for soil erosion information system of a hilly catchment.
313
monsoon period. In addition, it is also important to map the agricultural landuse pattern
of both monsoon Žkharif. and post-monsoon Žrabi. seasons together in the rural
catchments. Therefore, multi-seasonal remotely-sensed images of Landsat-TM digital
Žkharif—when most of the crops are matured and harvested. and IRS-LISS II FCC Žrabi
—cloud-free images showing vegetation in the upland areas. were used and classified
using the maximum likelihood classifier and standard visual interpretation techniques,
respectively. In addition, multi-thematic maps of soil, slope and drainage were also used
to generate multi-disciplinary knowledge-based rules to overcome the above inherent
problems in the hilly terrain. These maps were input and stored as separate layers in the
GIS for integrated analysis. The link between above GIS data layers and knowledge of
the multi-disciplinary group of Scientists and the field checks is provided by rules. Since
each satellite classification on its own was insufficient to discriminate all the necessary
classes, the rules were constructed by relating the GIS data layers and the aggregated
improved agricultural landuse pattern of the catchment to overcome some inherent
problems, viz. to discriminate the vegetation categories and to identify the information
classes under the cloudrshadow regions in kharif season ŽAdinarayana and Rama
Krishna, 1996.. Given that all the data sources were now in a common coordinate
system, a number of logical combinations could be performed on a per pixel basis to
consolidate the agricultural landuse pattern of this heterogeneous highland. To give a
few examples:
Ža. If Landsat-TM ŽKharif class. s Thick Forest and IRS LISS-II ŽRabi class. s Thick
Forest; Then class s Thick Forest.
Žb. If Kharif class s Sparse Forest and Rabi class s Open Žwithrwithout scrub. Area
or Degraded Pasture; Then class s Open Žwithrwithout scrub. Area or Degraded
Pasture, respectively.
Žc. If Kharif s Sparse Forest and Rabi s Cropped; Then class s Cropped.
Žd. If Kharif s Cloud Cover and Rabi s Thick Forest; Then class s Thick Forest.
Že. If Kharifq Rabi s Vegetation, Drainage densitys Low, Soil s Typic Chromusterts and slope s- 58; Then class s Cropped.
These informed-opinions, codified in the form of ‘if–then’ type of rules, were used to
refine the classification accuracy. Initially, the correspondence between the LISS
II-based visual and TM-based digital classification maps was at best 45%. After
applying the rules on the classification output, the agreement between the maps
increased to 85–100% ŽAdinarayana and Rama Krishna, 1996.. In the process, the
visual and digital classification techniques interacted to improve the classification
system. Finally, the GIS-assisted improved landuse map consists of five major categories, which were selected from the conservation planning point of view: agriculture,
thick forest, sparse forest, degraded pastures and open Žwithrwithout scrub. lands. The
introduction of multi-disciplinary expert rules provides an extension to the traditional
methods of satellite-derived classification which should enable more relevant, and
accurate, landusercover map to be prepared.
A collective approach comprising satellite data ŽIRS. and topographic maps were
used to delineate physiographic units. The soil associations of each physiographic unit
were established by studying soil profiles in the landscape and by random checking
outside the sample strip areas, and finally classifying the soils according to an estab-
314
lished soil taxonomy ŽSoil Survey Staff, 1975.. In the end, a soil map at 1:50 000 scale
was prepared by combining the soil information in each delineated physiographic unit.
The contours and high points were digitised from 1:50 000 SOI topomaps Ž20 m
interval. to generate the Digital Elevation Model ŽDEM. and the slope map with the
help of the GRAM package. The catchment was divided into three major slope steepness
classes for the evaluation of soil erosion and priority sub-watersheds delineation survey:
Ž1. gently sloping lands Ž0–88.; Ž2. moderately sloping lands Ž8–188.; and Ž3. steeply
sloping lands Ž) 188.. Steeper slopes promote concentrated run-off and show exposed
rocks in some places. Moderate slopes support mostly degraded pastures or open scrub
lands and have moderate to high run-off potential. Gently sloping lands covered with
moderately deep cultivated black soils have low run-off potential. Once the contours are
digitised, GRAM allows one to categorise any number of slope steepness classes as
needed.
