🖼️ Satellite Image Processing and Analysis

This lesson builds core elective concepts in BSc Agriculture with practical applications and exam-oriented clarity.

---

Satellite Image Processing and Analysis

Image Processing Workflow

Processing a satellite image for agricultural application follows a structured pipeline:

Data Download → Preprocessing → Analysis → Interpretation → Output Product

Each stage is described in detail below.

---

Preprocessing

Raw satellite data must be corrected before analysis to remove sensor artifacts and environmental effects.

Radiometric Correction

Converts raw DN (Digital Number) values stored in the image file into physically meaningful units:

• DN → Radiance (W/m²/sr/μm): accounts for sensor gain and offset
• Radiance → Top-of-Atmosphere (TOA) Reflectance: removes differences due to sun angle and Earth-Sun distance
• TOA → Surface Reflectance: removes atmospheric effects (see atmospheric correction below)

Radiometric correction also removes:

• Striping: alternating bright/dark lines from detector failures
• Dropouts: single missing scan lines

Atmospheric Correction

Removes the effects of atmospheric scattering and absorption so that image values represent true surface reflectance:

• DOS (Dark Object Subtraction): Simple method; assumes darkest pixel in image should be zero reflectance (deep shadow or clear water); subtracts that value from all pixels; good enough for many vegetation analyses
• FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes): Physics-based model; requires atmospheric parameters (aerosol type, visibility, water vapour); standard in ArcGIS/ENVI
• Sen2Cor: Dedicated atmospheric correction processor for Sentinel-2; freely available; produces Level-2A (surface reflectance) product automatically

Geometric Correction

Removes spatial distortions due to sensor geometry, Earth's curvature, and terrain relief:

• GCP (Ground Control Points): Identifiable features with known coordinates (road intersections, bridges, corners of permanent structures) used to warp the image to a map projection
• Orthorectification: Uses DEM to remove relief displacement (mountains appear to "lean" in uncorrected images)
• Geometric accuracy of correctly processed Sentinel-2 is < 10 m RMSE; Landsat < 12 m RMSE

Mosaicking and Subsetting

• Mosaicking: Joining multiple overlapping image tiles (e.g., multiple Sentinel-2 granules covering a state) into one seamless image; requires radiometric balancing at seam lines
• Subsetting (Clipping): Extracting the study area from a larger scene to reduce file size and processing time

---

Visual Interpretation

Before digital analysis, image analysts interpret satellite images visually using eight key elements:

Interpretation Key	Description	Example
Tone / Colour	Brightness or hue of features	Dark blue = deep water; bright green = healthy crops
Size	Relative dimensions of objects	Large irregular polygon = plantation; small regular = paddy field
Shape	Geometric form of features	Circular = centre-pivot irrigation; rectangular = paddy bunds
Texture	Coarseness/smoothness of image pattern	Smooth = flat fields; rough = forest canopy
Pattern	Spatial arrangement of features	Orchard rows; strip cropping; windbreaks
Shadow	Dark areas cast by tall objects	Trees cast shadows → confirms tree presence and height
Site	Geographic position / topographic context	Paddy typically in low-lying areas near water
Association	Relationship to nearby features	Sugarcane near sugar mill; fish ponds near canals

---

Digital Image Classification

Image classification assigns each pixel to a land use / land cover (LULC) category based on its spectral values.

Supervised Classification

The analyst provides training samples — areas of known land use — by digitizing polygons over representative areas of each class in the image:

1. Collect training samples for each class (e.g., rice, wheat, cotton, fallow, water, urban)
2. The software extracts spectral statistics (mean and covariance) for each class from the training polygons
3. A classification algorithm assigns every pixel to the class it most likely belongs to

Classification Algorithms:

• Maximum Likelihood (ML): Assumes Gaussian spectral distribution; probabilistic; most common traditional method
• Minimum Distance: Assigns pixel to class with closest mean; simple and fast
• Support Vector Machine (SVM): Finds optimal hyperplane separating classes; excellent with limited training data; handles high-dimensional hyperspectral data well
• Random Forest (RF): Ensemble of decision trees; robust; handles non-normal distributions; widely used in current research
• Neural Networks / Deep Learning (CNN): Best accuracy with large training datasets; increasingly used for national-scale crop mapping

Unsupervised Classification

No training samples; algorithm groups pixels into clusters based on spectral similarity:

• K-Means: User specifies number of clusters (k); algorithm iteratively assigns pixels and recalculates cluster centres until stable
• ISODATA (Iterative Self-Organizing Data Analysis Technique): Extended K-Means; can split and merge clusters automatically; standard in ERDAS IMAGINE and ENVI

After unsupervised classification, the analyst interprets each cluster using ground truth knowledge and labels it as a land use class.

Comparison

Feature	Supervised	Unsupervised
Training data	Required	Not required
Analyst input	High (collect samples)	Moderate (interpret clusters)
Control over classes	Full — user defines classes	Limited — algorithm finds clusters
Risk	Training bias if samples unrepresentative	Clusters may not match meaningful classes
Best for	Known study area with good ground truth	Exploratory; unknown areas

---

Accuracy Assessment

After classification, accuracy is evaluated using an independent set of validation points (not used in training) compared to known ground conditions:

Confusion Matrix (Error Matrix)

A table comparing classified vs. actual (reference) land use for each pixel in the validation set:

	Classified: Rice	Classified: Wheat	Classified: Fallow	Row Total (Reference)
Reference: Rice	45	3	2	50
Reference: Wheat	2	43	5	50
Reference: Fallow	1	4	45	50
Column Total (Classified)	48	50	52	150

From the confusion matrix:

• Overall Accuracy = (diagonal sum) / (total pixels) = (45+43+45)/150 = 88.7%
• Producer's Accuracy (omission error) = correctly classified / reference total for a class; e.g., Rice = 45/50 = 90%
• User's Accuracy (commission error) = correctly classified / classified total for a class; e.g., Rice = 45/48 = 93.75%
• Kappa Coefficient (κ): Measures agreement accounting for chance; κ > 0.8 = strong agreement; κ 0.6–0.8 = substantial; κ < 0.4 = poor

Minimum 50 validation points per class is the standard for reliable accuracy assessment (Anderson et al., 1976).

