NIRS-Based Citrus Harumi Fruit Soluble Solid Content Detection Leveraging Invariant Representation Learning With Hash Consistency

Accurate detection of citrus Harumi fruit soluble solid content is critical for improving the storage management quality and grading accuracy. However, in near-infrared spectroscopy (NIRS) of heterogeneous Harumi fruit batches, both the redundancy caused by high collinearity and the influence of nontarget disturbances can significantly degrade the performance of existing detection instruments. To address these challenges, a detection instrument is designed, which consists of a novel intelligent detecting model HashNIRL, and an NIRS-based detecting hardware component. HashNIRL comprises two modules: a structure-inspired hash representation switch (SHRS) and an environment self-distillation module (ESM). To mitigate the redundancy in NIRS, SHRS exploits the spectral graph structure to extract band co-variation relationships and generates hash-based channel masks that activate stable channels within the spectral features, thereby separating invariant features from noisy ones. To distinguish perturbed features, ESM employs the soft hash codes produced by SHRS as anchors, and takes the invariant and perturbed features as soft channel compositions to guide the model to focus on consistent hash-based channel masks in the hash space, thus enhancing its ability. The hardware component integrates two halogen lamps to ensure high-quality spectral data acquisition. Experimental results on the self-built benchmark (NIROOD) demonstrate that HashNIRL outperforms existing state-of-the-art methods, achieving 6.57% and 3.23% reductions in average mean squared error (MSE) and mean absolute error (MAE), respectively. In practical scenarios, HashNIRL further lowers average MSE and fitting slope error by 10.66% and 40.24%. These findings indicate that the self-developed instrument provides a promising tool for robust fruit quality assessment.