Comparative Evaluation of Deep Learning Models for Environmental Image Classification Using TensorFlow and the Plant Seedlings Dataset

Abstract

This article compares the performance of four deep learning models for the classification of seedling plant images captured under natural conditions. A custom convolutional neural network (CNN) and three pre-trained architectures—MobileNetV2, ResNet50, and EfficientNetB0—were evaluated using the Plant Seedlings Dataset, which contains over 4,700 images across twelve different species. The methodology involved image preprocessing, data augmentation, training in Google Colab, and evaluation through accuracy, macro F1-score, and confusion matrices. Results indicate that transfer learning models significantly outperform the custom CNN in terms of accuracy and generalization. EfficientNetB0 achieved the best overall performance (98.51% accuracy), while MobileNetV2 stood out for its computational efficiency (5 ms per prediction). ResNet50 reached competitive accuracy (98.11%) but required higher training and inference costs. The study confirms that modern architectures enable highly accurate classification even in visually complex scenarios and provides a comparative guide for model selection based on precision and computational constraints. These findings are relevant for real-world applications of artificial intelligence in ecology, environmental monitoring, and precision agriculture.