Remote Sensing for GIS, Spatial Data Modeling and Environmental Management
Open in the Padhaiko app6.1 Spatial Data Modeling in GIS
Integrating remote sensing products into a Geographic Information System (GIS) requires representing both discrete human/natural features and continuous field measurements in mathematically consistent spatial data structures, coordinate systems, and indexes.
Vector and Raster Data Structures
| Aspect | Vector Data Structure | Raster Data Structure |
|---|---|---|
| Spatial Primitives | Discrete geometric objects: points , polylines (ordered vertices), and polygons (closed rings) | Regular 2D grid (matrix) of uniform cells (pixels) indexed by row and column |
| Attribute Storage | One feature ID linked to a multi-column relational attribute table | Each grid cell holds a single numeric DN, reflectance, elevation, or class code |
| Topology & Precision | Explicit adjacency, connectivity, and containment; exact sub-meter boundary precision | Implicit spatial adjacency via grid indices; precision bounded by cell size () |
| Ideal RS / GIS Use | Road networks, administrative parcels, river centerlines, training ROIs | Satellite imagery, DEMs, NDVI surfaces, land-cover grids, continuous interpolation |
| Overlay & Compute | Vector polygon intersection is computationally complex | Cell-by-cell map algebra is fast and directly vectorizable |
Coordinate Reference Systems: Geographic vs. Projected, UTM, and WGS84
To overlay satellite imagery with field GPS points and cadastral maps, every dataset must carry an unambiguous Coordinate Reference System (CRS):
- Datum & WGS84: A geodetic datum defines the size, shape, and origin of a reference ellipsoid approximating Earth's geoid. WGS84 (World Geodetic System 1984) is the global Earth-centered, Earth-fixed (ECEF) datum used by GPS/GNSS (semi-major axis , flattening ).
- Geographic Coordinate System (GCS): Locates points on the 3D ellipsoid using angular coordinates—latitude () and longitude () in decimal degrees (e.g., EPSG:4326 for WGS84). Angular units are non-uniform in ground length ( of longitude shrinks as toward the poles), so GCS is unsuitable for direct distance or area calculations.
- Map Projections & Projected Coordinate Systems (PCS): A map projection mathematically transforms onto a flat 2D plane in linear units (meters). Because a curved ellipsoid cannot be flattened without distortion, projections trade off four properties:
- Conformal: Preserves local angles and shapes (e.g., Transverse Mercator).
- Equal-Area (Equivalent): Preserves true surface area (e.g., Albers Equal-Area Conic).
- Equidistant / Azimuthal: Preserves true distance or direction from specific reference lines/points.
Universal Transverse Mercator (UTM) Coordinate System
UTM divides the globe between and into longitudinal zones, each wide, projected using a secant Transverse Mercator cylinder (scale factor along the central meridian to keep scale distortion below across the zone):
- Zone Number () from Longitude (, east positive, west negative):
- Central Meridian () of Zone :
- False Easting and False Northing (to keep coordinates positive):
- False Easting: at the central meridian.
- False Northing: in the Northern Hemisphere; at the equator in the Southern Hemisphere.
- (For example, Nepal spans UTM Zones 44N and 45N on WGS84: EPSG:32644 and EPSG:32645.)
Spatial Indexing Using R-Tree Concepts
Searching millions of vector footprints or satellite scene polygons sequentially () is too slow for interactive GIS. An R-tree (Rectangle-tree) is a height-balanced spatial index that groups nearby objects hierarchically by their Minimum Bounding Rectangles (MBRs):
- Leaf nodes store the MBR and pointer for each individual geometry (or raster tile).
- Internal nodes store a parent MBR enclosing all child MBRs underneath it.
- Spatial Query Pruning: During a bounding-box intersection or point-in-polygon query, the search descends only into branches whose MBR intersects the query window, skipping entire subtrees in average time. Exact polygon geometry tests are then run only on the small candidate set returned by the MBR filter.