An isohyetal map was compiled manually from 10 years of mean annual rainfall data
from six meteorological stations available in and around the catchment area. The
catchment was divided into two rainfall zones: a - 1500 mm low rainfall zone, and
a G 1500 mm high rainfall zone.
The drainage network was generated from SOI toposheet Žof 1975. and the high order
river course systems were updated from Landsat-TM imagery Žof 1989.. The drainage
density Žtotal length of streamsrunit area. was calculated for each sub-watershed and
the catchment was divided into low, medium and high drainage density zones. The
proximity analysis was done to establish the buffer zone along the streams to demarcate
the zones where erosion was likely to be important.
The catchment was divided into 33 sub-watersheds, which are smaller hydrologic
units and constitute individual tributaries of the lowest order, or a group of such
tributaries. The area of each sub-watershed was planometrically computed and ranged
from 1.4 to 7.0 km2 , which is considered as a viable working size for conservation
planning.
All the previously described thematic maps were digitised Žexcept the TM-based
digitally classified data which will be in the raster form., rasterized and brought to a
common coordinate system, which thus acted as database for GIS analysis to generate
the erosion map.
4. Database analysis
One of the most important characteristics of GIS is its capability for data analysis and
spatial modelling. GRAM includes conventional GIS data analysis and manipulation
capabilities, such as overlaying, reclassification, etc. The system incorporates a set of
decision-rules in which several knowledge-based rules are used to evaluate and input
data at different levels of abstraction.
4.1. Generation of integrated resources units
Given that all the data layers are now spatially part of a common coordinate system, a
number of useful combinations can be performed. The first step in this process is to
315
create a basic unit by overlay analysis, and create Integrated Resources Unit ŽIRU.. Each
IRU comprises the catchment resource data on soil, slope, landuse, rainfall distribution
and drainage buffers for one grid cell Ž30 m.. One IRU has been taken as the basic
strategic unit for assessing the soil erosion. Since they exhibit strong uniformity, they
can all be expected to respond similarly to given intensities of human use and
management strategies.
Each IRU descriptor was derived from catchment factors superimposed as layers in a
GIS. Adjacent sets of IRU attributes were then separated. However, unlike photomorphic boundaries traced from aerial photographs or satellite images, an IRU in this system
cannot be printed in map form, so each pixel was geocoded to a standard SOI topomap.
In other words, the IRU can be ‘delineated’ in spatial domains using digital data and
computer processing, since each pixel acts as an IRU.
Taking all the catchment environmental factors and their attributes into consideration,
270 IRU combinations should occur in the catchment, whereas only 228 IRU combinations appeared in the catchment and are considered in our overlay analysis.
4.2. Knowledge-based rules
Avoiding individual bias, in the form of giving equal weightings to each soil erosion
parameter Žsoil, slope, drainage, rainfall and landuse., and in logical combinations, is the
main basis for the rule-based approach in the study. The approach is more useful in
complicated erosion-prone heterogeneous fragile ecosystems of hilly catchments which
receives heavy rainfall. Hence, a top–down multi-disciplinary knowledge-based rules
were formulated to map the soil erosion risk zones of this hilly catchment.
Rules concerning environmental relationships cannot normally be expressed with
absolute certainty Žtrue or false.. In other way, a rule lies in a continuum between true
Žprobability of unit. and false Žprobability of zero., depending on how sure we are that
the rule is true Žor false.. In this study, the link between the GIS database layers and the
soil erosion intensity unit is provided by information-based rules. In the implementation
of these rules, the multi-disciplinary expertise and knowledge of the local terrain
parameters and field observations were taken into consideration, in addition to the
authors expertise in soils, land-use planning, remote sensing and GIS. In the present
study, the rules were expressed as the probability of an item of evidence occurring Žif
- 58 slope; other drainage area; Typic Chromusterts; - 1500 mm rainfall with thick
forest. given a particular hypothesis Žthen that pixel is slight to moderate erosion risk..
The rules are the most subjective aspect of the information-based system: they are
heuristic, estimated from the informed opinion of the experts ŽSkidmore et al., 1991..
