Remote Sensing: Digital Image Classification and Land Cover Change Detection

Land Cover Taxonomies and Thematic Mapping

• Thematic Mapping Purpose and Role:

  • Digital image classification translates continuous multispectral or hyperspectral measurements into categorical information classes corresponding to distinct land cover or land use types.

  • Thematic maps derived from remote sensing provide foundational baseline data for diverse environmental science and management applications:

  • Monitoring regional and global land cover changes over time.

  • Directly quantifying the environmental impacts of discrete events (e.g., wildfires, floods, volcanic eruptions) or human activities (e.g., urbanization, clear-cutting).

  • Establishing spatial priorities for land conservation and reserve design.

  • Guiding municipal, regional, and landscape planning.

  • Providing critical boundary conditions and land surface parameterizations for environmental models, including global and regional carbon budget simulations, biodiversity distribution models, meteorological and climatological forecasting, atmospheric chemistry and air pollution transport models, and hydrologic/water quality assessments.

• Design Criteria for Classification Schemes:

  • Mutually Exclusive: Every mapped point must belong unambiguously to only one thematic class; no "taxonomic" overlap is permitted between category definitions.

  • Exhaustive: All land cover types present across the geographical domain must be accounted for within the scheme without omission.

  • Hierarchical: Classification schemes are structured in nested tiers where fine-scale sublevel classes can be systematically aggregated into broader, higher-level categories to support multi-scale research.

• International Geosphere-Biosphere Program (IGBP) Classification System:

  • Standardized global classification system applied to moderate-resolution satellite datasets, including global $1\,\text{km}$ products derived from MODIS (Moderate Resolution Imaging Spectroradiometer) observations.

  • Organizes terrestrial land cover into three broad clusters comprising 17 distinct land categories:

  • Forests:

    • Evergreen Needleleaf Forest

    • Evergreen Broadleaf Forest

    • Deciduous Needleleaf Forest

    • Deciduous Broadleaf Forest

    • Mixed Forests

  • Shrublands, Grasslands, and Wetlands:

    • Closed Shrublands

    • Open Shrublands

    • Woody Savannas

    • Savannas

    • Grasslands

    • Permanent Wetlands

  • Agriculture, Urban, and Barren:

    • Croplands

    • Urban and Built-up

    • Cropland/Natural Vegetation Mosaic

    • Snow and Ice

    • Barren or Sparsely Vegetated

• National Land Cover Database (NLCD):

  • Provides a standardized, high-resolution thematic land cover map of the United States at $30\,\text{m}$ spatial resolution derived from Landsat satellite imagery.

  • Features a standardized numerical coding system comprising specific classification categories:

  • 11: Open Water

  • 12: Perennial Ice/Snow

  • 21: Developed, Open Space

  • 22: Developed, Low Intensity

  • 23: Developed, Medium Intensity

  • 24: Developed, High Intensity

  • 31: Barren Land (Rock/Sand/Clay)

  • 41: Deciduous Forest

  • 42: Evergreen Forest

  • 43: Mixed Forest

  • 51: Dwarf Scrub (restricted to Alaska)

  • 52: Shrub/Scrub

  • 71: Grassland/Herbaceous

  • 72: Sedge/Herbaceous (restricted to Alaska)

  • 73: Lichens (restricted to Alaska)

  • 74: Moss (restricted to Alaska)

  • 81: Pasture/Hay

  • 82: Cultivated Crops

  • 90: Woody Wetlands

  • 95: Emergent Herbaceous Wetlands

Principles of Digital Image Classification

• The Classification Process:

  • Digital classification algorithms ingest a multi-layer image stack containing multiple co-registered spectral bands (e.g., Band 1 through Band $n$).

  • Pixels are sorted and grouped into discrete categories based on distinct characteristics in their multispectral response (variations in digital numbers or surface reflectance across the wavelength spectrum).

  • The classified output assigns every individual pixel an integer representing an informational category, yielding a final thematic map and an area inventory quantifying class distributions.

