Enhancing agricultural data interpretability and visualization with TabNet-driven feature extraction and Local Biplots

Modern agriculture faces escalating challenges that demand predictive models balancing accuracy and interpretability to support informed decision-making. While precision agriculture has advanced through high-resolution remote sensing technologies capturing spectral, thermal, and soil data, interpreting these complex, nonlinear datasets remains a significant challenge for agronomists. This study introduces the TabNet-informed UMAP-based Local Biplot (UL-Biplot) framework, which combines TabNet's attention-based feature attribution with a Local Biplot technique from the widely recognized Uniform Manifold Approximation and Projection (UMAP). The method provides an intuitive representation of non-stationary and non-linear data patterns, enhancing both global and cluster-level explainability. Also, we highlight feature loadings based on TabNet's attention scores and evaluate the framework on a synthetic benchmark and two real-world datasets: a forage grasses trial for breeder score prediction and the WMBD wheat dataset for LAI estimation. Quantitative results demonstrate that our framework achieved an R 2 = 0.79 ± 0.01 for LAI and delivered more consistent performance for breeder scores ( R 2 = 0.77 ± 0.01 ) relative to standard machine learning models. Moreover, TabNet's latent embeddings improved significantly in trustworthiness (from 0.94 to 0.98) and in class-neighbourhood preservation (0.047 versus 0.018 with conventional methods). The UL-Biplot also revealed distinct agronomic clusters associated with irrigation-induced greenness indices and moisture stress responses, providing actionable insights for site-specific crop management. • TabNet-informed UL-Biplot integrates attention-based feature selection with dimensionality reduction for interpretable agricultural data. • Combines TabNet’s sequential masks with local affine biplots for traceable visual analytics. • TabNet’s attention masks filter and rank features for Local Biplot reduction and global-axis projection. • UL-Biplot outperforms SVD in neighbourhood fidelity and class cohesion.