Deep learning time-series prediction method for cotton nitrogen content based on small sample hyperspectral data
Shizhe Qin, Hong Ren, Shun Chen, Yiren Ding, Hang Li, Xin Lv, Ze Zhang, Lifu Zhang
Abstract
Nitrogen, an essential nutrient for cotton growth, requires accurate and timely assessment for effective fertilizer application. Despite advances in sensing technology and models, traditional machine learning monitoring models exhibit limited accuracy in assessing nitrogen content. Agronomic sample collection is challenging, and small datasets are unsuitable for conventional deep learning (DL) methods. Moreover, monitoring at specific time points cannot capture dynamic nitrogen changes within the crop, and time lags in decision-making can lead to mismatches between nitrogen supply and crop demand. Therefore, improving monitoring accuracy with small agronomic samples and effectively predicting future nitrogen content changes are crucial. Here, we used a hyperspectral technology to collect data, focusing on “Xinluzao53” cotton and established four nitrogen concentration gradients. Approximately 30 days after emergence, we conducted destructive and non-destructive sampling of the main stem leaves at regular intervals. To train the monitoring model, destructive sampling involved collecting hyperspectral data, followed by leaf cutting for nitrogen determination, whereas non-destructive sampling involved collecting hyperspectral data over time without leaf damage. We then constructed DL monitoring models suitable for small samples to estimate nitrogen levels. The optimal monitoring model was applied to non-destructive sampling, and the resulting nitrogen content time-series was cleaned and used as input for prediction models. The one-dimensional convolutional neural network monitoring model developed in this study achieved optimal accuracy, and the improved ensemble time-series prediction models demonstrated better predictive performance than single time-series models. These findings offer valuable insights for monitoring phenotypic parameters with limited sample sizes and predicting future changes.