Development of Artificial Neural Networks for Predicting Soil Micro-Nutrients Availability Under Rice-Based Cropping Systems of North-western India

Micro-nutrient viz., zinc (Zn), copper (Cu), iron (Fe) and manganese (Mn) availability and their transformations have a direct relationship with soil fertility and ecosystem’s productivity. The wide-spread adoption of highly input-intensive rice-based cropping systems (RBCSs) has depleted micro-nutrients, especially in the light-textured soils of north-western India. We quantified micro-nutrients’ pool and investigated their transformations influencing their availability in soils under five different RBCSs, viz. rice-potato-mungbean, rice-peas-maize, rice-potato-maize, rice-wheat and rice-potato-melon and developed and evaluated the accuracy of artificial neural networks (ANNs) in estimating micro-nutrients’ availability in these soils. These results revealed a significant difference (p < 0.05) in micro-nutrients’ pool in soils under different RBCSs with a large variation in micro-nutrients pool, e.g., DTPA-Zn (1.9–3.6 mg kg−1), DTPA-Cu (41.4–47.0 mg kg−1), DTPA-Fe (6.2–7.8 mg kg−1) and DTPA-Mn (0.2–0.5 mg kg−1). A sequential speciation technique elucidated water-soluble + exchangeable fraction as the smallest (~0.1–1.2%), while the residual as the largest (~50%) fraction of the total micro-nutrients’ pool. The oxide-bound Zn, Cu, Fe, and Mn comprised ~44.2%, 42.7, 10.3, and 46.7%, respectively of the total micro-nutrients’ pool in soils under different RBCSs. Total-Zn was significantly higher (by ~33.3%) under rice-potato-maize than the rice-potato-mungbean, while total-Mn was higher (by ~22.8%) in rice-potato-mungbean than the rice-potato-melon system. Conversely, total-Fe content was significantly higher under rice-potato-melon, while total-Cu was significantly lower in rice-potato-melon as compared to the others. The ANN architecture (13-20-2-1 structure) comprised of 13 variables in input layers, 2 hidden layers with 20 and 2 neurons, respectively and micro-nutrient (Zn, Cu, Fe, and Mn) content as output layer with Levenberg-Marquardt algorithm was identified as the best ANN topology. This ANN architecture showed the highest accuracy of micro-nutrients prediction with a coefficient of determination (R2) value of 0.750–0.921**, root mean square error (RMSE) value of 0.057–0.954 mg kg−1 and mean absolute percentage error (MAPE) value of −0.944 and 1.451%. The ANN model architecture produced highly promising results in predicting the micro-nutrients availability in soil under RBCSs. The inclusion of short-duration legume as well as deep-rooted crops helps in enhancing micro-nutrient pool for increased crop production and improved soil health of light textured soils of north-western India.