The different land uses of the dataset
\n\n| **Class** | **Coverage (km$^2$)** | **Label** |\n|:---------------|:-----------------|:-----------------|\n| **Background** | **125.75** | **0** |\n| **Building** | **1.85** | **1** |\n| **Woodland** | **72.02** | **2** |\n| **Water** | **13.15** | **3** |\n| **Road** | **3.5** | **4** |","metadata":{}},{"cell_type":"markdown","source":"To maintain the readability of this notebook, we have created a utility script called utils.py. This script contains all the necessary functions and classes that we will use throughout this notebook. By utilizing the custom function visualize_tif, we can generate an RGB image with the corresponding masks. We have used the same color map as the creators of this dataset to visualize the 5 different land uses; buildings are represented as dark gray, woodlands as middle gray, water as light gray, roads as white, and black pixels correspond to other land uses (background).","metadata":{}},{"cell_type":"code","source":"\"\"\"from utils import visualize_tif\n\nvisualize_tif(h = 8, w = 8, index = 2,\n save_title = \"VisualizeTif\") # Index = n corresponds to the n-th tif file in order located in \"images\" file\"\"\"","metadata":{"execution":{"iopub.status.busy":"2023-03-18T05:59:03.078025Z","iopub.execute_input":"2023-03-18T05:59:03.078447Z","iopub.status.idle":"2023-03-18T05:59:03.085246Z","shell.execute_reply.started":"2023-03-18T05:59:03.078413Z","shell.execute_reply":"2023-03-18T05:59:03.084121Z"},"trusted":true},"execution_count":3,"outputs":[{"execution_count":3,"output_type":"execute_result","data":{"text/plain":"'from utils import visualize_tif\\n\\nvisualize_tif(h = 8, w = 8, index = 2,\\n save_title = \"VisualizeTif\") # Index = n corresponds to the n-th tif file in order located in \"images\" file'"},"metadata":{}}]},{"cell_type":"markdown","source":"## 1.2 Image Cropping ","metadata":{}},{"cell_type":"markdown","source":"The images in their original size, which are around 9000 x 9000, are too large to be used for training a deep neural network due to the limited availability of GPU resources. As a result, the authors in [[1]](#ref1) have divided the images and their corresponding masks into 512 x 512 tiles and split them into training, validation, and test sets at 70%, 15%, and 15% respectively. They also provide a list of filenames for all the datasets, enabling users to test their models on the same sets and compare their results with the benchmark results.\n\nTo streamline the splitting process, we have created a custom function called split_images. The function takes in only one argument which is the desired size of the resulting squared images (default is set to 512). We have utilized this function to split the images and save the results in a folder called \"output\".","metadata":{}},{"cell_type":"markdown","source":"### 1.2.1 Squared images of size 512 \n\nIn this section, we have divided the images into smaller ones with a size of 512x512. Additionally, we have provided some statistical information regarding the resulting dataset. To visualize the obtained dataset, we have used a custom function called visualize_dataset, and some samples can be seen below.","metadata":{}},{"cell_type":"code","source":"# Splitting the images\nfrom utils import split_images\n\nIMAGE_SIZE = 512\nsplit_images(TARGET_SIZE = IMAGE_SIZE)","metadata":{"execution":{"iopub.status.busy":"2023-03-18T05:59:32.547730Z","iopub.execute_input":"2023-03-18T05:59:32.548756Z","iopub.status.idle":"2023-03-18T06:02:18.696310Z","shell.execute_reply.started":"2023-03-18T05:59:32.548711Z","shell.execute_reply":"2023-03-18T06:02:18.694397Z"},"trusted":true},"execution_count":4,"outputs":[{"name":"stdout","text":"Splitting the images...