Mapping Lantana camara Invasion in the Nilgiris: GBIF Occurrence Data and MaxEnt Habitat Modelling

Mapping Lantana camara Invasion in the Nilgiris is an M.Sc botany dissertation that turns free public biodiversity data into a defensible habitat-suitability map for one of India's worst invasive shrubs — with the model's assumptions and weaknesses stated rather than hidden.

Lantana camara, introduced to India as an ornamental, has invaded forest floors, plantations and roadsides across the Western Ghats, suppressing native regeneration and altering fire behaviour. The Nilgiri Biosphere Reserve — sholas, tea estates and moist deciduous forest interleaved — is a textbook case. Field surveys of such terrain need permits, vehicles and months; yet thousands of occurrence records already exist in the Global Biodiversity Information Facility (GBIF), from herbaria, citizen-science platforms and published studies. The problem is not access but quality: coordinates sit on herbarium sheets, at preserve centroids, in oceans, and along roads.

This study treats that problem as its core. Occurrence records for L. camara within the reserve extent are downloaded from GBIF with dataset DOIs recorded, cleaned with the CoordinateCleaner battery (duplicates, centroids, zero coordinates, ocean points) and spatially thinned to reduce sampling bias. Nineteen WorldClim bioclimatic variables plus elevation describe the environment; a MaxEnt 3.4 maximum-entropy model is fitted with regularisation and feature classes selected by cross-validation, and background points drawn with a target-group approach to counter bias. Model quality is reported with AUC and its caveats, response curves and jackknife importance. The deliverable is a suitability map the forest department's language can absorb, plus an honest chapter on what occurrence-data modelling can and cannot claim.

Lantana camara suppresses native regeneration, changes fire regimes and reduces the usability of grazing land across the Nilgiris, yet control effort is allocated by impression — where forest staff happen to see dense stands — rather than by modelled habitat suitability. The occurrence data needed for that model already exist in public repositories, but they are uncleaned, spatially biased and rarely used: coordinates sit at herbarium-sheet precision, cluster along roads and towns, and occasionally fall in the ocean. Using such data raw produces maps that look authoritative and are wrong in exactly the places management cares about.

There is a need for a reproducible modelling workflow that states its data sources with citations, cleans coordinates with documented, repeatable rules, corrects sampling bias in both presence and background data, selects model settings by cross-validation rather than default, and reports model quality with its interpretation caveats. Delivering that workflow for L. camara in the Nilgiri Biosphere Reserve produces a suitability map the forest department can read, a variable-importance analysis ecologists can argue with, and a methods template other invasive-species projects in the department can reuse — all from data that is already public.

Abstract

Invasive shrubs reshape forest ecosystems faster than field surveys can track them. Lantana camara, among the most damaging invasive plants in the Western Ghats, is mapped here through a reproducible species-distribution-modelling workflow: GBIF occurrence records cleaned and bias-corrected, WorldClim bioclimatic and elevation layers as predictors, and a MaxEnt model fitted, evaluated and projected across the Nilgiri Biosphere Reserve. The output is a habitat-suitability map with a documented uncertainty story, produced at zero cost from open data.

Introduction

The Nilgiri Biosphere Reserve interleaves protected forest, shola patches, tea and coffee estates and dense road networks — exactly the disturbed-mosaic landscape Lantana camara exploits. Managers know the shrub is everywhere, but 'everywhere' is not a management plan: control programmes need to know where habitat is most suitable, where spread is likely next, and which environmental gradients drive suitability. Field surveys answer those questions slowly; public occurrence data plus distribution modelling answer them cheaply and repeatably.

Existing approaches and the gap

Occurrence records for L. camara exist in thousands within GBIF, but they are heterogeneous: museum and herbarium specimens with imprecise georeferences, citizen-science observations clustered near towns and roads, and published checklists whose coordinates are coarse. Studies that skip cleaning inherit those errors; studies that skip bias treatment let the road network pose as climate response. Many Indian student projects also quote AUC as if a single number settled model quality. The gap this project fills: a fully documented, standards-based workflow — GBIF citation practice, CoordinateCleaner flagging with a decision log, spatial thinning, target-group background sampling, regularisation chosen by cross-validation, and evaluation that pairs AUC with its honest interpretation — applied to a landscape where the result has direct management meaning.

Proposed study

• Data: L. camara occurrence records within the Nilgiri Biosphere Reserve bounding extent, downloaded from GBIF with dataset DOIs archived.
• Cleaning: CoordinateCleaner tests (duplicates, country and preserve centroids, zero and ocean coordinates, per-species outliers), with every flagged point's fate recorded; spatial thinning at a fixed minimum distance.
• Environment: WorldClim 2.1 bioclimatic variables and elevation, cropped to the study extent, checked for collinearity (r > 0.85 drops one of each pair).
• Modelling: MaxEnt 3.4 with feature-class and regularisation selection by cross-validated AUC, bootstrap replicates, target-group background points; response curves and jackknife importance per predictor.
• Output: habitat-suitability raster and classified map for the reserve, variable-contribution table, and a limitations chapter that treats sampling bias as a finding, not an embarrassment.

