Predicting the spatial distribution of invasive species richness using the combination of machine learning and geostatistical algorithms

Spatial predictions of invasive plant species richness across large geographic areas are essential for assessing the effects of management practices across environmental gradients. Traditional spatial statistical algorithms often struggle with large-scale, high-dimensional datasets in the era of big data, limiting their analytical capabilities and predictive accuracy. Simultaneously, measuring and exploring the drivers of invasive species richness is crucial for both ecological research and land management aimed at controlling invasions. To address these challenges, two novel hybrid machine learning-geostatistical algorithms, elastic net regression residual combined with ordinary kriging (ElaNetOK) and elastic net regression residual combined with inverse distance weighting (ElaNetIDW), were constructed and applied across six ecoregions of mixed forests in the United States. Datasets of 29,228 mixed forest plots together with 51 ecological variables from the WorldClim 2.1 and the Forest Inventory and Analysis (FIA) were used to evaluate the performance of traditional geostatistical models and the two hybrid machine learning-geostatistical algorithms. Results demonstrate that ElaNetOK and ElaNetIDW effectively handle high-dimensional spatial datasets, reduce prediction uncertainty, and consistently outperform traditional geostatistical methods in both accuracy and stability. Quantitatively, ElaNetIDW achieved accuracy improvements over conventional IDW ranging from 0.88% to 6.90% across six ecoregions, with an average improvement of 3.25%. Similarly, ElaNetOK outperformed OK, with regional improvements between 1.10% and 7.11%, averaging 3.33%. Furthermore, spatial prediction maps reveal that the hybrid models preserve large-scale environmental gradients while capturing enhanced local spatial heterogeneity, producing more ecologically realistic spatial patterns. These findings demonstrate the potential of hybrid machine learning–geostatistical frameworks for large-scale ecological prediction and invasive species management. • Hybrid algorithms integrating machine learning and geostatistical algorithms were developed to predict invasive species richness. • Large-scale and high-dimensional variables extracted from WorldClim 2.1 and FIA datasets were used as auxiliary information for prediction. • The proposed hybrid algorithms outperformed traditional geostatistical methods in terms of prediction accuracy and stability. • Key variables influencing invasive species richness were identified and ranked by the hybrid algorithm, providing scientific support for invasive species management.