Dissolved organic matter evolution and straw decomposition rate characterization under different water and fertilizer conditions based on three-dimensional fluorescence spectrum and deep learning

Abstract

Straw returning is a sustainable way to utilize agricultural solid waste resources. However, incomplete decomposition of straw will cause harm to crop growth and soil quality. Currently, there is a lack of technology to timely monitor the rate of straw decomposition. Dissolved organic matter (DOM) is the most active organic matter in soil and straw is mainly immersed in the soil in the form of DOM. In order to formulate reasonable straw returning management measures , a timely monitoring method of straw decomposition rate was developed in the study. Three water treatment (60%-65%, 70%-75% and 80%-85% maximum field capacity) and two fertilizer (organic fertilizer and chemical fertilizer) were set up in the management of straw returning to the field. Litterbag method was used to monitor the weight loss rate of straw decomposition under different water and fertilizer conditions in strawberry growth stage. The changes of DOM components were determined by three-dimensional fluorescence spectroscopy (3D-EEM). From the faster decomposition period to the slower decomposition period, the main components of DOM changed from protein-like components to humus-like components. At the end of the experiment, the relative content of humus-like components under the treatment of organic fertilizer and moderate water was the highest. Convolutional neural network (CNN) combined with 3D-EEM was used to identify the decomposition speed of straw. The classification precision of neural network validation set and test are 85.7% and 81.2%, respectively. In order to predict the decomposition rate of straw under different water and fertilizer conditions, 3D-EEM data of DOM were used as the input of CNN, parallel factor analysis (PARAFAC) and fluorescence region integral (FRI), and dissolved organic carbon data were used as the input of dissolved organic carbon linear prediction. The prediction model based on CNN had the best effect (R² = 0.987). The results show that this method can effectively identify the spectral characteristics and predict the decomposition rate of straw under different conditions of water and fertilizer, which is helpful to promote the efficient decomposition of straw.

Keywords: Convolutional neural network; Spectral analytical method; Straw decomposition rate; Water and fertilizer management.