An intercomparison of national and global land use and land cover products for Fiji

Abstract

Here, a methodology to generate national-scale annual 10 m spatial resolution land use and land cover maps for Fiji (Fiji LULC) is presented. A training dataset of 13,419 points with a LULC label across three years from 2019 to 2021 was generated alongside a nationally representative test dataset of 834 points. These data were used to train a random forests model to convert an image stack of pre-processed Sentinel-2 surface reflectance data and topographic spatial layers into an annual categorical LULC map. When evaluated against the test dataset, the model has an overall accuracy of 83 % (SE: 2.1 %).The Fiji LULC map was compared to three global 10 m spatial resolution land cover products: Google’s Dynamic World, ESRI LULC, and ESA’s WorldCover v200. These maps were compared statistically using the independent test dataset and in several case study applications (e.g. agricultural monitoring and disaster impacts mapping). The Fiji LULC had a higher overall accuracy than the three global LULC products and aligned more closely with a high-quality field survey of over 2500 rice fields (i.e. Fiji LULC classified 88 % of the rice fields as agricultural compared to 60.6–15.7 % in the global LULC products). A comparison of the overlap between the agricultural class of the four LULC maps with a flood mask following Tropical Cyclone Yasa indicated that dataset choice has a substantial impact on estimates of the area of flooded croplands. The Fiji LULC map tends to capture agricultural land covers and smaller scale landscape features with more accuracy than the global products. This analysis illustrates the importance of assessing the performance of global LULC products in particular locations and for specific applications. As demonstrated here, the choice of LULC product could impact subsequent analysis and monitoring tasks. To support these LULC product comparisons, an open-source Python package for computing performance metrics for LULC maps when reference data have different strata to map classes has been published. Further, the training data, test data, and national-scale maps for Fiji have been produced for 2019 to 2022 and are available as open source products on the Pacific Data Hub.

Keywords

• Land cover
• Fiji
• Sentinel-2
• Machine learning