Remote Sensing Change Detection and Image Analysis

Principal Component Analysis for Change Detection

When applied to Principal Component Analysis (PCA) to detect changes, a multitemporal file is generated with the bands corresponding to the two dates. PCA is then applied to this file. The primary components collect image stability, while the secondary components recognize the information that does not change. Two common techniques used in this context are multitemporal regression and change vectors.

Scatter Plots and Band Correlation

A scatter plot (dispersograma) is used to analyze the correlation between bands and assist in band selection. It reveals the presence of clusters. For instances of no apparent reflectance, factors include vertical observation of the atmosphere. Without a reflector, surface topography is considered Lambertian.

Vegetation Indices and Biophysical Models

Vegetation Indices enhance the contribution of vegetation in the spectral response of specific areas. They work by mitigating factors such as soil and atmosphere. These indices serve as input variables for monitoring biophysical models and allow for detailed vegetation analysis.

Kappa Index for Accuracy Assessment

The multitemporal Kappa index compares ratings given by different methods to see if they differ in their degree of adjustment to reality.

Stages of Geometric Correction

1. Location: Identifying points common to both the image and the map (Control Points).
2. Calculation: Processing functions between the coordinates of the image and the reference.
3. Transfer: Moving the original data to a new position after the defined transformation.

Confusion Matrix in Change Detection

The confusion matrix is used in the verification of results, particularly in detecting change, because it relates the two images. The format includes:

• 1. Areas detected as stable if they truly are.
• 2. Verification if detected changes are real.
• 3. Stable areas within a specific category.
• 4. Areas with no changes in reality.
• 5. Instances where no real change occurred and was not discriminated against.
• 6. Changes that were incorrectly identified.

Unsupervised vs Supervised Classification

Unsupervised Methods: Spectral Class to Information Class

The computer defines spectral classes using clustering algorithms to determine the statistical similarity of the pixels. The analyst defines the number of classes beforehand and relates the spectral classes (clusters) with information classes. In this method, the phases of training and assignment are often merged.

Supervised Methods: Information Class to Spectral Class

The analyst identifies representative samples of each class, known as training areas. The computer extracts numerical information based on the strength of these samples to create spectral signatures. During the allocation phase, each pixel is compared with these signatures and labeled with the class that fits best.

Change Detection Techniques and Prerequisites

Prerequisites: Radiometric and geometric adjustment.

Visual Analysis: Multitemporal color composite.

1. Digital analysis techniques for continuous variables: Subtractions and ratios, PCA, regression, and change vectors.
2. Categorized variables: Multitemporal cross-tabulation and MultiTOS classification.