MoistureVision: Rapid non-destructive prediction of cotton seed moisture using hyperspectral imaging and machine...

Rapid, non-destructive and accurate prediction of cotton seed moisture content is important for assessing seed vigour and improving storage capacity. In this study, a prediction approach for cotton seed moisture content was developed based on machine learning (ML) and hyperspectral imaging. Using the cultivar Jinken 1161 as the experimental material, spectral data in the range of 935-1720 nm were acquired. Outliers were removed using the Isolation Forest algorithm, and the samples were divided into calibration and prediction sets using the spectral-physicochemical value coordinate algorithm. The raw spectra were pre-processed using four methods, including Savitzky-Golay (SG) smoothing and standard normal variate transformation, before constructing traditional ML models [partial least square regression and multiple linear regression (MLR)] and deep learning (DL) models [convolutional neural network and long short-term memory network]. To reduce data redundancy and improve computational efficiency, feature wavelengths related to moisture content were selected using the successive projection algorithm and the least absolute shrinkage and selection operator (LASSO). Comparative analysis of different algorithmic combinations identified the optimal model, which was subsequently applied to hyperspectral images for pixel-wise prediction. This application enabled visualisation of the spatial distribution of moisture within individual cotton seeds. The results showed that, given the current sample size, ML models outperformed DL models. The SG-LASSO-MLR model achieved the best performance, with a prediction correlation coefficient (R2 p), root mean square error of prediction and residual predictive deviation of 0.9557, 0.908 and 4.77, respectively. These findings provide a feasible and effective technical solution for rapid and non-destructive detection of cotton seed moisture content. These outcomes offer valuable insights for seed quality evaluation and intelligent crop monitoring.