This project aims to classify land use and cover datasets obtained from Sentinel-2 satellite images using a deep Convolutional Neural Network (CNN) implemented with PyTorch. Two different training approaches are employed: one utilizing transfer learning with the pre-trained VGG19 model, and the other constructing a CNN from scratch.
EuroSAT: Land Use and Land Cover Classification with Sentinel-2
The dataset used in this project consists of Sentinel-2 satellite images labeled with corresponding land use and cover categories. It provides a comprehensive representation of various land features. The dataset comprises 27,000 labeled and geo-referenced images, divided into 10 distinct classes. It is available in two versions: RGB and multi-spectral. The RGB version consists of images encoded in JPEG format, representing the optical Red, Green, and Blue (RGB) frequency bands. These images provide color information in the visible spectrum. The multi-spectral version of the dataset includes all 13 Sentinel-2 bands, which retains the original value range of the Sentinel-2 bands, enabling access to a more comprehensive set of spectral information.
- RGB (The employed one in this project)
- Multi-spectral
For transfer learning, I leveraged the pre-trained VGG19 model, which has been trained on a large-scale image dataset. By fine-tuning the model on our specific land use and cover dataset, I can take advantage of the knowledge and feature extraction capabilities learned by VGG19.
To train the model using transfer learning, follow the steps below:
- Preprocess the dataset by resizing the images, normalizing pixel values, and splitting into training and validation sets.
- Load the pre-trained VGG19 model.
- Freeze the weights of the convolutional layers to prevent further training.
- Replace the classifier layer of VGG19 with a new fully connected layer tailored to the number of land use and cover categories in our dataset.
- Train the model on the training set while monitoring the validation performance.
- Evaluate the trained model on the test set and analyze the classification accuracy and other relevant metrics.
-
Model : The Pretrained VGG19 Model
-
Loss Function: Cross-Entropy
-
Optimizer: Adam
-
Learning Rate: 0.001
-
Batch Normalization: True
-
Apply Dropout: True
-
Number of training epochs: 25
-
Traing Data Size: 22000
-
Test Data Size: 5000
-
Model Summary:
Layer (type:depth-idx) Output Shape Param ========================================================================================== ├─Sequential: 1-1 [-1, 512, 2, 2] -- | └─Conv2d: 2-1 [-1, 64, 64, 64] (1,792) | └─ReLU: 2-2 [-1, 64, 64, 64] -- | └─Conv2d: 2-3 [-1, 64, 64, 64] (36,928) | └─ReLU: 2-4 [-1, 64, 64, 64] -- | └─MaxPool2d: 2-5 [-1, 64, 32, 32] -- | └─Conv2d: 2-6 [-1, 128, 32, 32] (73,856) | └─ReLU: 2-7 [-1, 128, 32, 32] -- | └─Conv2d: 2-8 [-1, 128, 32, 32] (147,584) | └─ReLU: 2-9 [-1, 128, 32, 32] -- | └─MaxPool2d: 2-10 [-1, 128, 16, 16] -- | └─Conv2d: 2-11 [-1, 256, 16, 16] (295,168) | └─ReLU: 2-12 [-1, 256, 16, 16] -- | └─Conv2d: 2-13 [-1, 256, 16, 16] (590,080) | └─ReLU: 2-14 [-1, 256, 16, 16] -- | └─Conv2d: 2-15 [-1, 256, 16, 16] (590,080) | └─ReLU: 2-16 [-1, 256, 16, 16] -- | └─Conv2d: 2-17 [-1, 256, 16, 16] (590,080) | └─ReLU: 2-18 [-1, 256, 16, 16] -- | └─MaxPool2d: 2-19 [-1, 256, 8, 8] -- | └─Conv2d: 2-20 [-1, 512, 8, 8] (1,180,160) | └─ReLU: 2-21 [-1, 512, 8, 8] -- | └─Conv2d: 2-22 [-1, 512, 8, 8] (2,359,808) | └─ReLU: 2-23 [-1, 512, 8, 8] -- | └─Conv2d: 2-24 [-1, 512, 8, 8] (2,359,808) | └─ReLU: 2-25 [-1, 512, 8, 8] -- | └─Conv2d: 2-26 [-1, 512, 8, 8] (2,359,808) | └─ReLU: 2-27 [-1, 512, 8, 8] -- | └─MaxPool2d: 2-28 [-1, 512, 4, 4] -- | └─Conv2d: 2-29 [-1, 512, 4, 4] (2,359,808) | └─ReLU: 2-30 [-1, 512, 4, 4] -- | └─Conv2d: 2-31 [-1, 512, 4, 4] (2,359,808) | └─ReLU: 2-32 [-1, 512, 4, 4] -- | └─Conv2d: 2-33 [-1, 512, 4, 4] (2,359,808) | └─ReLU: 2-34 [-1, 512, 4, 4] -- | └─Conv2d: 2-35 [-1, 512, 4, 4] (2,359,808) | └─ReLU: 2-36 [-1, 512, 4, 4] -- | └─MaxPool2d: 2-37 [-1, 512, 2, 2] -- ├─AdaptiveAvgPool2d: 1-2 [-1, 512, 1, 1] -- ├─Sequential: 1-3 [-1, 10] -- | └─Flatten: 2-38 [-1, 512] -- | └─Linear: 2-39 [-1, 128] 65,664 | └─ReLU: 2-40 [-1, 128] -- | └─Dropout: 2-41 [-1, 128] -- | └─Linear: 2-42 [-1, 10] 1,290 ========================================================================================== Total params: 20,091,338 Trainable params: 66,954 Non-trainable params: 20,024,384 Total mult-adds (G): 1.61 ========================================================================================== Input size (MB): 0.05 Forward/backward pass size (MB): 9.25 Params size (MB): 76.64 Estimated Total Size (MB): 85.94 ========================================================================================== -
Results
Train Accuracy: 99.54%'
Test Accuracy: 89.88%'
In addition to transfer learning, I also implement a CNN architecture from scratch to classify the land use and cover dataset. This approach allows the model to learn the relevant features directly from the images without relying on pre-trained weights.