The well-developed and moderately deep Ž50–90 cm. black cultivated soils generally
occupy the valleys and plains of the catchment. Undulating uplands and subdued hills
have degraded pastures or open areas and partly cultivated lands with shallow soil depth
Ž20–50 cm.. Undeveloped and very shallow Ž- 20 cm. soils cover the hilly regions with
steep slopes and are mainly forest areas. The drainage density, in general, will be high,
moderate and low in cases of forests, degraded pasturesropen Žwithrwithout scrub.
areas and agriculture regions, respectively. These general natural conditions, in addition
to the useful logical combinations of catchment environmental factors, were taken into
316
consideration in formulating the knowledge-based rules to map the soil erosion risk
zones of this heterogeneous hilly catchment.
The knowledge-based rules for each IRU and their probability of erosion risk factor
have been worked out. Two contrasting examples are:
Ža. If 58 slope, other drainage area, Typic ChromustertsU , - 1500 mm rainfall with
thick forest; then class s slight to moderate erosion risk.
Žb. If ) 188 slope, 1–2 drainage orders, Typic UstorthentsUU , ) 1500 mm rainfall
zone with open areas; then class s very severe erosion.
ŽU . Piedmont black soils-deep, poorly drained, silty clay to clay, calcareous. ŽUU . Hilly
greyish brown soils-shallow to very shallow, well drained, gravely sandy loam.
Thus, these simple and useful logical combinations on a per pixel basis, enable
mapping of the erosion intensity that is likely to occur at any given location, without any
subjective bias. Finally, the soil erosion map was generated with four categories: slight
to moderate, moderate to severe, severe and very severe soil erosion intensity units.
An attempt has been made to use the multi-source data and GIS to estimate the soil
loss by USLE ŽWeischmeir and Smith, 1978. for environmental conservation planning
of this hilly catchment ŽAdinarayana and Gopal Rao, 1996.. The soil loss was computed
for each pixel as per the attributes and their values, and summed up for each
sub-watershed independently making use of the sub-watershed map. The soil loss yield
of all the sub-watersheds were added to obtain the total sediment yield of the catchment.
The rate of soil loss by USLE method of the catchment was 28.11 t hay1 yeary1 and
closely matching with the field observed value of 24.67 t ha yeary1 . The range of soil
loss values Žtons per hectare per year. for the erosion intensity categories was: Ž1. slight
to moderate Ž- 20., Ž2. moderate to severe Ž20–30., severe Ž30–50., and very severe
Ž) 50..
An integrated priority delineation survey, for taking up preferential soilrwater
conservation measures, by Sediment Yield Index ŽSYI. model ŽAdinarayana et al.,
1995., carried out in the catchment, grouped the catchment into very high, high, medium
and low priority classes. Apparently, the areal extent of high and medium priority
classes occupy about 80% of the catchment, which compares well with the occurrence of
severe and moderate to severe erosion intensity unit categories by this rule-based model.
5. Conclusions
The case study presented here suggests that the use of extensive multi-disciplinary
knowledge-based rules helps in mapping the spatial pattern of magnituderseverity of
soil erosion without suffering from individual bias by giving equal weightings to each
catchment environmental factor. The soil erosion information system thus generated,
identifies the targeted problem areas for catchment conservation planning. The methodology described will be more useful in hilly catchments with heterogeneous landscape
features where the rainfall erosivity is a major environmental threat to the sustainable
development of the area.
The raster based GIS provides an easy method of integrating the catchment resource
data into useful maps. Once in raster form, per pixel logical combinations are simple and
could almost be handled in a standard database.
317
The knowledge bases, with further research and testing on different and larger areas,
generated for the purpose of soil erosion assessment could lead to form, in future,
decision-support systems for conservation planning on catchment basis.
Sophisticated methodologies, such as image processing and GIS, are useful tools for
the assessment of soil erosion at regional scales. The model could be further improved
by finer resolution terrain analysis, incorporating many more ecological variables and
informed opinions from the experts and local people Žthe experiential knowledge-bases.,
and employing effective extrapolation of expert systemrneural network analysis. Eventually, with this improved approach, the local governmental bodies should be able to
possess a highly versatile and responsive system that enables their resource managers to
assess the breadth and depth of their needs in a geographically precise and consistent
manner.
The study also ensures the ability to monitor the dynamic changes of the catchment,
particularly the landuse pattern by remote sensing data with periodic intervals, and to
redefine the conservation planning strategies on the basis of detected change. This
rule-based approach for hilly catchments with variable rainfall regimes and severe land
degradation problems is likely to provide more sustainable and cost-effective conservation management strategies.
