<![CDATA[Nonlinear prediction faces its main challenge in the form of overfitting, which leads to inaccurate predictions. The problem becomes evident when researchers attempt to apply advanced deep learning systems to small agricultural data collections. Researchers use knowledge discovery in databases (KDD) to evaluate regularization methods with ensemble techniques, but have not sufficiently explored how structured data augmentation interacts with MaxNorm regularization. The research explores how a sliding-window transformation, together with bootstrap augmentation methods, works when ElasticNet, Bayesian, and MaxNorm regularization techniques are combined into an LSTM-XGBoost prediction system to predict Tikog grass demand. The research showed that data augmentation techniques helped reduce model overfitting, thereby improving performance on prediction tasks. Among the regularization strategies applied to LSTM, MaxNorm achieved the largest reduction in error, with testing MSE decreasing from 0.060472 to 0.002090 after augmentation. A comparative evaluation further shows that LSTM-XGBoost achieved the highest overall performance (R² = 0.997806), while deep learning models showed greater sensitivity to augmentation and regularization strategies. These findings highlight that structured time-series augmentation combined with norm-based regularization enhances generalization capability, particularly for high-capacity sequence models trained on limited agricultural data.]]>
