This project is about image classification of satellite imagery, which was a problem presented in the Schneider Electric datascience hackathon. Our team has performed data analysis and visualizations, trained a Deep Convolutional Neural Network called EfficientNetV2S, and achieved competitive results.
The problem was to classify satellite images into one of the three possible classes. The dataset provided by Schneider Electric had a total of 1700 images for training the model. The challenge was to train a model that could accurately classify these images.
We started by exploring the dataset, checking for class imbalances, and performing some visualizations. We found that the dataset was imbalanced similar color disributions which where problems we tackled to boost the performance. We also plotted some sample images for each class to get a better understanding of the data.
We decided to use EfficientNetV2S, a state-of-the-art deep learning architecture that has shown excellent performance on image classification tasks. We trained the model on the dataset using the TensorFlow library, with a batch size of 32 and a learning rate of 0.1. We used the categorical cross-entropy loss function and the Adam optimizer.
We trained the model for 75 epochs and achieved a validation accuracy of 61%, which is competitive compared to other participants in the hackathon.
Our future steps include applying transfer learning using pre-trained models to further improve the model's performance. We plan to use pre-trained models such as ResNet, VGG, and Inception, and fine-tune them on our dataset.
In conclusion, the Schneider Electric datascience hackathon was a great experience for our team. We were able to apply our data analysis, visualization, and deep learning skills to solve a real-world problem. Our model achieved competitive results, and we look forward to applying transfer learning to further improve its performance.