​

Satellite Image Classification

Presented By:

Aarya Parikh - 40262787

Dev Pandya - 40268577

Priyansh Bhuva - 40269498

Harshvardhansingh Rao - 40268567

Introduction

• Satellite image classification harnesses advanced deep learning techniques, particularly Convolutional Neural Networks (CNNs) to analyze imagery.
• Wildfires, icebergs and vast range of data has been collected from datasets.
• CNN architectures are employed to automatically extract relevant features from the imagery including color, shape, texture and spatial patterns which addresses the pressing concerns of environmental degradation.
• Using labeled data and employing optimization techniques alongside adjusting hyperparameters led to enhancements in both accuracy and robustness of the models.
• Concurrently, diligent evaluation and record-keeping of performance metrics were maintained throughout the process.
• In this project, three datasets were used on three architectures along two transfer learning.

​

Datasets Used

• Wildfire
• SeaIce
• RESISC45 (Remote Sensing Image Scene Classification

CNN Architectures

Optimization Algorithm

Loss Function

Results: Experimental Setup

Data Collection

Data Preprocessing

Model Training

Model Optimization

Model Testing

Ablative Study

The following shows the varied hyperparameters on SeaIce Dataset on different models.

Model Comparison on the Datasets

The three images provides the details about the training accuracy vs epochs graphs for the Wildfire, SeaIce and RESISC45 dataset respectively. The bar chart below represents the comparison of all three datasets on each models with their accuracy.

​

SeaIce

RESISC45

Wildfire

Model Comparison on the Datasets

The t-SNE visualization helps us to explore high dimensional data by categorizing similar and dissimilar datapoints into clusters and a class activation map is used to indicate the area of an image that a CNN model analyzes to predict the image class. Examples shown below:

Wildfire

SeaIce

RESISC45

Results: Transfer Learning

Transfer Learning was carried out on two architectures for SeaIce dataset:

​

• MobileNet
• ShuffleNet

Bibliography

1. Deep learning for satellite image classification. in advances in intelligent systems and computing. https://doi. org/10.1007/978-3-319-99010-1_35.
2. Deep learning for understanding satellite imagery: An experimental survey. https://doi.org/10.3389/frai. 2020.534696.
3. Health effects of wildfire. https://www.canada. ca/en/health-canada/news/2023/06/publichealth-update-on-the-health-effects-ofwildfires.html.
4. Learning rate scheduler test. https://github.com/ ozwin / Snake _ Species _ Recognition / blob / main / Output / Practice % 20 % 26 % 20failures / learning_rate_calculation_for_spliting_ epochs_for_cosine.ipynb.
5. Resnet paper. https://arxiv.org/pdf/1512. 03385.pdf.
6. Review of deep learning methods for remote sensing satellite images classification: experimental survey and comparative analysis. https://doi.org/10.3390/ rs10071119. 8 [11] Satellite image classification using deep learning. https: //www.doi.org/10.56726/IRJMETS35775.