• Quantitative Area Inventory Example:

  • In a representative maximum likelihood classification of a regional scene ($1,580,706$ total pixels; $395,176.50\,\text{km}^2$):

  • Soil: $73,945$ pixels; $18,486.25\,\text{km}^2$; $4.68\%$

  • Snow: $110,697$ pixels; $27,674.25\,\text{km}^2$; $7.00\%$

  • Herbaceous: $686,120$ pixels; $171,530.00\,\text{km}^2$; $43.41\%$

  • Shrublands: $438,187$ pixels; $109,546.75\,\text{km}^2$; $27.72\%$

  • Forested: $166,335$ pixels; $41,583.75\,\text{km}^2$; $10.52\%$

  • Urban: $89,905$ pixels; $22,476.25\,\text{km}^2$; $5.69\%$

  • Water: $15,517$ pixels; $3,879.25\,\text{km}^2$; $0.98\%$

  • Total: $1,580,706$ pixels; $395,176.50\,\text{km}^2$; $100.00\%$

• Spectral Classes vs. End-User (Informational) Classes:

  • Spectral Categories: Distinct clusters of uniform radiometric brightness across several spectral bands identified by clustering or pattern recognition algorithms.

  • Informational/End-User Categories: Land cover or land use labels defined by human interpretation (e.g., Deciduous Forest, Irrigated Crops, Water, Rainfed Crops, Urban Areas).

  • In an idealized system, there is a simple 1:1 mapping between spectral classes and informational classes.

  • In operational environments, two primary complexities emerge:

  • One-to-many relationship: A single target informational class may comprise multiple distinct spectral classes due to physical variations such as terrain shading, topographic slope/aspect, canopy cover density, soil moisture variation, and water turbidity or depth.

  • Many-to-one relationship: Two or more distinct target informational classes may display identical spectral signatures in available bands (e.g., pasture vs. rainfed crops), preventing direct discrimination without ancillary data or thematic class merging.

• Fundamental Rule: Spectral Separability:

  • Land cover classification rests on the hypothesis that surface targets exhibit statistically separable signatures based on variance in pixel brightness across spectral wavelengths.

  • Spectral separability is evaluated using 1D histograms, univariate probability distributions, and multi-dimensional scatter plots ("feature space plots").

  • Examining multi-band combinations (such as Red versus Near-Infrared) reveals spatial clustering not visible in single-band histograms, forming the foundation for multidimensional classification algorithms.

Supervised Classification and Training Site Protocol

• Taxonomy of Classification Paradigms:

  • Supervised Classification:

  • Relies on a priori knowledge regarding the identity and geographic location of specific land cover types.

  • Requires selecting representative sample pixels ("training sites") across the image to calculate statistical profiles for each class.

  • Spectral properties of these training sites train mathematical decision rules that assign all unclassified pixels across the full scene.

  • Unsupervised Classification:

  • Operates without a priori knowledge of ground conditions.

  • Algorithms automatically aggregate pixels into natural mathematical groupings ("clusters") based purely on numerical proximity in spectral space.

  • The analyst re-labels these spectral clusters a posteriori into meaningful informational classes.

• Selection Criteria for Training Sites:

  • Sourcing: Derived from known optical properties, targeted field surveys, georeferenced aerial and ground photography, or existing validated cartographic records.

  • Homogeneity: Training sites must encompass pure, uniform stands of the target class and strictly avoid transitional boundaries, ecotones, or mixed pixels.

  • Sample Size and Distribution: Multiple non-contiguous training sites must be delineated for each category to capture natural geographic variance, differing canopy closures, and varying soil backgrounds.

  • Representing Within-Class Heterogeneity: Real environmental variation (e.g., differences in illumination caused by rugged topography, varying water depth or sediment load, varying stand ages) must be explicitly sampled by assigning distinct sub-classes during the training phase.

