Materials and Methods

Materials & Method

Fig. 1 outlines the method used to assess LU transitions in Karnataka state using the Fuzzy AHP MCA modeling technique, considering various scenarios. The method incorporates (i) spatial data acquisition and processing, (ii) LULC extraction, and (iii) hybrid modeling.

Fig. 1. Method of LU transitions assessment

Study area

Karnataka state is located in the southwest region of India, bounded by Goa, Maharashtra, Telangana, Andhra Pradesh, Tamil Nadu, and Kerala with a spatial extent of 1,91,791 km². The region has 320 km of coastline with significant forest cover and a rich natural resources base, extending 760 km N–S (11˚34’ N and 18˚27’ N) and 420 km E–W (74˚3’ E and 78˚ 34’ E). Karnataka is demarcated into 30 administrative districts consisting of 178 sub-districts (taluks) for decentralized governance, comprising over 27,481 villages and 367 towns with a population of 64.06 million (density: 320 persons per km²) (Fig. 2). The forest ecosystem of Karnataka is unique and highly diverse. The different forest ecosystems have resulted from an interplay of topographic, climatic, and edaphic factors, influenced by altitude and distance from the sea. The forest types include tropical evergreen, semi-evergreen, moist deciduous, dry deciduous, thorny scrubs, sholas, and coastal mangroves. Karnataka state is a repository of rich biodiversity with more than 1.2 lakh known species, including 4,500 flowering plants, 800 fishes, 600 birds, 160 reptiles, 120 mammals, and 1,493 medicinal plants; half the biodiversity of Western Ghats is present in Karnataka. The forests support a wide range of endemic flora and fauna through a network of well-connected and protected wildlife sanctuaries and national parks. There are five national parks and 30 wildlife sanctuaries covering an area of 9,586.02 km² (Ramachandra et al. 2018). Apart from these, there are 15 conservation reserves; there is one community reserve comprising 652.369 km².

Fig. 2. Study area: Karnataka state, India

Data

The process of data acquisition involves primary data collection such as temporal RS data (USGS Earth Explorer (https://earthexplorer.usgs.gov/) has a repository of Landsat data starting from 1984 – to date), field data through sampling.

The ancillary data included cadastral maps (1:6000), topographic maps (1:50000 and 1:250000) of the Survey of India (SOI), and vegetation maps of South India (1:250000) developed by the French Institute of Pondicherry (Pascal 1986). Ground control points (GCPs), digitized from the topographic maps, were used to geo-register the scanned paper maps and geo-rectify RS data. Various forest cover types were digitized using the vegetation map of South India (1:250000) (Pascal 1986) to classify RS data of the 1980s. Other ancillary data included land cover maps, administration boundary data and transportation data (road networks). False colour composites (FCCs) helped in digitizing training data corresponding to heterogeneous patches distributed across the scene. Pre-calibrated GPS (global positioning system – Garmin GPS) units were used to collect attribute data of digitized training data (polygons), required for RS data classification and validation. This was supplemented with vector layers of ancillary point, line, and polygon data from virtual online spatial repositories such as Google Earth (http://earth.google.com) and Bhuvan (http://bhuvan.nrsc.gov.in). Table 1 lists the data used for analyses.

Table 1. Data used for assessing LU dynamics and description

The agents of LU transitions, given in Table 2, were digitized from reference maps as points, lines, and polygons. The proposed large-scale project details were compiled from the Infrastructure Development & Inland Water Transport Department, Government of Karnataka (https://idd.karnataka.gov.in/).

Table 2. Description of various agents and constraints considered for modeling

Method

Quantification of temporal ecosystem extent

The LU analyses involved: (a) downloading Landsat data of 2005; (b) procuring IRS MSS data for Karnataka state from the National Remote Sensing Centre (NRSC), Hyderabad (iii) geo-rectification of RS data scenes; (iv) generating FCCs of RS data (bands – green, red, and near-infrared (NIR), depending on the data LANDSAT/IRS) (Fig. 3). The FCCs aided in digitizing the training polygons corresponding to heterogeneous landscape elements; (v) digitizing the training polygons (uniformly distributed over the study region) covering 10% of the study region and loading these polygon coordinates into pre-calibrated GPS. The GPS aided in locating polygons during the field survey; (vi) collecting attribute data (LU) of these polygons from the field; (vii) augmenting attribute information corresponding to these training polygons from online portals (Bhuvan, Google Earth) and validating from the field using GPS. The training data was used for classification (60%), and the remaining 40% was used for accuracy assessment, and (viii) classifying LU through a supervised classifier considering training data (collected from the field) based on a Gaussian maximum likelihood (GML) algorithm.