---

Vegetation Indices

Vegetation indices are mathematical combinations of spectral bands that enhance the vegetation signal and reduce the effect of soil background, atmosphere, and viewing angle.

NDVI — Normalized Difference Vegetation Index

Formula: NDVI = (NIR − Red) / (NIR + Red)

• Range: −1 to +1
• Interpretation:
  • Healthy vegetation: 0.4–0.8 (depends on crop, season, growth stage)
  • Sparse vegetation / senescent crops: 0.1–0.3
  • Bare soil: 0.05–0.15
  • Water bodies: < 0 (negative)
  • Clouds/snow: negative or near zero
• Limitations: Saturates in dense canopies (LAI > 3); sensitive to soil background at low vegetation cover; affected by atmosphere in uncorrected images
• Uses: Crop condition monitoring, drought early warning, crop calendar mapping, yield forecasting, Land use change detection

SAVI — Soil Adjusted Vegetation Index

Formula: SAVI = [(NIR − Red) / (NIR + Red + L)] × (1 + L)

• L: Soil correction factor; L = 0.5 for intermediate vegetation density; L = 1 for very sparse; L = 0 reduces to NDVI
• Purpose: Corrects NDVI for soil brightness effects in sparse canopy conditions (< 30% canopy cover) — common in dryland farming, early crop season
• Advantage over NDVI: Less sensitive to soil colour variations (dark vs. light soils)

EVI — Enhanced Vegetation Index

Formula: EVI = 2.5 × [(NIR − Red) / (NIR + 6×Red − 7.5×Blue + 1)]

• Reduces atmospheric contamination (uses blue band to correct aerosol effects)
• Reduces canopy background variation
• Does not saturate as readily in dense canopies (Amazon, tropical crops)
• Used in MODIS standard vegetation products (MOD13A2)

Other Important Indices

Index	Formula	Application
RVI (Ratio Vegetation Index)	NIR / Red	Simple vegetation measure; early satellite work
NDWI (Normalized Difference Water Index)	(Green − NIR) / (Green + NIR)	Water body detection, extent mapping
LSWI (Land Surface Water Index)	(NIR − SWIR) / (NIR + SWIR)	Leaf/canopy water content; irrigation scheduling
NDSI (Normalized Difference Snow Index)	(Green − SWIR) / (Green + SWIR)	Snow cover mapping
NBR (Normalized Burn Ratio)	(NIR − SWIR2) / (NIR + SWIR2)	Post-fire damage severity assessment
NDRE (Normalized Difference Red Edge)	(NIR − Red Edge) / (NIR + Red Edge)	Crop nitrogen status; early stress (Sentinel-2)
Salinity Index	sqrt(Red × SWIR)	Soil salinity mapping

---

Change Detection

Multitemporal analysis uses images from two or more dates to detect changes on the ground:

• Post-classification comparison: Classify each date separately → compare classifications pixel-by-pixel
• Image differencing: Subtract band values of date-1 from date-2; significant positive or negative differences indicate change
• CVA (Change Vector Analysis): Direction and magnitude of change in feature space
• Time series analysis: NDVI time series (e.g., MODIS 16-day composites) to track seasonal crop development and detect anomalies

---

Crop Identification Using Multi-Temporal NDVI

Different crops have distinct NDVI temporal profiles due to different planting dates, growth rates, and harvest times:

Crop	NDVI Pattern
Kharif rice (India)	Planted June–July; NDVI peaks August–September (~0.7); drops sharply after harvest October
Rabi wheat	Planted November; NDVI rises January–February; peaks March (~0.75); harvested April
Sugarcane	Slow rise; stays high for 10–12 months; distinctive long-duration high NDVI
Cotton	Planted June; moderate NDVI (0.4–0.6); longer season than rice

The FASAL programme (Ministry of Agriculture) uses MODIS NDVI time series combined with LISS-III for advance crop area and production estimates for rice, wheat, rapeseed, cotton, and jute — providing forecasts months before harvest.

---

Overview

Image processing transforms raw satellite data through radiometric correction, atmospheric correction, and geometric rectification into analysis-ready surface reflectance products. Visual interpretation uses eight keys (tone, colour, size, shape, texture, pattern, shadow, association). Digital classification is either supervised (using training samples with ML, SVM, Random Forest algorithms) or unsupervised (K-Means, ISODATA). Accuracy is reported via the confusion matrix — Overall Accuracy and Kappa coefficient. Vegetation indices (NDVI, SAVI, EVI, NDRE) amplify the vegetation signal for crop monitoring, while multi-temporal NDVI profiles allow crop type discrimination — the basis of India's FASAL crop forecasting programme.

---

Summary Cheat Sheet

Topic	Key takeaway
Main focus	Image preprocessing, visual interpretation, supervised and unsupervised classification, accuracy assessment, and vegetation indices including NDVI, SAVI, and EVI.
Section context	Revise this lesson with the rest of Remote Sensing for stronger conceptual continuity.