\n\nProcessed M-33-20-D-c-4-2 1/41\nProcessed M-33-20-D-d-3-3 2/41\nProcessed M-33-32-B-b-4-4 3/41\nProcessed M-33-48-A-c-4-4 4/41\nProcessed M-33-7-A-d-2-3 5/41\nProcessed M-33-7-A-d-3-2 6/41\nProcessed M-34-32-B-a-4-3 7/41\nProcessed M-34-32-B-b-1-3 8/41\nProcessed M-34-5-D-d-4-2 9/41\nProcessed M-34-51-C-b-2-1 10/41\nProcessed M-34-51-C-d-4-1 11/41\nProcessed M-34-55-B-b-4-1 12/41\nProcessed M-34-56-A-b-1-4 13/41\nProcessed M-34-6-A-d-2-2 14/41\nProcessed M-34-65-D-a-4-4 15/41\nProcessed M-34-65-D-c-4-2 16/41\nProcessed M-34-65-D-d-4-1 17/41\nProcessed M-34-68-B-a-1-3 18/41\nProcessed M-34-77-B-c-2-3 19/41\nProcessed N-33-104-A-c-1-1 20/41\nProcessed N-33-119-C-c-3-3 21/41\nProcessed N-33-130-A-d-3-3 22/41\nProcessed N-33-130-A-d-4-4 23/41\nProcessed N-33-139-C-d-2-2 24/41\nProcessed N-33-139-C-d-2-4 25/41\nProcessed N-33-139-D-c-1-3 26/41\nProcessed N-33-60-D-c-4-2 27/41\nProcessed N-33-60-D-d-1-2 28/41\nProcessed N-33-96-D-d-1-1 29/41\nProcessed N-34-106-A-b-3-4 30/41\nProcessed N-34-106-A-c-1-3 31/41\nProcessed N-34-140-A-b-3-2 32/41\nProcessed N-34-140-A-b-4-2 33/41\nProcessed N-34-140-A-d-3-4 34/41\nProcessed N-34-140-A-d-4-2 35/41\nProcessed N-34-61-B-a-1-1 36/41\nProcessed N-34-66-C-c-4-3 37/41\nProcessed N-34-77-A-b-1-4 38/41\nProcessed N-34-94-A-b-2-4 39/41\nProcessed N-34-97-C-b-1-2 40/41\nProcessed N-34-97-D-c-2-4 41/41\nExecution completed in 2.0 minutes and 43.63 seconds.\n","output_type":"stream"}]},{"cell_type":"code","source":"from utils import visualize_dataset\n\nvisualize_dataset(num_samples = 8, w = 12, h = 12, pad = 1.4,\n save_title = \"Visualize_dataset\", indices = [0,1,17,20,29,5,6,7])","metadata":{"execution":{"iopub.status.busy":"2023-03-18T06:04:59.213615Z","iopub.execute_input":"2023-03-18T06:04:59.214037Z","iopub.status.idle":"2023-03-18T06:05:04.370050Z","shell.execute_reply.started":"2023-03-18T06:04:59.214002Z","shell.execute_reply":"2023-03-18T06:05:04.369012Z"},"trusted":true},"execution_count":7,"outputs":[{"output_type":"display_data","data":{"text/plain":"\n\n
![](\"https://github.com/ChrisNick92/Geospatial-Data/blob/main/Lab6/Images/u-net-architecture.png?raw=true\")
\n\n\nThe only difference from the architecture presented in the original paper is that we use BarchNorm in every conv block and we do not change the size of the during the operations of a conv block.","metadata":{}},{"cell_type":"code","source":"from utils import UNet\n\nmodel = UNet(in_channels = 3, out_channels = 5).to(device)","metadata":{"execution":{"iopub.status.busy":"2023-03-18T06:12:13.213733Z","iopub.execute_input":"2023-03-18T06:12:13.214812Z","iopub.status.idle":"2023-03-18T06:12:16.511748Z","shell.execute_reply.started":"2023-03-18T06:12:13.214743Z","shell.execute_reply":"2023-03-18T06:12:16.510718Z"},"trusted":true},"execution_count":18,"outputs":[]},{"cell_type":"code","source":"# Training starts!