Land-Use / Land-Cover Change Detection for Coimbatore District Using Landsat Imagery and Random Forests

Land-Use / Land-Cover Change Detection for Coimbatore District is an advanced M.Sc (IT) study that measures how a fast-industrialising Tamil Nadu district actually changed on the ground between two satellite epochs, using free Landsat imagery and a supervised random-forest classifier — with the accuracy of every claim assessed, not assumed.

Coimbatore is a textbook study area: a manufacturing and textile hub ringed by irrigated cropland, coconut plantations and the Western Ghats' forest edge. Two epochs are analysed — 2014 and 2024 — using Landsat 8/9 Collection 2 Level-2 Surface Reflectance scenes downloaded from USGS Earth Explorer, complemented by NRSC's Bhuvan LULC 1:50,000 maps as reference layers.

The method chain is deliberately classical and examinable: cloud masking from the QA_PIXEL band, median seasonal composites, spectral indices (NDVI, NDBI, NDWI), training samples built from the Bhuvan reference plus careful visual interpretation, a random-forest classification into six classes (built-up, cropland, plantation, forest, water, other/fallow), and a formal accuracy assessment — confusion matrix, overall accuracy, kappa and per-class recall — against an independent validation sample. A post-classification change matrix then quantifies the decade: hectares of cropland converted to built-up, plantation loss, forest stability at the ghats' edge.

The comparison against a maximum-likelihood classifier and the cross-check of class areas against the Bhuvan 50K layer give the study the measured rigour expected of an advanced project. All software is free; all data is free; the study is fully reproducible from the documented steps.

Coimbatore district is industrialising rapidly, but its conversion of cropland and plantation to built-up land is documented only fragmentarily — in sanctioned plans, news reports and occasional academic studies whose classifications carry no accuracy assessment. Planning bodies and researchers therefore argue from impressions, and national products like Bhuvan's 1:50,000 LULC maps, though excellent, are too coarse in space and time to answer a district-specific question about what changed in a particular decade.

Existing local change-detection studies typically report eye-catching 'urban expansion' figures without independent validation samples, without kappa, without per-class accuracy, and without comparing against any baseline classifier — which means their headline numbers cannot be trusted and their errors are invisible. There is a need for a methodologically complete, reproducible change-detection study: two properly matched Landsat epochs, a documented training regime, a formal accuracy assessment with an independent sample, a benchmarked classifier comparison, a post-classification change matrix with impossible-transition screening, and every step documented so that another student can reproduce or extend it for a different district.

Abstract

Land-use and land-cover (LULC) change is the visible ledger of development: fields become layouts, plantations become industrial plots, forest edges hold or retreat. This study quantifies that ledger for Coimbatore district between 2014 and 2024 using two epochs of free Landsat 8/9 Surface Reflectance imagery, a supervised random-forest classifier with spectral indices, and a formal accuracy assessment, followed by a post-classification change matrix. NRSC's Bhuvan LULC 1:50,000 maps serve as both reference for training and independent sanity check for results. The deliverables are reproducible notebooks, accuracy-assessed change maps, class-area statistics and a dissertation ready for the viva.

Introduction

Municipal and planning records describe what was approved to be built; satellite imagery records what actually exists. For Coimbatore — industrial corridors, an airport ring, highway expansion and a forest boundary under pressure — the difference between the two is precisely what a change-detection study contributes. Free Landsat archives provide a consistent, calibrated 30 m record across decades, and Collection 2 Level-2 Surface Reflectance removes the atmospheric-correction step that used to block student projects.

Existing system and gap

Bhuvan's national LULC mapping is authoritative but periodic and generalised at 1:50,000 — it cannot answer a district-specific question about which decade a particular conversion happened. Published local studies in Indian journals frequently skip the parts that make a classification claim scientific: independent validation samples, kappa statistics, per-class accuracy, or any comparison against a baseline classifier. Many also mix epochs, sensors or seasons carelessly, so their 'change' includes phenology rather than land use. The gap this study fills: a documented, accuracy-assessed, single-sensor, season-matched two-epoch change detection for one district, with the classifier benchmarked against maximum likelihood, errors analysed rather than hidden, and the workflow reproducible from free data.

Proposed study

• Data: Landsat 8/9 C2 L2 Surface Reflectance scenes for the 2014 and 2024 winter seasons (minimum cloud, stable phenology); Bhuvan LULC 1:50K as reference.
• Preprocessing: cloud/shadow masking from QA_PIXEL, per-scene scaling, median compositing per epoch, index generation (NDVI, NDBI, NDWI).
• Classification: random forest (scikit-learn) with a stratified training sample built from Bhuvan-derived polygons plus visual interpretation on high-resolution basemaps; six information classes; benchmarked against maximum likelihood (implemented in Python) on identical samples.
• Assessment: independent stratified validation sample; confusion matrix, overall accuracy, kappa, per-class producer's/user's accuracy; misclassification analysis by class pair.
• Change analysis: post-classification comparison 2014 → 2024; change matrix in hectares; plausibility screening of impossible transitions against imagery; class-area cross-check against Bhuvan.

