STAR: Soil texture analysis recognizer integrating domain-adaptive transfer learning with NIR spectroscopy

Soil texture is a critical parameter influencing water retention, nutrient dynamics, and plant growth, with direct implications in agriculture, environmental management, and land-use planning. While near-infrared (NIR) spectroscopy has emerged as a rapid, non-destructive and environmental friendly alternative to conventional lab-based methods, its transition to practical deployment remains limited by constrained model generalization, high dependency on annotated data, and cross-domain inconsistencies. To address these challenges, here we introduce a Soil Texture Analysis Recognizer (STAR), a compact and intelligent NIR-based device for precise soil texture classification. STAR employs a domain-adaptive deep learning modeling strategy that incorporates two novel algorithmic components: a transfer learning-oriented spectral preprocessing method, Transfer Multiplicative Scatter Correction (TMSC), to reduce spectral distributional shifts, and the Selective Enhanced Transfer Adaptive Boosting (SETAB) framework to enhance model adaptability under cross-regional and inter-domain imbalance conditions. Validated using local soil samples from Sichuan Province, STAR achieved an overall classification accuracy of 85.0 % with a Kappa coefficient of 0.78 across 5 local soil texture classes, and successfully identified previously unseen soil texture types, including loamy sand and sandy loam, with accuracies of 100.0 % and 66.7 %, respectively. These findings underscore STAR's robust generalization capability and its practical utility for soil texture analysis. Beyond this, STAR platform provides a scalable pathway for broader NIR-based soil property measurement, and the proposed modeling strategy offers a feasible bridge between deep learning-assisted spectral modeling and real-world applications.