• Evaluation of Training Statistics Prior to Scene Classification:

  • Prior to running full classification algorithms, training statistics must be reviewed to confirm statistical separability and avoid redundant spectral dimensions.

  • Univariate Histograms and Summary Statistics: Class means, standard deviations, and variances are calculated for every band. Overlapping histograms indicate poor class separability in that wavelength.

  • Bi-Modal Distributions: If a single training class presents a bi-modal histogram, it indicates that two distinct spectral sub-types have been incorrectly combined, requiring division into two separate training classes.

  • Multi-Dimensional Feature Space Plots: Ellipsoids or bounding boxes defined by training statistics are plotted in 2D or 3D feature spaces to identify overlapping classes that require band substitutions or additional sampling.

Mathematical Classification Algorithms

• Minimum Distance to Mean Classifier:

  • Assigns an unknown pixel vector to the class whose mean vector is nearest in spectral space using Euclidean distance.

  • The Euclidean distance $d_{x, A}$ from unknown pixel $x$ to class centroid $A$ across $m$ spectral bands is calculated as:     $d_{x, A} = \sqrt{\sum_{i=1}^{m} (DL_{x, i} - DL_{A, i})^2}$

  • Where $DL_{x, i}$ is the brightness value (digital number or reflectance) of pixel $x$ in band $i$.

  • $DL_{A, i}$ is the mean brightness value of training class $A$ in band $i$.

  • $m$ is the total number of spectral bands.

  • The algorithm calculates $d_{x, k}$ for all candidate classes $k = 1, \dots, c$ and assigns pixel $x$ to the class that yields the minimum distance.

  • Characteristics: Computationally fast and mathematically straightforward, but insensitive to differences in class variance and covariance (it treats all classes as symmetrical, spherical clusters of equal size).

• Parallelepiped Classifier:

  • Constructs multidimensional rectangular boundaries ("boxes") in feature space around each class, defined by user-specified dispersion ranges around the mean (typically based on the class standard deviation, such as $\pm 1\sigma$, $\pm 2\sigma$, or empirical minimum and maximum bounds).

  • Decision rule for assigning unknown pixel $x$ to category $A$:     $DL_{A, i} - R_{A, i} \le DL_{x, i} \le DL_{A, i} + R_{A, i}$

  • Where $R_{A, i}$ represents the spectral dispersion range or standard deviation threshold of class $A$ in band $i$.

  • This condition must be simultaneously satisfied across all evaluated bands $i = 1, \dots, m$.

  • Limitations:

  • Pixels falling outside all defined bounding boxes remain permanently unclassified.

  • High spectral correlation between bands creates diagonally oriented distributions, leading rectangular boxes to overlap in feature space; pixels falling within overlapping regions are assigned ambiguous, multi-class labels.

• Nearest Neighbor and $k$-Nearest Neighbor ($k$-NN):

  • Rather than summarizing training data into aggregate parametric statistics (e.g., class means), the algorithm measures distances between the unknown pixel and every individual training pixel in multidimensional feature space.

  • In standard Nearest Neighbor classification, the pixel is assigned to the class of the single closest training pixel.

  • In $k$-Nearest Neighbor ($k$-NN), the algorithm identifies the $k$ closest training pixels and assigns the unknown pixel to the majority class through voting, providing robustness against isolated training outliers.

• Maximum Likelihood Classifier (MLC):

  • Parametric probabilistic classifier that calculates the statistical probability that an unknown pixel belongs to each candidate class, assigning the pixel to the class with the highest probability.

  • Operates on the assumption that training data for each category follow a multivariate normal (Gaussian) distribution.

  • The univariate probability density function for class $\omega_i$ is defined as:     $p(x | \omega_i) = \frac{1}{\sigma_i \sqrt{2\pi}} \exp\left( -\frac{(BV - \mu_i)^2}{2\sigma_i^2} \right)$

  • Where $BV$ is the brightness value of the unknown pixel.