Fig. 3. FCC and digitized training polygons across agroclimatic zones

The GML algorithm quantitatively evaluates both the variance and covariance of the category spectral response patterns when classifying an unknown pixel of RS data, assuming Gaussian distribution of data points. The statistical probability of a given pixel value being a member of a particular class was computed. After evaluating the probability of each category, the pixel was assigned to the most likely LU category/class (highest probability value). LU analysis was carried out using the free and open-source GRASS (geographical resources analysis support system) GIS software (http://wgbis.ces.iisc.ernet.in/grass/index.php); this is a robust support for processing both vector and raster data. The temporal, spatial data acquired from space-borne sensors were classified based on a GML classifier using available temporal ground truth information. The classification of RS data (of earlier times) was done using the information compiled from historical published vegetation maps, revenue maps, topographic maps, and land records.

The accuracy of LU classification was assessed through an error matrix (or confusion matrix) with overall (producer's and user's) accuracies and kappa (κ) statistics. The producer’s accuracy measures omission errors, and the user’s accuracy (UA) measures commission errors. Kappa compares two or more matrices and weighs cells in the error matrix according to the magnitude of misclassification.

Modeling of landscape dynamics

The modeling of landscape dynamics was carried out using the hybrid Fuzzy AHP MCA technique using temporal LU details and integrating agents with distance-based relationships of driving forces. Hybrid models aid in ascertaining criteria (n) and alternatives (m) in interactive decision-making by assessing the relative significance of two elements (criteria or alternatives) i and j for a given pairwise comparison matrix of the likelihood of events for all possible alternative ranking outcomes. This technique integrates expert knowledge of spatial data to determine the weight of each factor influencing land suitability classification, which helps to evaluate various decision-making policies. The urbanizing agents, constraints, and classified data were considered as base layers for modeling. The constraints exclude areas from being considered and are represented by Boolean images, where ‘0’ is assigned to areas that are not suitable, and ‘1’ is given to suitable areas for a type of LU. The constraints considered in this study include water bodies, forest zones, protected areas, and catchment areas based on regional and city development plans. The factors are assigned a relative degree of suitability represented by a fuzzy scale for each location. For example, land suitability for urban development is enhanced by a factor of proximity to roads. The driving factors used in modeling include road and railway networks, bus stops and railway stations, industries, educational institutes, places of religious importance, and service centers (such as hospitals, hotels, police stations, and shopping malls). Fig. 4 depicts the agents of LU changes, and Fig. 5 shows the constraints considered for modeling. The identification of these factors and constraints was made by developing attribute information using Google Earth data.

Fig. 4. Distance maps of various urban growth factors

Fig. 5. Constraints of LU changes

Prediction based on policy scenarios

Policy scenarios were considered for geo-visualization of likely LU in 2033, based on various complex factors provided by the development plans of the Karnataka region, 2033. These were business as usual (BAU), agent-based land use transition (ALT), reserve forest protection (RFP), afforestation (AF), and sustainable development plan (SDP) scenarios. Many researchers have successfully demonstrated the use of BAU scenario, which considers past and recent growth patterns and factors such as population, socio-economic trends and urban density (Samie et al. 2017; Hamad et al. 2018; Guzman et al. 2020). The scenario-based prediction is also known as the ‘what-if” prediction. Here, the simulation conducted is based entirely on the assumption of certain layers and factors, which implies ‘what will happen if factor-a, factor-b, and factor-c are considered” (Singh 2003). A knowledge of the impact of LU changes and possible placement of new infrastructure, green area conservation, regulation of zones, upgradation of existing transport corridors, and related changes in LU categories inside the perimeter of the study region will be highly significant for municipal authorities for urban planning. The details of chosen scenarios to visualize (using modelling) likely LU in 2033 are as follows.

Scenario 1 – Business as usual (BAU): BAU assumes that the historical trend of LU changes from 1985–2019 will continue during 2019-2033, without any change in the environmental and economic development policies.

Scenario 2 – Agent-based land use transition (ALT): This scenario evaluates the role of various drivers (agents) such as proposed (new) developments by regulatory agencies, existing industries, linear projects, urbanization, slope, core built-up areas, and special economic zones (SEZs).

Scenario 3 – Reserve forest protection (RFP): This scenario assumes a policy to protect reserve forests and protected areas from further degradation and allow growth in regions other than reserve forest areas.

Scenario 4 – Afforestation (AF): The AF (high conservation) scenario accounts for afforested areas by the Karnataka forest department. It assumes the protection and afforestation activities as a prime variable in the model to account for the policy decision.

Scenario 5 – Sustainable development policy (SDP): This scenario integrates reserve forest protection with stringent norms and afforestation of degraded lands (scenarios 3 and 4).

Table 3 gives details of the chosen policy scenarios with the modeling technique to visualize likely LU in 2033.

Table 3. Details of modeling techniques in the visualization of policy scenarios