Integrating Sentinel-2 Derivatives to Map Land Use/Land Cover in an Avocado Agro-Ecological System in Kenya

Abstract

Reliable, readily available, and appropriate land use/land cover (LULC) information is fundamental for coherent land and natural resources management, especially in data-scarce environments that are complex and heterogeneous. This study took a holistic approach for evaluating the classification accuracy of LULC classes in an avocado production system in Kenya using different classification scenarios and the random forest (RF) machine learning (ML) algorithm in Google Earth Engine (GEE). We integrated sentinel-2 (S2) spectral bands, vegetation indices (VIs), and phenological variables in two classification routines, pixel- and polygon-based procedures, and assessed their performance and importance in mapping LULC classes. To assess the LULC classification accuracy, a confusion matrix and a pattern-based assessment were used. This study demonstrated that the polygon-based classification procedure was the best (overall accuracy > 75% for confusion matrix and > 0.7 for pattern-based accuracy assessment methods) in mapping out complex landscapes when compared to the pixel-based classification procedures. Combining S2 reflectance with vegetation indices, red-edge (RE) vegetation indices, and phenological metrics can considerably improve LULC classification accuracy.