  • $\mu_i$ is the mean brightness value for class $\omega_i$.

  • $\sigma_i$ is the standard deviation for class $\omega_i$.

  • Multi-dimensional extension: Uses the full mean vector and band-to-band covariance matrix, defining equiprobability ellipsoids in multidimensional space that account for spectral correlation between bands.

  • Decision Rule: An unknown measurement vector $X$ is assigned to class $\omega_i$ if and only if $p(X | \omega_i) > p(X | \omega_j)$ for all candidate classes $j \ne i$.

• Other Modern Classifiers:

  • Non-parametric and machine learning alternatives include decision tree classifiers, random forests, artificial neural networks (ANN), and specialized hyperspectral classification algorithms (e.g., Spectral Angle Mapper).

Change Detection Prerequisites and Environmental Coherence

• Objectives of Change Detection:

  • Identifies and quantifies spatial and spectral alterations over time caused by natural successional dynamics, abrupt disturbances, or anthropogenic activity.

  • Addresses quantitative monitoring questions:

  • Area of forest stands cleared for agricultural grazing.

  • Volumetric and spatial retreat of glacial ice sheets.

  • Linear and areal extent of coastal shoreline erosion.

  • Surface area loss in depleting lake and reservoir systems.

• Mandatory Observation Coherence Criteria:

  • Reliable automated change detection requires controlling for extraneous environmental and sensor differences across image dates.

  • Multi-date image pairs must share:

  • Identical Spectral Bands: Matched sensor bandwidths and wavelength centers.

  • Identical Viewing Geometry: Constant sensor view angles and orbital trajectories to avoid variations caused by bidirectional reflectance distribution functions (BRDF).

  • Identical Geographical Registration: High-accuracy spatial coregistration at sub-pixel levels to prevent false change signals along high-contrast boundaries.

  • Consistent Temporal Windows: Near-anniversary acquisition dates to minimize confounding differences in sun angle, solar zenith, soil moisture, and plant phenological states.

  • Consistent Sky and Atmospheric Conditions: Cloud-free, haze-free conditions supported by rigorous absolute radiometric calibration and atmospheric correction to surface reflectance.

Image Algebra, Differencing, and Anomaly Mapping

• Mathematical Foundations of Image Algebra:

  • Mathematical operations (subtraction, ratioing) applied on a pixel-by-pixel basis between two coregistered images acquired at dates $t_1$ and $t_2$.

  • Operates directly on calibrated surface reflectance, top-of-atmosphere (TOA) reflectance, raw digital numbers, or derived spectral vegetation/water indices.

• Normalized Burn Ratio (NBR) and Delta NBR ($\Delta\text{NBR}$):

  • Leverages contrasting responses of Near-Infrared (NIR) and Shortwave-Infrared (SWIR) wavelengths to detect canopy destruction and exposed charred soils.

  • Normalized Burn Ratio equation:     $\text{NBR} = \frac{\text{NIR} - \text{SWIR}}{\text{NIR} + \text{SWIR}}$

  • Delta NBR ($\Delta\text{NBR}$) quantifies burn severity across wildfire events by subtracting the post-fire image from the pre-fire image:     $\Delta\text{NBR} = \text{NBR}_{\text{pre-fire}} - \text{NBR}_{\text{post-fire}}$

  • Higher positive differences directly correlate with increasing wildfire burn severity.

• Vegetation Index Differencing:

  • Calculated by subtracting Normalized Difference Vegetation Index (NDVI) images:     $\Delta\text{NDVI} = \text{NDVI}_{t_2} - \text{NDVI}_{t_1}$

  • Evaluated over tropical deforestation regions (e.g., Acre state, Brazil from 1990 to 2010), where sharp declines in NDVI highlight rapid forest clearance and agricultural expansion.

• Histogram Thresholding for Binary and Trinary Change Mapping:

  • In a difference image ($t_2 - t_1$), stable, unchanged pixels form a Gaussian-like distribution centered close to zero.