To train the CNN from scratch, follow these steps:
- Preprocess the dataset in the same manner as mentioned for transfer learning.
- Design a CNN architecture suitable for the task of land use and cover classification. Consider using convolutional layers, pooling layers, and fully connected layers.
- Initialize the model with random weights.
- Train the model on the training set while monitoring the validation performance.
- Evaluate the trained model on the test set and analyze the classification accuracy and other relevant metrics.
-
Model:
def conv_block(in_channels, out_channels, kernel_size=3, stride=1, padding=1): return nn.Sequential( nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding), nn.BatchNorm2d(out_channels), nn.ReLU(), nn.MaxPool2d(kernel_size=2, stride=1) ) def build_model(): model = nn.Sequential( conv_block(3, 64,3), conv_block(64, 128,3), conv_block(128, 256,3), conv_block(256, 512,3), nn.AdaptiveAvgPool2d((1,1)), nn.Flatten(), nn.Linear(512, 128), nn.ReLU(), nn.Dropout(0.3), nn.Linear(128, 64), nn.ReLU(), nn.Dropout(0.3), nn.Linear(64, 10), ) -
Loss Function: Cross-Entropy
-
Optimizer: Adam
-
Learning Rate: 0.001
-
Batch Normalization: True
-
Apply Dropout: True
-
Number of training epochs: 25 and 50
-
Traing Data Size: 22000
-
Test Data Size: 5000
-
Model Summary:
========================================================================================== Layer (type:depth-idx) Output Shape Param # ========================================================================================== ├─Sequential: 1-1 [-1, 64, 63, 63] -- | └─Conv2d: 2-1 [-1, 64, 64, 64] 1,792 | └─BatchNorm2d: 2-2 [-1, 64, 64, 64] 128 | └─ReLU: 2-3 [-1, 64, 64, 64] -- | └─MaxPool2d: 2-4 [-1, 64, 63, 63] -- ├─Sequential: 1-2 [-1, 128, 62, 62] -- | └─Conv2d: 2-5 [-1, 128, 63, 63] 73,856 | └─BatchNorm2d: 2-6 [-1, 128, 63, 63] 256 | └─ReLU: 2-7 [-1, 128, 63, 63] -- | └─MaxPool2d: 2-8 [-1, 128, 62, 62] -- ├─Sequential: 1-3 [-1, 256, 61, 61] -- | └─Conv2d: 2-9 [-1, 256, 62, 62] 295,168 | └─BatchNorm2d: 2-10 [-1, 256, 62, 62] 512 | └─ReLU: 2-11 [-1, 256, 62, 62] -- | └─MaxPool2d: 2-12 [-1, 256, 61, 61] -- ├─Sequential: 1-4 [-1, 512, 60, 60] -- | └─Conv2d: 2-13 [-1, 512, 61, 61] 1,180,160 | └─BatchNorm2d: 2-14 [-1, 512, 61, 61] 1,024 | └─ReLU: 2-15 [-1, 512, 61, 61] -- | └─MaxPool2d: 2-16 [-1, 512, 60, 60] -- ├─AdaptiveAvgPool2d: 1-5 [-1, 512, 1, 1] -- ├─Flatten: 1-6 [-1, 512] -- ├─Linear: 1-7 [-1, 128] 65,664 ├─ReLU: 1-8 [-1, 128] -- ├─Dropout: 1-9 [-1, 128] -- ├─Linear: 1-10 [-1, 64] 8,256 ├─ReLU: 1-11 [-1, 64] -- ├─Dropout: 1-12 [-1, 64] -- ├─Linear: 1-13 [-1, 10] 650 ========================================================================================== Total params: 1,627,466 Trainable params: 1,627,466 Non-trainable params: 0 Total mult-adds (G): 5.82 ========================================================================================== Input size (MB): 0.05 Forward/backward pass size (MB): 55.84 Params size (MB): 6.21 Estimated Total Size (MB): 62.09 ========================================================================================== -
Results (in case of training over 25 epochs)
Achieved Train Accuracy: 93.12%'
Achieved Test Accuracy: 91.99%'
Trained with NVIDIA GeForce GTX 1660 Ti
- Results (in case of training over 50 epochs)
Achieved Train Accuracy: 98.25%
Achieved Test Accuracy: 96.39%'
Trained with NVIDIA GeForce GTX 1660 Ti. It took around three hours, 12 minutes.
Contributions to this project are welcome! If you encounter any issues or have suggestions for improvements, please feel free to open an issue or submit a pull request.
This project is licensed under the MIT License.