Simultaneous hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning across VNIR and SWIR spectral regions
Smart Agricultural Technology · 2026-12 · 15 · 102744
DOI: 10.1016/j.atech.2026.102744 ↗Kurz gefasst
Fachartikel in der Zeitschrift Smart Agricultural Technology, 2026-12. Verfasst von Niharika Vullaganti und 4 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: USDA Agricultural Research Service.
Abstract (Originalsprache)
Rapid and scalable assessment of soil nutrient variability is critical for precision nutrient management but remains constrained by conventional laboratory methods. This study investigates whether multitask deep learning can simultaneously classify soil nitrogen (N) and phosphorus (P) from hyperspectral data, and whether task-specific expert gating improves performance over shared representations. Six models combining three backbone architectures (CNN, ResNet-18, DenseNet) with shared-backbone multitask and multi-gate mixture-of-experts (MMoE) frameworks were evaluated using 284 samples from two agricultural sites in North Dakota, scanned in the laboratory with VNIR (400–1000 nm) and SWIR (900–1700 nm) cameras. Models were compared with repeated stratified 5-fold cross-validation at the sample level, and the final model was evaluated on a second set of cross-validation splits of the same samples that was not used for model selection. The VNIR, SWIR and fused inputs and the six architectures performed similarly, with no statistically significant differences, and the MMoE framework did not improve on the standard multitask models. Spectral preprocessing and feature selection also did not improve classification. This simplest configuration, a multitask CNN with VNIR data, achieved sample-wise accuracies of 0.53 ± 0.07 for N and 0.50 ± 0.05 for P with most errors falling into an adjacent class (within-one-class accuracy of 0.85 and 0.92), and performed as well as separate single-task models. When the model was trained on one site and tested on the other, accuracy fell to 0.37 for N and 0.15 for P. The model therefore learned relationships that held within the sampled fields but did not transfer to a new site. Validation across more sites and under field conditions is required before practical use.
Förderung
- USDA Agricultural Research Service
Zitieren
APA
Vullaganti, N., Zhang, X., Aderholdt, W., Overby, P., & Sun, X. (2026). Simultaneous hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning across VNIR and SWIR spectral regions. Smart Agricultural Technology, 15, 102744. https://doi.org/10.1016/j.atech.2026.102744
Vancouver
Vullaganti N, Zhang X, Aderholdt W, Overby P, Sun X. Simultaneous hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning across VNIR and SWIR spectral regions. Smart Agricultural Technology. 2026;15:102744. doi:10.1016/j.atech.2026.102744
BibTeX
@article{Vullaganti2026,
author = {Vullaganti, Niharika and Zhang, Xiaomo and Aderholdt, William and Overby, Paul and Sun, Xin},
title = {Simultaneous hyperspectral classification of soil nitrogen and phosphorus using multitask deep learning across VNIR and SWIR spectral regions},
journal = {Smart Agricultural Technology},
year = {2026},
volume = {15},
pages = {102744},
doi = {10.1016/j.atech.2026.102744}
}
EUVIMED