A Proof-of-concept for non-destructive lettuce nutrient assessment using vision transformers and residual networks on raw hyperspectral imagery

• First application of Vision Transformers to hyperspectral nutrient regression • Achieved a test R² of 0.65 for nitrogen without radiometric calibration or preprocessing • Attention maps reveal biochemically relevant spectral regions (NIR and SWIR) • Comparable performance to ResNet architectures on small dataset (928 images) • End-to-end learning eliminates preprocessing barriers for agricultural deployment Hyperspectral imaging (HSI) shows promise for non-destructive measurement of tissue nutrient concentrations in plants. However, many HSI analysis approaches rely on extensive preprocessing, traditional statistical methods, and pixel averaging that require subject matter expertise and sacrifice spatial information. This reduces their prospects for adoption in precision agricultural applications, where automated, end-to-end approaches are needed. To address this gap, we explored end-to-end deep learning (DL) to process raw hyperspectral data (400-1100nm). This study presents a novel application of vision transformers (ViT) for a regression task in precision nutrient management, comparing it to popular residual network (ResNet) architectures (ResNet-34, ResNet-50, and ResNet-101) and providing a proof-of-concept for nutrient status estimation from raw proximal hyperspectral imagery. To explore the applicability of these approaches, we conducted a pilot-scale longitudinal hydroponic lettuce growth study with destructive sampling, imaging plants grown with different applied nutrient concentrations under pilot-scale greenhouse conditions. Results reveal considerable variability within and between these models under the same data splits, with ViT and ResNet-34 showing particular promise for N estimation, achieving an average test R 2 of 0.65 and 0.73 R 2 without radiometric calibration and with limited preprocessing (resizing and normalization). These results demonstrate the potential for developing DL approaches that support end-to-end nutrient control automation in precision agriculture.