Prediction of nano, fine, and medium colloidal phosphorus in agricultural soils with machine learning

Soil colloids have been shown to play a critical role in soil phosphorus (P) mobility and transport. However, identifying the potential mechanisms behind colloidal P (Pcoll) release and the key influencing factors remains a blind spot. Herein, a machine learning approach (random forest (RF) coupled with partial dependence plot analyses) was applied to determine the effects of different soil physicochemical parameters on Pcoll content in three colloidal subfractions (i.e., nano- (NC): 1–20 nm, fine- (FC): 20–220 nm and medium-sized colloids (MC): 220–450 nm) based on a regional dataset of 12 farmlands in Zhejiang Province, China. RF successfully predicted Pcoll content (R2 = 0.98). Results showed that colloidal- organic carbon (OCcoll) and minerals were the major determinants of total Pcoll content (1–450 nm); their critical values for increasing Pcoll release were 87.0 mg L−1 for OCcoll, 11.0 mg L−1 for iron (Fecoll) or aluminium (Alcoll), 2.6 mg L−1 for calcium (Cacoll), 9.0 mg L−1 for magnesium (Mgcoll), 2.5 mg L−1 for silicon (Sicoll), and 1.4 mg L−1 for manganese (Mncoll). Among three colloidal subfractions, the major factors determining Pcoll were soil Olsen-P (POlsen; 125.0 mg kg−1), Cacoll (2.5 mg L−1), and colloidal P saturation (21.0%) in NC; Mncoll (1.5 mg L−1), Mgcoll (6.8 mg L−1), and POlsen (135.0 mg kg−1) in FC; while Mncoll (1.5 mg L−1), Alcoll (2.5 mg L−1), and Fecoll (3.8 mg L−1) in MC, respectively. OCcoll had a considerable effect in the three fractions, with critical values of 80.0 mg L−1 in NC or FC, and 50.0 mg L−1 in MC. Our study concluded that the information gleaned using the RF model can be used as crucial evidence to identify the key determinants of different size fractionated Pcoll contents. However, we still need to discover one or more easy-to-measure parameters that can help us better predict Pcoll.