  • Pixels experiencing real physical changes reside within the extended positive and negative distribution tails.

  • Analysts inspect the differenced image histogram and define manual or statistical threshold values ($-$ and $+$) to categorize the landscape into discrete thematic change states:

  • Negative change (e.g., vegetation or water loss).

  • Stable / no change.

  • Positive change (e.g., vegetation gain or revegetation).

• Direct Single-Band Spectral Differencing:

  • Applied to individual spectral bands sensitive to specific environmental targets (e.g., subtracting NIR bands between Landsat ETM+ from May 2000 and ASTER from April 2003 over Lake Mead, NV).

  • Because water absorbs NIR strongly while surrounding rock/soil is highly reflective, decreases in reservoir water volume appear as distinct, measurable positive difference zones.

• Vegetation Anomaly and Long-Term Trend Mapping:

  • Vegetation Anomalies: Compare a specific time point against a long-term multi-year baseline average:

  • California NDVI anomaly: Evaluated in January 2014 relative to the long-term 2004–2013 January baseline, mapping statewide vegetative drought stress.

  • Southeastern Australia NDVI anomaly: Recorded in November 2006 relative to the 2000–2005 baseline, showing acute negative anomalies during severe drought periods.

  • Linear Trends: Dense temporal satellite series are fitted with linear regressions over decades (e.g., multi-decadal NDVI trends from 2001 to 2021 capturing extensive greening trajectories across Arctic tundra biomes).

Multitemporal Color Composites

• Multi-Date Display Mechanics:

  • Standard false-color composites display three different spectral bands from a single date across the Red, Green, and Blue (RGB) color channels.

  • Multitemporal color composites assign the same spectral band or index from two or three distinct dates into the RGB color channels.

  • Areas with no spectral change display as balanced neutral gray or white tones ($R = G = B$).

  • Areas experiencing physical change display as saturated primary or secondary colors driven by the date that dominates the spectral response.

• Two-Date Color Composite Implementations:

  • Lake Mead, NV (NIR Band Evaluation):

  • Landsat ETM+ (May 2000) assigned to Green and Blue; ASTER (April 2003) assigned to Red.

  • Areas where NIR reflectance increased display as Red.

  • Areas where NIR reflectance decreased display as Cyan ($\text{Green} + \text{Blue}$).

  • Unchanged terrain displays in balanced shades of Gray.

• North Rim, Grand Canyon (Landsat TM NDVI Composites):

  • Landsat TM 1990 NDVI vs. Landsat TM 2000 NDVI.

  • Areas exhibiting an increase in NDVI show up Green.

  • Wildfire burn scars with marked decreases in NDVI display prominently as Magenta ($\text{Red} + \text{Blue}$).

• New Orleans Coastal Wetland Erosion (NDWI Composites):

  • Normalized Difference Wetness Index (NDWI) is computed to enhance open surface water against adjacent terrestrial soils and vegetation:     $\text{NDWI} = \frac{\text{Green} - \text{NIR}}{\text{Green} + \text{NIR}}$

  • Comparing NDWI from 2003 against NDWI from 2021 within an RGB color composite isolates storm-driven wetland loss and sea-level inundation paths.

• Three-Date Multitemporal Composite Dynamics (Acre, Brazil Case Study):

  • Combines three distinct temporal intervals into an RGB composite: Date 1 (1990) = Red, Date 2 (2010) = Green, Date 3 (2024) = Blue.

  • Stable Forest: Displays in neutral white/gray tones ($R \approx G \approx B$ across all three observation dates).

  • Early-Stage Deforestation: Displays Red (high vegetation index in 1990 [$\text{Red} \uparrow$], followed by low indices in 2010 and 2024 [$\text{Green} \downarrow$, $\text{Blue} \downarrow$]).

  • Mid-Stage Deforestation: Displays Yellow (high vegetation index in 1990 [$\text{Red} \uparrow$] and 2010 [$\text{Green} \uparrow$], followed by clearance in 2024 [$\text{Blue} \downarrow$]).

