🐿️ Spatial Ecology: Predicting Red and Grey Squirrel Distribution

🧩 Overview

This project analyses the spatial distribution of red and grey squirrels in Scotland using environmental and occurrence data.
It integrates GIS-based spatial processing and machine learning models to understand how habitat features and species interactions shape their range.

The study demonstrates how spatial autocorrelation can bias model accuracy and how spatial cross-validation (spCV) provides a more realistic measure of predictive performance compared to standard k-fold CV.

---

🎯 Key Highlights

• Objective: Identify the influence of land-cover and proximity to grey squirrels on red squirrel presence.
• Data:
  • Species occurrence data (Sciurus vulgaris & Sciurus carolinensis).
  • UK Land Cover Map (LCMUK) raster data.
• Spatial scale optimisation: Tested buffer radii from 100 m – 2000 m, found optimal at 1200 m.
• Model comparison:
  • Random Forest (RF) vs Support Vector Machine (SVM).
  • Both under CV and spatial CV frameworks.
• Result:
  • RF accuracy dropped from 0.81 (CV) → 0.60 (spCV), showing over-optimism in standard validation.
  • SVM improved with parameter tuning, reaching ~0.73 under spatial CV.
• Reproducibility: Uses relative paths via {here} and standard project structure.

---

⚙️ Tech Stack

Category	Tools & Packages
Language	R
Spatial processing	terra, sf, raster, rgdal
Machine learning	mlr, randomForest, kernlab
Cross-validation	k-fold CV and Spatial CV (SpCV, SpRepCV)
Reproducibility	here, structured project folders
Visualisation	Base R plotting, raster maps, histograms

---

📁 Project Structure

├── data/                # occurrence & raster data
├── code.R               # R script
└── README.md

Data Directory

Place the following files here before running the analysis:

• Sciurus.csv – Red squirrel occurrence records
• Grey_squirrel_records.csv – Grey squirrel occurrence records
• LCMUK.tif – UK Land Cover Map raster
• Sciurus_SA.shp – Study area boundary shapefile

Note: These datasets are not included in the repository due to size and licensing restrictions.

---

🧠 Insights

• Spatial autocorrelation can inflate predictive accuracy if ignored.
• Spatial CV is essential for realistic model evaluation in ecological studies.
• Ecological processes often operate at intermediate spatial scales (~1 km), as shown by the 1200 m optimal buffer.
• Machine learning models like RF and SVM, when combined with sound spatial design, provide interpretable and generalisable predictions.