2.3 Datasets

This section summarizes the data source, spatial resolution and other specification details of input geospatial data used for modelling travel time and geographic coverage in the region. All input datasets were clipped to the administrative boundaries of the district, as derived from NASA’s Socioeconomic Data and Applications Center. Data layers were projected consonantly into the spatial reference frame, WGS84/UTM Zone 44 N and all rasters used in travel time analysis were attuned specifically at 30-mtrs resolution.

Health facilities represented in study site are stratified as: i) Second tier District hospital and Community health centers(DH and CHCs) providing secondary healthcare ii) First tier Primary health centers(PHCs) serving as first point of contact between population and qualified doctors and iii) Sub-centers operating as peripheral health institution available to rural population. The health system of entire district comprises of 137 subcenters, 44 primary health centers (including upgraded), 3 community health centers and 1 district hospital. All of these facilities provide ambulatory and immunization care, whereas, only 34 and 28 facilities are equipped for service provisioning of delivery and inpatient related services. Geographic coordinates of all the facilities (accuracy of ± 10 meters) were collected by conducting field visit to these facilities. A vector layer of spatial location of facilities was created which was superimposed into final land-cover grid. One facility was located on the cell considered to be waterbody, which was then, manually moved to nearest cell.

Land use and land cover mapping for 2017–18 on 1:250K scale using multi-temporal Resourcesat- 2 terrain corrected Linear Imaging Self Scanning Sensor (LISS-III) satellite data was obtained from National Remote Sensing Center, Indian Space Research Organization upon request. This raster dataset is constructed by conflating 3 seasons–Monsoon–Kharif: August-October, Post Monsoon–Rabi: December–March and Pre Monsoon–Zaid: April–May. The land-cover data comprised 18 land cover classes which was reclassified into 7 major generic classes—i) Forest land ii) Grassland iii) Cropland iv) Settlement/Built area v) Wetlands vi) Waste-land/Fallow/Other land-cover and vii) Snowcapped land. Road network dataset, obtained from Open Street Map and further digitized using Google Earth was reclassified into road classes encompassing i) National Road ii) Secondary Roads and iii) Local tertiary Roads. Tertiary roads were further prorated into i) Four wheeler passable roads ii) Two wheeler passable roads and iii) Walking only roads. The data was rectified to connect segments of roads omitted through digitization and deleting those extending into waterbodies. Rivers and waterbodies mask embodying physical barriers was extracted from Land use and land cover map. Both road network and river dataset were rasterized to 30 meters gridded cells and were then superimposed on land-cover raster dataset to create merged land-cover dataset with 12 unique land feature classes.

It is pertinent to consider DEM for analyzing movement of patients across varying topography. The DEM is used as the reference grid for each individual analysis and the extent, projection and resolution of all other layers used in the project hinges upon DEM. Raster format layer of Digital elevation model (DEM) extracted from CartoSAT-1, a high resolution satellite data was retrieved from Bhuvan website [13] for 2.5 m spatial resolution in track stereo. The resolution of raster layer was then changed to 30m using resampling technique of Cubic convolution interpolation.

The gridded population distribution of 1 hectare estimated in continuous raster surface was created by dasymetric modelling approach. The design conflated detailed census estimates for 2015 (Projected from Indian census, 2011) and widely available, remotely sensed and geospatial ancillary data in a flexible Random Forest estimation technique (model version 2c) following methodology described in [14] and [15]. A Random Forest algorithm was used to generate gridded population density estimates that were subsequently used to dasymetrically disaggregate population counts from administrative units into grid cells of 1 hectare spatial resolution. Population density response variable and legion of covariates were computed at administrative level which was then used to fit Random Forest model for predicting population density at grid cell level (generating dasymetric weighting layer in conjunction with covariates coalescing from following ancillary data–i) LULC raster layer at 30 m spatial resolution ii) Lights at night raster data at 15 arc- second derived from imagery collected by Suomi National Polar-orbiting Partnership (NPP) Visible Infrared Imaging Radiometer Suite (VIIRS) sensor iii) Mean annual temperature and mean annual precipitation raster data from WorldClim/BioClim 1950–2000 at 30 arc-second, mosaicked and subset to match extent of land-cover data of the region iv) Buildings, residential and infrastructure vector layer from Open Street Map v) Protected areas vector layer from World Database on Protected Areas vi)River and waterbodies network vector layer vii) Road network vector layer viii) Built-area raster layer ix) Elevation and derived slope raster layer. These collated set of covariates were used for model fitting and prediction of the final layer.