Improving Machine Learning Models of Land Cover Change Using Spatial Sample Weights Based on Patch‐Level Landscape Metrics

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Machine learning (ML) is prevalent in land change modeling for automating transition rule specification. However, aspatial strategies are typically used to mitigate the imbalances characterizing multitemporal land cover (LC) datasets. While spatialized cost‐sensitive learning strategies including spatial sample weights (SSWs) have demonstrated benefits for ML models, the spatial pattern variations linked to LC change events have not been considered. Therefore, the goal of this study is to enhance SSW strategies by regulating LC change sample importance with a novel patch weight (PW) term derived from patch‐level landscape metrics. An application example demonstrates Extreme Gradient Boosting Machine (XGB) models used to forecast LC change for the City of Kelowna, Canada, under various SSW configurations. The results showed numerous PW settings facilitated realistic quantities of change overall and for each LC class while reducing allocation errors compared to baseline configurations. This study advances spatialized cost‐sensitive learning strategies to enable land change modeling using ML.
Funding Information
  • Natural Sciences and Engineering Research Council of Canada (RGPIN‐2023‐04052)

This publication has 72 references indexed in Scilit: