<p>Soil amendment is significant for maintaining soil health, enhancing crop yield, and promoting sustainable agriculture. In order to improve soil quality for agricultural site preparation, there is a need of advanced technologies to address effective soil amendment procedure. In this context, an efficient deep neural network (DNN) approach is proposed to identify appropriate soil amendments and predict its required amount to maintain soil carbon level in agricultural sites. Unlike conventional single-task learning methods that handle amendment type and quantity prediction separately, the proposed multi-task learning (MTL) framework efficiently learns both tasks in parallel, utilizing shared representations to enhance overall accuracy and reduce prediction error. The proposed DNN model efficiently utilizes a MTL concept to handle the interconnected tasks. This model is developed through hyperparameter tuning and employs a learning rate scheduler to enhance the model performance by ensuring faster convergence, maintaining model smoothness, and reducing the likelihood of being stuck at local minima for the loss function. The proposed model improves a 2.05% accuracy and a 31.25% reduction in root mean square error, instead of existing works. This improvement is attributed to its efficient information-sharing capability for performing both related tasks, such as determining the type of soil amendment and predicting its necessary quantity. Furthermore, machine learning interpretation techniques, such as the feature importance score (F-score) and Sobol index, are employed to identify key factors affecting soil carbon maintenance. The simulation results also indicate additional benefits, including increased organic matter, enhanced fertility, microbial promotion, and stable soil potential of hydrogen levels.</p>