8.4 Image Processing, Classification and LiDAR

Image Processing, Classification & LiDAR

Modern photogrammetry and remote sensing workflows culminate in digital image processing, automated automated land-cover classification, and high-density active 3D point cloud generation using Light Detection and Ranging (LiDAR). Mastery of these analytical techniques is essential for SURCON candidates involved in modern environmental, forestry, urban, and engineering surveying.

---

Digital Image Preprocessing

Raw satellite and airborne sensor imagery contains systematic radiometric, atmospheric, and geometric distortions that must be corrected before quantitative spatial analysis.

1. Radiometric Correction

Adjusts digital numbers (DN) to compensate for sensor detector variation, solar illumination geometry, and solar zenith angle variations over time. DN values are converted to Top-of-Atmosphere (TOA) Radiance ($L_\lambda$) using calibrated gain and bias values.

2. Atmospheric Correction

Solar radiation passing twice through Earth's atmosphere undergoes scattering (Rayleigh and Mie scattering) and atmospheric absorption by water vapor, ozone, and aerosols. Atmospheric correction converts TOA Radiance into true Surface Reflectance (Bottom-of-Atmosphere / BOA).

• Dark Object Subtraction (DOS): An empirical image-based correction method that assumes deep clear water or shadow pixels should have near-zero reflectance, subtracting the minimum pixel value across the image.
• Radiative Transfer Models: Physics-based algorithms (e.g., FLAASH, 6S) that model atmospheric parameters using elevation, humidity, and atmospheric aerosol models.

3. Geometric Rectification & Orthorectification

Aligns image coordinates with ground geographic coordinates. Orthorectification uses a DEM and ground control points (GCPs) to correct for camera tilt and terrain relief displacement on a pixel-by-pixel basis, producing an orthorectified image.

---

Image Classification Methods

Image classification converts continuous multispectral pixel data into discrete thematic land cover maps (e.g., agriculture, water, forest, built-up area).

                    ┌───────────────────────────────────────────────┐
                    │         Multispectral Image Data              │
                    └───────────────────────┬───────────────────────┘
                                            │
            ┌───────────────────────────────┴───────────────────────────────┐
            ▼                                                               ▼
┌───────────────────────┐                               ┌───────────────────────┐
│ Supervised Methods    │                               │ Unsupervised Methods  │
│ (Requires Training)   │                               │ (Automated Clustering)│
├───────────────────────┤                               ├───────────────────────┤
│ • Maximum Likelihood  │                               │ • K-Means Clustering  │
│ • Random Forest (ML)  │                               │ • ISODATA Algorithm   │
│ • Support Vector (SVM)│                               │                       │
└───────────────────────┘                               └───────────────────────┘

Supervised Classification

The analyst defines representative training sites (polygons) of known land-cover classes on the image. The statistical properties of these training samples are used to train a classifier algorithm:

• Maximum Likelihood Classifier (MLC): Assumes normal (Gaussian) distribution of pixel values per class. Calculates the probability that a given pixel belongs to a specific class and assigns the pixel to the class with maximum probability.
• Random Forest (RF): An ensemble machine learning algorithm that constructs dozens of decision trees on random subsets of training data, assigning classes based on majority voting. RF handles complex non-linear relationships without assuming Gaussian distribution.

Unsupervised Classification

Automated algorithms group image pixels into natural spectral clusters without prior analyst training data:

• K-Means: Aggregates pixels into $K$ specified clusters based on Euclidean spectral distance, iteratively updating cluster mean vectors.
• ISODATA (Iterative Self-Organizing Data Analysis Technique): An advanced clustering method that allows clusters to split, merge, or be discarded dynamically based on standard deviation thresholds.

---

Classification Accuracy Assessment

A classification is incomplete until its accuracy is quantified using an independent validation dataset displayed in a Confusion Matrix (Error Matrix).

• Overall Accuracy: Total diagonal correct pixels divided by total sample size $N$.
• Producer's Accuracy: Measures omission error (probability that a ground truth reference pixel is correctly classified on the map).
• User's Accuracy: Measures commission error (probability that a pixel classified on the map actually represents that class on the ground).
• Kappa Coefficient ($\kappa$): Quantifies classification accuracy corrected for chance agreement:

$\kappa = \frac{N \sum_{i=1}^{k} n_{ii} - \sum_{i=1}^{k} (n_{i+} \cdot n_{+i})}{N^2 - \sum_{i=1}^{k} (n_{i+} \cdot n_{+i})}$

---

LiDAR Fundamentals

LiDAR (Light Detection and Ranging) is an active airborne or terrestrial remote sensing technology that measures distances using high-frequency laser pulses (typically $1,064\text{ nm}$ for topographic systems, $532\text{ nm}$ green laser for bathymetric systems).

Time-of-Flight Ranging Equation

Distance $D$ from aircraft laser scanner to target surface is derived from pulse time-of-flight $t$:

$D = \frac{c \cdot t}{2}$

Where $c$ is the speed of light ($3.0 \times 10^8\text{ m/s}$) and $t$ is the elapsed round-trip time between pulse emission and return detection.

Discrete Return vs. Full Waveform LiDAR

• Discrete Return LiDAR: Records discrete return echoes when the laser pulse strikes physical surfaces (e.g., First Return = canopy top; Intermediate Returns = branches; Last Return = bare ground).
• Full Waveform LiDAR: Digitizes the continuous energy return wave across time, capturing complex canopy structure density in forestry applications.

Deliverables: Point Density and Canopy Models

• Point Density: Measured in points per square meter ($\text{pts/m}^2$). High-density LiDAR ($> 10-20\text{ pts/m}^2$) penetrates dense vegetation canopies to capture ground terrain.
• Ground Filtering: Algorithms (e.g., Cloth Simulation Filtering - CSF, Progressive Morphological Filtering) separate ground point returns from non-ground returns.
• Canopy Height Model (CHM): Quantifies true vegetation height above terrain by subtracting the bare-earth DEM from the top-surface DSM:

$\text{CHM} = \text{DSM} - \text{DEM}$