Feasibility

• Technical: everything is open — GBIF, WorldClim, MaxEnt, QGIS, R — and the data volume is trivial for a laptop; the methods are standard and examinable.
• Economic: zero cost; no field campaign, permits or equipment.
• Operational: the workflow is schedulable in 20 weeks; the main risk (thin coverage in parts of the reserve) is handled by a pre-declared fallback to a Western Ghats extent model with the Nilgiris as projection area.

In scope

• One species (Lantana camara) in one landscape: the Nilgiri Biosphere Reserve and its bounding extent in Tamil Nadu, Kerala and Karnataka.
• GBIF-mediated occurrence records as the presence dataset; WorldClim 2.1 (19 bioclim variables plus elevation) as the environmental stack.
• MaxEnt as the single modelling algorithm, with cross-validated setting selection, bias-aware background sampling and standard evaluation outputs.
• A suitability raster, classified map, variable-contribution tables and a scripted, re-runnable workflow.

Out of scope

• New field surveys: no ground-truthing campaign is possible within the slot, and absence of field validation is a stated limitation rather than a hidden one.
• Population dynamics, spread-rate forecasting and management-cost optimisation — the model maps suitability, not time-to-invasion.
• Other invasive species beyond a discussion of reusing the workflow.
• Future-climate projections using CMIP scenarios — noted as the natural next phase.
• Genotypic or cytotype-level analysis of the L. camara complex.

Research design: a quantitative secondary-data study with an ecological-modelling deliverable — presence-only species distribution modelling under a written protocol with pre-declared cleaning and evaluation rules.

Materials and analysis plan: GBIF occurrence data (cited per dataset), WorldClim 2.1 at 30 arc-seconds, MaxEnt 3.4, QGIS, and R packages CoordinateCleaner, terra, dplyr and ENMTools. Cleaning rules are fixed in the protocol: duplicate removal by coordinate and date, centroid and ocean tests, per-species outlier flagging, then thinning — every rule applied mechanically with a logged decision for each flagged record. Background points follow a target-group approach (backgrounds drawn from occurrences of a reference set of well-recorded forest plants in the same extent) to counter road-and-town bias; the collinearity screen keeps one variable per correlated pair, chosen for ecological interpretability and recorded in the protocol. Model performance is reported as mean cross-validated AUC across replicates with its spread, and the suitability raster is classified into four management-facing categories whose thresholds are fixed before the map is made.

The workflow is a four-stage pipeline — acquire, clean, model, map — with a decision log beside every stage, so that each map pixel is traceable back through the model, the cleaned records and the raw downloads.

flowchart TD
  A["GBIF download: Lantana camara occurrences within the reserve extent"] --> B["Raw archive with dataset citations and DOIs"]
  B --> C["CoordinateCleaner: duplicates, centroids, zeros, ocean, outliers"]
  C --> D{"Flagged record: keep, correct or discard?"}
  D -->|decision logged| E["Cleaned presence set"]
  E --> F["Spatial thinning at fixed minimum distance"]
  G["WorldClim 2.1 bioclims + elevation"] --> H["Crop to extent; collinearity screen r > 0.85"]
  H --> I["Predictor shortlist (fixed in protocol)"]
  F --> J["MaxEnt 3.4: feature classes + regularisation by cross-validation"]
  I --> J
  K["Target-group background points"] --> J
  J --> L["Evaluation: cross-validated AUC, response curves, jackknife"]
  J --> M["Suitability raster -> classified map of the reserve"]
  L --> M
  M --> N["Limitations chapter: sampling bias, transferability"]

Three design decisions shape the model's credibility. First, every flagged occurrence has a recorded fate: the cleaning stage is a documented filter, not an invisible one, because silent deletions are how distribution models acquire convenient biases. Second, the background is treated as carefully as the presence data — target-group backgrounds share the road-and-town bias of the occurrence records, so the model contrasts the species' habitat against the sampling effort's habitat, which is the comparison the statistics actually require. Third, model settings are chosen, not defaulted: feature classes and regularisation are selected by cross-validated performance on this landscape, and the thresholds that turn the continuous suitability raster into management categories are fixed before the map exists, so the classification cannot be tuned to a preferred picture.

• Project the fitted model onto WorldClim future-climate layers to ask where suitability is heading under warming scenarios.
• Add the forest department's own field observations as an independent validation set where they can be shared.
• Extend the workflow to co-occurring invasives (Chromolaena odorata, Ageratina adenophora) and compare their niche overlaps with ENMTools.
• Couple the suitability map with a road-distance layer to model accessibility of control effort rather than suitability alone.
• Submit the final records back to GBIF through a department dataset, closing the loop between using and contributing open biodiversity data.