Comparing the performance of multivariate and univariate time series models in detecting global land cover change

By Duyên Bùi

Abstract
In the past two decades, numerous land cover change algorithms have emerged, aiding land managers and policymakers in gaining more insights into the impact of land cover change on climate. Multivariate time-series detection models recently gained attention, yet their performance in a global land cover context remains unclear. This research evaluates the effectiveness of multivariate models compared to univariate models in detecting global land cover change, testing three models: BFAST Lite (univariate model), COLD (multivariate model), and DRMAT (multivariate model). Using an eight-year-long Landsat time series dataset, each model was thoroughly calibrated: the validation set contained over 30,000 sample sites worldwide. A modified BFAST Lite version for multiple band inputs was developed. All models tended to significantly overpredict, with low F1 scores when applying NDVI input; however, BFAST Lite and COLD performed better than DRMAT. The performance of the tested models (BFAST Lite and DRMAT) improved with six band inputs, especially for multivariate BFAST Lite, making it the most suitable model for detecting global land cover change. However, COLD showed a decreased accuracy, possibly due to inadequate cloud masking. Multivariate BFAST Lite also excelled in detecting changes across six thematic land cover classes: crop, herbaceous, woody, vegetation, and non-vegetation. Further research can fully unveil the potential of multivariate BFAST Lite by finding suitable pre/postprocessing methods to address inter-correlation between bands.

Keywords: multivariate models; univariate models; global land cover change; change detection.