Raster Map Algebra
Map algebra (Dana Tomlin) expresses spatial raster analysis as algebraic operations on gridded layers ():
| Class | Spatial Scope | Mathematical Form / Mechanism | Remote Sensing & GIS Example |
|---|---|---|---|
| Local | Single cell across or more stacked rasters | NDVI band math, image differencing, slope thresholding | |
| Focal | Neighborhood window around (e.g., ) | Low-pass smoothing, Sobel edges, terrain slope/aspect from DEM | |
| Zonal | All cells sharing the same zone ID in a zone raster | Mean NDVI or total burned area per watershed or district polygon | |
| Global | Entire raster extent simultaneously | Euclidean distance to roads, cost-distance, watershed flow accumulation |
Every georeferenced raster links pixel indices to map coordinates via a 6-parameter affine geotransform:
where is the upper-left corner, is pixel width, and is negative pixel height for north-up images ().
6.2 Modern Geospatial Architectures
Modern Earth-observation archives hold petabytes of imagery on cloud object storage (S3, GCS). Downloading whole ZIP archives before processing does not scale; cloud-native geospatial architectures allow clients to discover, stream, and compute on exact regions of interest in place.
Cloud-Optimized GeoTIFFs (COGs)
A Cloud-Optimized GeoTIFF (COG) is a standard backward-compatible GeoTIFF file organized internally for efficient HTTP GET Range requests:
- Header & Image File Directories (IFDs) first: Placed at the beginning of the file so a client reads geospatial metadata and tile byte offsets in one tiny initial request.
- Internal Tiling: Pixels are stored in rectangular blocks (e.g., or ) rather than full scanline strips, allowing a client to fetch only the tiles covering a small study area.
- Internal Overviews (Pyramids): Downsampled resolutions () are embedded inside the same file so zoomed-out web maps read low-resolution tiles rapidly without scanning full-resolution data.
SpatioTemporal Asset Catalogs (STAC)
STAC is an open JSON specification for indexing and discovering Earth-observation assets across space, time, and sensors:
- STAC Catalog: Top-level directory linking sub-catalogs, collections, or items.
- STAC Collection: Extends Catalog with dataset-wide spatial/temporal extents, license, and summaries (e.g.,
sentinel-2-l2a). - STAC Item: A single GeoJSON
Featurerepresenting one scene/granule with geometry footprint, acquisitiondatetime, sensor properties (eo:cloud_cover), and links to individual Assets (e.g., COG URLs for Red, NIR, QA mask). - STAC API: RESTful search endpoint supporting spatiotemporal queries (
bbox,datetime,query={"eo:cloud_cover": {"lt": 15}}) that return matching COG URLs for direct streaming.
Multidimensional Geospatial Arrays
Time-series remote sensing, climate reanalysis, and hyperspectral cubes are naturally modeled as labeled -dimensional arrays (data cubes) with dimensions such as (time, band, y, x). Unlike plain NumPy arrays indexed only by integer positions, multidimensional geospatial formats (Zarr, NetCDF4/HDF5) attach named coordinate axes (timestamps, wavelengths, projected ), physical units, CRS metadata, and chunking layouts for parallel out-of-core computation.