Feasibility study

• Technical: downloads, masking and classification all run on a standard laptop; rasterio/GDAL handle the imagery and scikit-learn trains the forest in minutes. The image-processing and machine-learning coursework maps directly onto the method chain.
• Economic: zero cost — Landsat data is free and unrestricted; Bhuvan thematic downloads are free with registration; every software component is open source.
• Operational: no field campaign is needed; validation samples come from imagery and published maps. The main risk — monsoon cloud — is managed by winter scenes and multi-month compositing.

In scope

• Coimbatore district (Tamil Nadu) as the study area; two epochs (2014, 2024) with winter-season Landsat 8/9 imagery as the primary sensor record.
• Six information classes at 30 m Landsat resolution; post-classification change detection; accuracy assessment with independent samples.
• Deliverables: reproducible notebooks, accuracy-assessed maps, change matrix, dissertation and viva deck.

Out of scope

• Sub-30 m mapping (Sentinel-2 at 10 m or drone imagery) — noted as future work to keep the two epochs sensor-consistent.
• Continuous change models and trajectory analysis across more than two epochs.
• Causes-of-change modelling (economic or demographic drivers) — the study measures and locates change, not its determinants.
• Real-time or near-real-time monitoring; this is a retrospective, dual-epoch study.

Research design: quantitative, image-based Earth-observation study with a formal accuracy-assessment protocol; no human participants, so no ethics clearance is required. Landsat and Bhuvan data are used under their published open-use terms with attribution.

Work division (team of 2): Member A owns data acquisition, preprocessing and QGIS composition; Member B owns sampling, classification and assessment; both co-own the change analysis and write the dissertation in jointly agreed chapters — and each member can demonstrate every stage, because the viva examines individuals.

Analysis protocol (fixed before classification): winter-season scenes only; a single classification scheme across epochs; stratified random validation sample independent of training polygons; overall-accuracy target ≥ 85% with per-class user's accuracy ≥ 75% for the built-up and cropland classes that drive the headline claims; every rejected scene and merged sample recorded in a decision log.

Although this is a study rather than a software product, the workflow is an engineered pipeline, and drawing it keeps the analysis honest — every box is a documented, rerunnable step.

flowchart TD
  A["USGS Earth Explorer: Landsat 8/9 C2 L2 SR scenes (2014, 2024, winter)"] --> B["Cloud/shadow mask from QA_PIXEL + reflectance scaling"]
  B --> C["Median composites per epoch (winter window)"]
  C --> D["Spectral indices: NDVI, NDBI, NDWI"]
  E["Bhuvan LULC 1:50K + visual interpretation"] --> F["Stratified training samples (6 classes)"]
  D --> F
  F --> G["Random forest (scikit-learn)"]
  F --> H["Maximum-likelihood baseline"]
  D --> G
  D --> H
  G --> I["Classified maps 2014 and 2024"]
  H --> I
  I --> J["Independent validation: confusion matrix, kappa, per-class accuracy"]
  J --> K{"OA >= 85% and key classes >= 75%?"}
  K -->|no| F
  K -->|yes| L["Change matrix 2014 to 2024 (hectares)"]
  L --> M["Impossible-transition screening + Bhuvan area cross-check"]
  M --> N["Figures, tables and dissertation chapter"]

Three design decisions deserve defence in the viva. First, one sensor family and one season per epoch: Landsat 8 in 2014 and Landsat 8/9 in 2024, both winter, so that 'change' reflects land use rather than phenology or sensor differences. Second, the validation sample is independent and stratified — drawn after classification from imagery interpretation, never reused from training — which is what makes the confusion matrix a statement about the map rather than a memory test. Third, the benchmark comparison against maximum likelihood converts a subjective 'our classifier is good' into a measured result on identical inputs, exactly as a benchmarking exercise should.

• Add Sentinel-2 (10 m) for a recent epoch to quantify how much of the mixed-pixel error at 30 m disappears, keeping Landsat for the historical anchor.
• Extend to a three-epoch trajectory (2004–2014–2024) to distinguish steady conversion from recent acceleration.
• Move classification to Google Earth Engine for all-district processing at higher temporal frequency without local storage constraints.
• Couple the change map with road-network buffers to test whether conversion clusters along highways.
• Package the sampling and assessment protocol as a reusable template so other students can replicate the study for their own districts.