Modeling Apparent Nitrogen Mineralization under Field Conditions Using Regressions and Artificial Neural Networks

Abstract

Soil N mineralization is an important source of N for grain crops, but its estimation under field conditions is usually very difficult. Our objective was to develop models suitable for predicting N mineralization during the growing seasons of wheat (Triticum aestivum L.) and corn (Zea mays L.) under field conditions. Fifty-eight field experiments were performed with wheat, and 35 with corn, along three growing seasons, in which soil apparent N mineralization was estimated by the mass balance approach. Apparent nitrogen mineralized from decomposing residues (ANMR) or soil humic substances (ANMH) were estimated separately. Two empirical modeling techniques were tested, linear regression and artificial neural networks, using as independent variables or inputs some environmental variables. Both techniques allowed the development of suitable models for N mineralization prediction (R2 > 0.68), but neural networks gave slightly better results. The ANMR ranged from −42 to 64 kg N ha−1, increasing as residue mass and N concentration increased. An average ANMR of 15 to 16 kg N ha−1 was produced both during wheat and corn growing seasons. The ANMH ranged from −80 to 328 kg N ha−1, being on average four times greater during corn growing cycle than during wheat season (127 vs. 34 kg N ha−1). The ANMH decreased as initial mineral N content of the soil, remaining residue mass or fine particles content of the soil increased, and it was greater in soils of higher organic matter level and mineralization potential, as determined by an incubation test. Increases in temperature and rainfall also determine greater ANMH. The methodology developed for apparent N mineralization estimation may be applied to other crops and production regions.