Spatial and spectral analysis of fairy circles in Namibia on a landscape scale using satellite image processing and machine learning analysis

Abstract

Fairy circles (FCs) are a unique phenomenon characterized by circular patches, 4–10 m in diameter, of bare soil within a vegetated matrix. This project aimed to study the spatial and spectral characteristics of FCs on a landscape scale in Namibia. The specific objectives of this research are (1) processing satellite observations to explore the FCs distributions by applying statistical analysis and deep machine learning algorithms; (2) analyzing the FCs' geometric attributes to retrieve their spatial patterns regarding topographic features nearby. The FCs were classified within 25 km2 by processing 15 input layers through a convolutional neural network (CNN) model. The layers include four WorldView2 spectral bands, derived vegetation, biocrust, and mineral indices, and textural characteristics. The FCs’ geometry was extracted, and spatial autocorrelation was performed. By labeling 1600 FCs and using the CNN model, 14,536 FCs were mapped with 0.97% accuracy and a binary cross-entropy loss function value of only 0.01. Field measurements and laboratory analysis justified the need to use spectral indices for the model. Unique elongated FCs, clustered by hotspot analysis, were quantified and mapped along watercourses in alluvial fans with notable connectivity. On a landscape scale that has not yet been studied, spatial and spectral analyses became possible only with valuable remote sensing retrievals, deep statistical analysis, and machine learning algorithms.

Keywords

• CNN model
• Spectral indices
• Texture variables
• Spatial autocorrelation