Array Processing with Python: rasterio, geopandas, and xarray
| Library | Primary Abstraction | Core Capabilities |
|---|---|---|
rasterio | GDAL-backed raster datasets (DatasetReader) | Windowed/COG reading via HTTP Range requests, affine transforms, CRS reprojection, rasterization, and writing GeoTIFFs |
geopandas | Spatial tabular GeoDataFrame (Shapely + Pandas) | Reading/writing Shapefile/GeoPackage/GeoJSON, CRS reprojection (to_crs), spatial joins, buffers, and polygon clipping masks |
xarray | Labeled -D DataArray and Dataset (often with rioxarray & Dask) | Dimension-aware indexing (.sel(time=..., band=...)), temporal composites (.resample, .groupby), and lazy chunked array math |
import geopandas as gpd
import rasterio
from rasterio.mask import mask
import xarray as xr
# 1. Read district boundary vector and reproject to raster CRS
aoi = gpd.read_file("study_area.gpkg")
with rasterio.open("https://example.org/sentinel2_b04_cog.tif") as src_red:
aoi_proj = aoi.to_crs(src_red.crs)
red_arr, _ = mask(src_red, aoi_proj.geometry, crop=True)
with rasterio.open("https://example.org/sentinel2_b08_cog.tif") as src_nir:
nir_arr, _ = mask(src_nir, aoi_proj.geometry, crop=True)
# 2. Wrap into labeled xarray DataArrays and compute NDVI
red = xr.DataArray(red_arr[0].astype("float32"), dims=("y", "x"))
nir = xr.DataArray(nir_arr[0].astype("float32"), dims=("y", "x"))
ndvi = (nir - red) / (nir + red + 1e-6)
print("Mean AOI NDVI:", float(ndvi.mean()))6.3 Environmental Management Applications
Multi-Temporal Land-Use/Land-Cover (LULC) Monitoring
Tracking LULC dynamics (deforestation, urban sprawl, glacial retreat, agricultural expansion) compares co-registered, radiometrically normalized multi-date imagery:
- Image Differencing (Algebraic Change Detection): Subtracts a single band or spectral index between two dates (, e.g., ). Pixels exceeding from the difference histogram mean flag probable change, though it does not identify which land-cover transition occurred.
- Post-Classification Change Detection: Independently classifies imagery at dates and into the same thematic classes, then overlays the two maps pixel-by-pixel to construct a Change Matrix (Transition Matrix) :
- Entry gives the area (or pixel count) that belonged to class at and transitioned to class at .
- Diagonal entries represent persistence (unchanged area); off-diagonal entries () specify exact "from–to" conversions (e.g., Forest Cropland).
- Row sums give totals (); column sums give totals (); net change for class is .
- (Caveat: Classification errors from both dates compound—if each map has accuracy, joint two-date accuracy is roughly .)
Disaster Assessment: Flood Mapping and Wildfire Burn Severity
- Flood Inundation Mapping via SAR Backscatter Thresholding: Because floods occur under heavy monsoon clouds, optical sensors are frequently blocked. Synthetic Aperture Radar (SAR) penetrates clouds day and night. Calm open floodwater acts as a specular reflector, bouncing radar pulses away from the side-looking antenna and returning very low backscatter (typically below in VV/VH).
- Pre- and Post-Event Analysis: Comparing a post-flood SAR image against a dry-season baseline excludes permanent water bodies (rivers, lakes) and isolates newly inundated floodwater pixels via Otsu or empirical thresholding plus slope masking (removing radar shadows on steep hillsides).
- Wildfire Burn-Severity Assessment using NBR: Healthy vegetation reflects strongly in NIR and weakly in SWIR2, whereas charred ash and exposed dry soil reflect poorly in NIR and strongly in SWIR2:
| Range | Burn Severity Class | Ecological Impact |
|---|---|---|
| Enhanced Regrowth | Post-fire vegetation recovery exceeds pre-fire greenness | |
| Unburned | No detectable fire damage | |
| Low Severity | Surface litter burned; canopy largely intact | |
| Moderate Severity | Partial canopy scorch and understory consumption | |
| High Severity | Complete canopy mortality and deep soil charring |
Thermal Remote Sensing, LST, and Urban Heat Islands
Thermal infrared (TIR) sensors (, e.g., Landsat TIRS, MODIS, ASTER) measure emitted thermal radiance to retrieve Land Surface Temperature (LST).
Brightness Temperature, Emissivity Correction, and Split-Window LST
- At-Sensor Brightness Temperature (, in Kelvin): Inverting Planck's law assuming blackbody emissivity :
where are band-specific thermal calibration constants.
- Single-Channel Emissivity-Corrected LST (): Real surfaces have spectral emissivity (e.g., water , dense vegetation , bare concrete/soil ). Correcting for emissivity at effective wavelength :
- Split-Window (SW) Algorithms: Use two adjacent TIR bands inside the atmospheric window (e.g., near and near ). Because atmospheric water vapor absorbs slightly more strongly at than at , the brightness-temperature difference directly parameterizes atmospheric attenuation without external radiosonde profiles:
where , , is atmospheric water vapor, and are regression coefficients.
