This research investigates the application of convolutional neural networks (CNNs) for detecting excess grass in playgrounds, with a focus on evaluating the impact of varying network structures on performance. Images captured by security cameras around playgrounds served as the primary dataset, providing realistic scenarios for model training and testing. CNN architectures with one, two, three, four, and five layers were trained to identify areas with excess grass. Performance metrics, including accuracy, precision, recall, F1-score, and AUC, were calculated to evaluate the models using a confusion matrix. The results revealed that deeper models with three or more layers captured intricate features but suffered from overfitting, impacting generalization. The two-layer CNN emerged as the optimal configuration, achieving an accuracy of 99% and an F1-score of 0.99, striking the best balance between computational complexity and prediction accuracy. The single-layer model trained quickly, with the highest accuracy, whereas models with three, four and five-layer model demonstrated diminishing returns in performance. These findings suggest that the one or two-layer architecture is most suitable for automated grass detection tasks, offering insights for future applications in landscape analysis and optimization. This study highlights the importance of selecting appropriate CNN architectures for specific real-world image recognition tasks.