  • Deforestation Followed by Regrowth/Restoration: Displays Magenta (high in 1990 [$\text{Red} \uparrow$], cleared in 2010 [$\text{Green} \downarrow$], and actively revegetated by 2024 [$\text{Blue} \uparrow$]).

Multi-Date Layer Stacking and Post-Classification Comparison

• Multidate Layer Stacking Paradigm:

  • Coregisters and concatenates individual spectral bands from two or more observation dates into a single unified raster stack (e.g., combining Bands 1, 2, 3 from 1990 with Bands 1, 2, 3 from 2010 into a 6-band file).

  • Standard supervised or unsupervised classification algorithms (or transformation techniques such as Principal Component Analysis [PCA]) are run directly on this combined dataset.

  • The classification assigns pixels directly into dynamic transition classes (e.g., Forest-to-Forest, Forest-to-Crops, Forest-to-Burned, Crop-to-Forest, Water-to-Crop).

• Post-Classification Comparison Methodology:

  • Independent classification is performed on images from each date ($t_1$ and $t_2$) using individual thematic category maps.

  • The resulting classified maps are compared on a pixel-by-pixel basis using a mathematical cross-tabulation matrix ("From-To Change Detection Matrix").

  • Every unique class-to-class trajectory is assigned a distinct integer and an associated color lookup table (LUT) to visualize discrete land surface conversions.

• Detailed Analysis of the Fort Moultrie, SC Case Study (1982 to 1988):

  • Evaluates Landsat TM scenes over Fort Moultrie, SC acquired on November 9, 1982, and December 19, 1988.

  • Delineates 9 distinct land cover classes:

  • 1: Developed/Exposed Land

  • 2: Cultivated Land

  • 3: Grassland

  • 4: Upland Forest

  • 5: Estuarine Emergent Wetland

  • 6: Riverine Aquatic Beds

  • 7: Palustrine Forested Wetland

  • 8: Water

  • 9: Estuarine Unconsolidated Bottom

  • Creates an $81$-cell cross-tabulation change matrix (cells 1 to 81):

  • Diagonal cells (1, 11, 21, 31, 41, 51, 61, 71, 81) represent unchanged, stable land cover, mapped to Black ($\text{RGB} = 0, 0, 0$).

  • Unselected transitional classes are mapped to White ($\text{RGB} = 255, 255, 255$).

  • New Developed/Exposed Land: Cells 10, 19, 28, 37, 46, 55, 64, 73 represent conversions from non-developed classes into Developed/Exposed Land, mapped to bright Red ($\text{RGB} = 255, 0, 0$).

  • New Estuarine Unconsolidated Bottom: Cells 9, 18, 27, 36, 45, 54, 63, 72 capture conversions to Estuarine Unconsolidated Bottom, mapped to Yellow ($\text{RGB} = 255, 255, 0$).

• Urban Expansion Application (Phoenix, AZ Case Study):

  • Evaluated across the Phoenix metropolitan area comparing independent NLCD classifications between 2001 and 2011.

  • Maps new low-, medium-, and high-intensity urban build-out (highlighted in red) converting natural desert shrubland and agricultural parcels.

• Comparative Trade-Offs of Post-Classification Comparison:

  • Advantages:

  • Intuitive, easily understood, and widely implemented across natural resource agencies.

  • Less vulnerable to inter-date sensor calibration mismatches or slight atmospheric disparities because each image is classified independently.

  • Yields an explicit "from-to" transition matrix detailing exact conversion pathways.

  • Disadvantages:

  • Multiplicative Error Propagation: The final change map compounds errors present in both underlying classification maps. If Map 1 has an accuracy of $80\%$ and Map 2 has an accuracy of $80\%$, the composite change map accuracy can drop to roughly $64\%$ ($0.80 \times 0.80$).

