Prediction to Prevention: Comparative Machine Learning Models for Early Intervention of Dactyloctenium radulans in Australia

Weeds pose a significant threat to agricultural productivity, causing economic damage. A comprehensive understanding of the distribution of weed species is essential for formulating efficient management strategies. Despite its ecological significance, limited research has been conducted on the native Australian weed species Button grass ( Dactyloctenium radulans ). In this study, four species distribution models (SDMs) were employed to estimate its potential habitat suitability across the entirety of Australia. The four SDMs used in this study were Maxent, random forest (RF), boosted regression trees, and artificial neural networks (ANN). We conducted and collected occurrence records from field surveys and also used historical occurrence records from GBIF with environmental and soil variables to build robust predictive models. Our results suggested that seasonal temperature and soil pH were critical determinants of D. radulans distribution, contributing up to 37% of permutation importance. Among the four SDMs, Maxent outperformed other models in predictive accuracy (area under the curve = 0.90), while RF and ANN demonstrated distinct advantages in specific contexts. The integration of field and GBIF datasets enabled us to broaden our scope, and we identified southeastern Australia as a high-risk zone for D. radulans proliferation. This research underscores the utility of machine learning algorithms, including Maxent in ecological modelling, providing actionable insights for weed management and habitat restoration.