The methodology attempted in the case study will be further tested in the other areas
of hilly catchment ecosystem in the Western Ghat mountainous zone. Although it is an
early airing of results obtained from the methodology described, the introduction of
multi-disciplinary expert’s informed opinion may provide an extension to the traditional
methods of soil erosion which should enable us to generate andror evaluate different
conservation scenarios.
Acknowledgements
This research project was financed by a research grant from the Department of
Science and Technology, Government of India Žthrough SERC ProgrammeESr23r077r89.. The authors wish to thank CSRE, for providing facilities. Thanks are
also due to our colleagues at CSRE, particularly Dr. G. Venkataraman ŽGeologist., for
useful help, discussions and comments on this paper.
References
Adinarayana, J., Gopal Rao, K., 1996. Prioritisation of watersheds based soil erosion studies–A remote
sensing approach. Project report, Centre of Studies in Resources Engineering, IIT-Bombay.
Adinarayana, J., Rama Krishna, N., 1996. Integration of multi-seasonal remotely-sensed images for improved
landuse classification of a hilly watershed using Geographical Information Systems. International Journal
of Remote Sensing 17 Ž9., 1679–1688.
Adinarayana, J., Flach, J.D., Collins, W.G., 1994. Mapping land use patterns of a river catchment using
Geographical Information Systems. Journal of Environmental Management 42 Ž1., 55–61.
Adinarayana, J., Rama Krishna, N., Gopal Rao, K., 1995. An integrated approach for prioritization of
watersheds. Journal of Environmental Management 44, 375–384.
318
Bocco, G., Valenzuela, C.R., 1991. Integration of GIS and remote sensing in land use and erosion studies. In:
Valenzuela, C.R. ŽEd.., Introduction to GIS. ITC Publication No. LIS-22, Enschede, The Netherlands, pp.
192–205.
De Roo, A.P.J., Hazelhoff, L., Burrough, P.A., 1989. Soil erosion modelling using ANSWERS and GIS. Earth
Surface Processes and Landforms 14, 517–533.
Harris, T.M., Boardman, J., 1998. Alternative approaches to soil erosion prediction and conservation using
Expert Systems and Artificial Neural Networks. In: Boardman, J., Favis-Mortlock, D.T. ŽEds.., Modelling
Soil Erosion by Water. Springer, NATO-ASI Series I-55, Berlin, pp. 461–478, in press.
Lee, T., Richards, J.A., Swain, P.H., 1987. Probabilistic and evidential approaches for multisource data
analysis. IEEE Transactions on Geoscience and Remote Sensing GE 25, 283–293.
Moore, D.M., Lees, B.G., Davey, S.M., 1983. A new method for predicting vegetation distributions using
decision tree analysis in a GIS. Journal of Environmental Management 15, 59–71.
Sengupta, S., Patil, R.L., Venkatachalam, P., 1996. Assessment of population exposure and risk zones due to
air pollution using geographical information systems. Computers, Environment and Urban Systems 20 Ž3.,
191–199.
Skidmore, A.K., Ryan, P.J., Dawes, W., Short, D., O’Loughlin, E., 1991. Use of expert system to map forest
soils from a GIS. International Journal of Geographical Information Systems 5 Ž4., 431–445.
Soil Survey Staff, 1975. Soil Taxonomy—A basic system of soil classification for making and interpreting soil
survey. USDA Handbook No. 436, Washington DC.
Spanner, M.A., Strahler, A.H., Estes, J.E., 1983. Soil loss prediction in a GIS format. Proceedings of 17th
International Symposium on Remote Sensing of Environment. Ann Arbor, MI, 9–13 May 1983, pp.
89–100.
Weischmeir, W.H., Smith, D.D., 1978. Predicting rainfall erosion losses—A guide to conservation planning.
United States Department of Agriculture Handbook No. 537. Washington DC.
Young, R.A., Onstad, C.A., Bosch, D.D., Anderson, W.P., 1994. Agricultural Non-Point Source Pollution
Model, version 4.03. AGNPS User’s Guide Žftp: soils.mrsars.usda.gov..