- Urban Heat-Island (UHI) Analysis: Impervious asphalt, concrete rooftops, low-albedo pavements, reduced evapotranspiration, and anthropogenic waste heat make cities warmer than surrounding vegetated countryside. Surface UHI (SUHI) intensity is quantified as:
Combining TIR-derived LST with NDVI and impervious-surface fraction guides urban cooling interventions such as tree-canopy corridors, parks, and high-albedo cool roofs.
Worked Out Examples
6.1
- A ground station in Kathmandu lies at longitude , latitude . Determine its UTM zone number and central meridian .
- A north-up Landsat raster in that zone has upper-left origin with , (). Compute the map coordinates of pixel .
Determining UTM zone and central meridian:
- UTM zone number:
- Central meridian of Zone :
Applying the affine geotransform to :
- Kathmandu at lies in UTM Zone 45N with central meridian .
- Pixel maps to .
6.2
- A post-classification LULC change matrix (values in ) between (rows) and (columns) across a basin is:
| Forest | Cropland | Urban | Total | |
|---|---|---|---|---|
| Forest | ||||
| Cropland | ||||
| Urban | ||||
| Total |
Compute (a) total unchanged (persistent) area and percentage, and (b) net change in Forest and Urban areas.
- In the same region, a forest plot before a wildfire had , and after the fire had . Compute and classify the burn severity.
Analyzing the LULC change matrix:
- Total persistent area (diagonal sum):
- Net change:
- Forest: ( net loss, driven mainly by converted to Cropland and to Urban, partially offset by reforestation).
- Urban: ( net expansion).
Computing pre-fire NBR, post-fire NBR, and dNBR:
Since , the plot suffered High Severity burn.
- Persistent area is (); net change is for Forest and for Urban.
- , , and (High Severity).
6.3
Landsat 8 TIRS Band 10 () measures at-sensor brightness temperatures over an urban concrete roof () and over a surrounding vegetated rural park (). Using :
- Compute the emissivity-corrected Land Surface Temperature () in Kelvin and for both sites.
- Calculate the Surface Urban Heat-Island (SUHI) intensity .
Emissivity-corrected LST for the urban concrete site ():
With and :
Emissivity-corrected LST for the rural vegetation site () and SUHI intensity: With :
Surface Urban Heat-Island intensity:
- Urban LST (); Rural LST ().
- Surface UHI intensity .
Exercises
Exercise — RS for GIS, Geospatial Architectures & Environmental Management
Distinguish between vector and raster data structures and between a Geographic Coordinate System (WGS84) and a Projected Coordinate System (UTM). Explain how R-tree spatial indexing accelerates GIS queries. For Pokhara at longitude , find its UTM zone number and central meridian.
Vector vs. raster, GCS vs. PCS, and R-trees: Vector models represent discrete points, lines, and polygons with explicit topology and exact boundary coordinates, whereas raster models represent continuous fields as regular grids suited to fast map algebra. A GCS (e.g., WGS84 EPSG:4326) uses angular degrees on a 3D ellipsoid, while a PCS (e.g., UTM) projects onto flat meters for accurate area and distance calculations. An R-tree groups nearby objects hierarchically inside Minimum Bounding Rectangles (MBRs), pruning non-intersecting subtrees in time during spatial queries.
UTM zone and central meridian for :
Vector stores discrete topology while raster stores continuous grids; WGS84 GCS uses degrees while UTM PCS uses meters; R-trees prune spatial searches via hierarchical MBRs. Pokhara lies in UTM Zone 44N with central meridian .
Explain the four classes of raster map algebra (local, focal, zonal, global) with one remote-sensing example each, and describe how Cloud-Optimized GeoTIFFs (COGs) and SpatioTemporal Asset Catalogs (STAC) enable scalable cloud-native geospatial workflows in Python (rasterio, geopandas, xarray).