  • Ineffective at mapping subtle, gradual canopy alterations or continuous ecological modifications (e.g., slight pest infestations or canopy thinning).

Advanced Change Analysis: Vector Analysis and Dense Time Series

• Spectral Change Vector Analysis (SCVA):

  • Models transitions of land cover features as mathematical vectors operating in a multidimensional spectral coordinate space (e.g., Red versus Near-Infrared).

  • Every vector is characterized by two distinct metrics:

  • Vector Direction: Identifies the qualitative nature or type of land cover conversion based on relative shifts in band brightness.

  • Vector Magnitude (Intensity): Quantifies the overall physical severity or extent of the change.

  • Directional Trajectories in Red-NIR Feature Space:

  • Growing Vegetation / Revegetation: Increasing NIR reflectance accompanied by decreasing Red reflectance ($\Delta\text{NIR} > 0$, $\Delta\text{Red} < 0$).

  • Wildfire Burn / Forest Clearance: Decreasing NIR reflectance accompanied by increasing Red reflectance ($\Delta\text{NIR} < 0$, $\Delta\text{Red} > 0$).

  • Flooding / Standing Water Inundation: Decreasing reflectance in both Red and NIR wavelengths ($\Delta\text{NIR} < 0$, $\Delta\text{Red} < 0$) due to strong specular water absorption.

• Physiological Spectral Foundations of Leaf Desiccation:

  • Laboratory spectral analyses of Magnolia grandiflora leaves undergoing progressive drying ($100\%$, $75\%$, $50\%$, $25\%$, and $5\%$ relative water content across $0.4\,\mu\text{m}$ to $2.4\,\mu\text{m}$) provide the physical baseline for change vector directions.

  • As moisture content decreases, reflectance in water absorption bands (particularly near $1.4\,\mu\text{m}$ and $1.9\,\mu\text{m}$ in the SWIR) rises dramatically, shifting vegetative pixels along predictable desiccation vectors.

• Dense Satellite Time Series Analysis:

  • Historically limited to coarse spatial resolution sensors (e.g., AVHRR, MODIS), dense temporal analysis at $30\,\text{m}$ spatial resolution became universally viable following the free, open-access release of the complete USGS Landsat archive in October 2008.

  • Evaluates multi-year, multi-date pixel trajectories across decades (e.g., continuous tracking from January 1994, through September 2005, to June 2011) to capture forest thinning, clear-cuts, and subsequent agricultural conversion.

• Algorithmic Temporal Modeling and Change Points (Kennedy et al., 2007):

  • Replaces discrete bi-temporal snapshot comparisons with continuous trajectory fitting algorithms that detect mathematical "change points" and structural break points in reflectance profiles (e.g., Landsat Band 5 SWIR reflectance).

  • Delineates four distinct temporal trajectory archetypes:

  • a) Abrupt Disturbance: Stable pre-disturbance baseline $p_1$, followed by an abrupt, step-like increase in Band 5 SWIR reflectance occurring at break point year $p_0$, stabilizing at elevated post-disturbance level $p_2$.

  • b) Disturbance and Revegetation: Abrupt step increase at break point $p_0$ to peak reflectance $p_2$, followed by a steady, negative recovery slope $p_3$ as canopy vegetation regrows toward post-recovery baseline $p_4$.

  • c) Gradual Revegetation: Continuous long-term decline in SWIR reflectance over time starting from initial degraded state $p_1$ along slope $p_0$ toward vegetated baseline $p_2$.

  • d) Revegetation to a Stable State: Gradual decline from initial clear-cut state $p_1$ along recovery trajectory $p_0$ until reaching a definitive transition point $p_3$, where it levels off into a stable, mature closed canopy state $p_2$.

  • Dense time series analysis allows researchers to separate permanent, anthropogenic land conversions (e.g., forest converted permanently to row-crop agriculture displaying persistent seasonal oscillations) from temporary, cyclical forest clearing and vegetative regeneration.