Fine-Tuning ResNet50 for Land Use and Land Cover Classification in the Philippines

This study focuses on employing ResNet50 as a base model for classifying LULC from the satellite dataset. Unlike conventional ML methods such as decision trees, RF and SVM, convolutional neural network (CNN) has emerged as a viable solution for image classification, with its automatic feature extraction and effective spatial feature capture. However, the complexity of building models from scratch can lead to overfitting, prompting the adoption of transfer learning. Transfer learning involves retraining a pre-trained model base, significantly enhancing CNN model training accuracy, while reducing the need for a large dataset. The objectives of this study include preprocessing Sentinel-1 synthetic aperture radar (SAR) dataset, fine-tuning the ResNet50 model, and assessing model performance using confusion matrix together with other performance metrics. The model-building pipeline includes ResNet50 input preparation, adding custom layers on top of the base model, compiling the model, and implementing early stopping during training. Hyperparameter tuning is performed using the keras tuner. Results exhibit the efficacy of the method, indicated by high training and testing accuracies and low losses. The final evaluation metrics indicate a well-performing model that satisfies the evaluation metrics. This study successfully applies transfer learning with ResNet50 to Sentinel-1 SAR data for LULC classification, providing insights into effective remote sensing applications.