Four classes of raster map algebra:
- Local: Cell-by-cell math at (e.g., computing NDVI from NIR and Red bands).
- Focal: Neighborhood window around (e.g., smoothing or DEM slope).
- Zonal: Summary statistics grouped by polygon/zone ID (e.g., mean rainfall per watershed).
- Global: Depends on the entire grid (e.g., distance to nearest road or flow accumulation).
COG, STAC, and Python stack: STAC indexes satellite scenes as searchable JSON Items with spatiotemporal extents and Asset links, while COGs place IFD headers upfront and store internally tiled, multi-resolution overviews so rasterio can fetch only intersecting tiles via HTTP Range requests. geopandas handles vector AOIs and CRS reprojection, and xarray performs labeled multi-dimensional array calculations across space, bands, and time.
Map algebra spans local, focal, zonal, and global scopes. STAC provides standardized spatiotemporal metadata discovery and COGs enable partial HTTP Range reads of tiled overviews, processed together by geopandas (vectors), rasterio (tiled rasters), and xarray (labeled -D cubes).
Compare image differencing and post-classification comparison for multi-temporal LULC monitoring. A 2010–2020 change matrix (in ) across a watershed shows: Agriculture Agriculture , Agriculture Urban , Agriculture Forest ; Forest Agriculture , Forest Forest , Forest Urban ; Urban Urban (zero transitions from Urban to other classes). Compute the percentage of persistent area and the net change in Forest and Urban areas.
Image differencing vs. post-classification comparison: Image differencing subtracts co-registered single bands/indices () to locate change magnitude rapidly but cannot identify "from–to" land-cover transitions. Post-classification compares two independently classified maps via a change matrix to quantify every class-to-class conversion, though classification errors from both dates multiply.
Persistence and net change calculations:
- Persistent area: .
- 2010 vs. 2020 Forest area: ; .
- 2010 vs. 2020 Urban area: ; .
Image differencing detects magnitude without class transitions, whereas post-classification builds a full from–to change matrix. Persistent area is (); net changes are (Forest) and (Urban).
Explain why SAR backscatter thresholding is effective for flood inundation mapping during monsoon storms, and describe the pre- vs. post-event NBR workflow for wildfire burn severity. If a forest site has pre-fire and post-fire , compute and state the burn-severity class.
SAR flood mapping & NBR burn-severity principle: Microwave SAR penetrates thick monsoon clouds day and night; smooth open floodwater reflects radar pulses specularly away from the sensor, producing very low backscatter () compared to rough land. Comparing pre- and post-flood SAR isolates newly inundated terrain while masking out permanent rivers. For wildfires, burning destroys NIR-reflecting leaf cell structure and exposes SWIR2-reflecting ash/soil, lowering and creating a positive .
Computing , , and :
Since , this corresponds to Moderate (Moderate–High) Burn Severity.
SAR penetrates clouds and detects specular dark floodwater by low thresholding relative to pre-event baselines. For the wildfire site, , , and (Moderate Severity).
Explain how split-window thermal infrared algorithms retrieve Land Surface Temperature (LST) without external atmospheric profiles, and describe the physical drivers of Surface Urban Heat Islands (SUHI). If a TIR band at records brightness temperature over a surface with emissivity , calculate the emissivity-corrected LST () in Kelvin and ().
Split-window LST and SUHI drivers: Split-window algorithms use two adjacent TIR channels (around and ) that experience different water-vapor absorption; their brightness-temperature difference cancels atmospheric water-vapor attenuation directly. Urban Heat Islands arise from low-albedo dark pavements, high-heat-capacity concrete/asphalt, reduced vegetation evapotranspiration, urban canyon trapping, and anthropogenic heat release.
Emissivity-corrected LST calculation: With and :
In degrees Celsius:
Split-window algorithms exploit differential water-vapor absorption between adjacent TIR channels to remove atmospheric effects, and SUHI reflects impervious surfaces, low albedo, and reduced evapotranspiration. The corrected LST is ().