\ntraining_loop(model, train_dloader, val_dloader, epochs, lr, loss_fn, mod_epochs =1,\n regularization = \"L2\", reg_lambda = 1e-6, early_stopping = True,\n patience = 4, verbose = True, model_title = \"Vanilla UNet\", save = True,\n stopping_criterion = \"loss\")","metadata":{"execution":{"iopub.status.busy":"2023-03-18T06:12:16.513655Z","iopub.execute_input":"2023-03-18T06:12:16.514377Z","iopub.status.idle":"2023-03-18T10:51:22.945050Z","shell.execute_reply.started":"2023-03-18T06:12:16.514325Z","shell.execute_reply":"2023-03-18T10:51:22.943765Z"},"trusted":true},"execution_count":19,"outputs":[{"name":"stdout","text":"Training of Vanilla UNet starts!\nUsing loss as stopping criterion.\n\nEpoch: 1/80 Training Loss: 0.6967 Validation Loss: 0.6628\nValidation loss decreased (inf --> 0.662850). Saving model ...\nEpoch: 2/80 Training Loss: 0.6301 Validation Loss: 0.6004\nValidation loss decreased (0.662850 --> 0.600365). Saving model ...\nEpoch: 3/80 Training Loss: 0.5874 Validation Loss: 0.5526\nValidation loss decreased (0.600365 --> 0.552626). Saving model ...\nEpoch: 4/80 Training Loss: 0.5399 Validation Loss: 0.4870\nValidation loss decreased (0.552626 --> 0.487041). Saving model ...\nEpoch: 5/80 Training Loss: 0.5039 Validation Loss: 0.4484\nValidation loss decreased (0.487041 --> 0.448413). Saving model ...\nEpoch: 6/80 Training Loss: 0.4783 Validation Loss: 0.4522\nEarlyStopping counter: 1 out of 4\nEpoch: 7/80 Training Loss: 0.4495 Validation Loss: 0.4134\nValidation loss decreased (0.448413 --> 0.413400). Saving model ...\nEpoch: 8/80 Training Loss: 0.4331 Validation Loss: 0.3894\nValidation loss decreased (0.413400 --> 0.389407). Saving model ...\nEpoch: 9/80 Training Loss: 0.4244 Validation Loss: 0.3615\nValidation loss decreased (0.389407 --> 0.361521). Saving model ...\nEpoch: 10/80 Training Loss: 0.4110 Validation Loss: 0.3631\nEarlyStopping counter: 1 out of 4\nEpoch: 11/80 Training Loss: 0.3973 Validation Loss: 0.3646\nEarlyStopping counter: 2 out of 4\nEpoch: 12/80 Training Loss: 0.3955 Validation Loss: 0.3527\nValidation loss decreased (0.361521 --> 0.352738). Saving model ...\nEpoch: 13/80 Training Loss: 0.3908 Validation Loss: 0.3358\nValidation loss decreased (0.352738 --> 0.335812). Saving model ...\nEpoch: 14/80 Training Loss: 0.3807 Validation Loss: 0.3233\nValidation loss decreased (0.335812 --> 0.323259). Saving model ...\nEpoch: 15/80 Training Loss: 0.3669 Validation Loss: 0.3281\nEarlyStopping counter: 1 out of 4\nEpoch: 16/80 Training Loss: 0.3622 Validation Loss: 0.3379\nEarlyStopping counter: 2 out of 4\nEpoch: 17/80 Training Loss: 0.3604 Validation Loss: 0.3263\nEarlyStopping counter: 3 out of 4\nEpoch: 18/80 Training Loss: 0.3570 Validation Loss: 0.3203\nValidation loss decreased (0.323259 --> 0.320345). Saving model ...\nEpoch: 19/80 Training Loss: 0.3486 Validation Loss: 0.3102\nValidation loss decreased (0.320345 --> 0.310211). Saving model ...\nEpoch: 20/80 Training Loss: 0.3378 Validation Loss: 0.3180\nEarlyStopping counter: 1 out of 4\nEpoch: 21/80 Training Loss: 0.3408 Validation Loss: 0.3091\nValidation loss decreased (0.310211 --> 0.309116). Saving model ...\nEpoch: 22/80 Training Loss: 0.3387 Validation Loss: 0.3033\nValidation loss decreased (0.309116 --> 0.303265). Saving model ...\nEpoch: 23/80 Training Loss: 0.3398 Validation Loss: 0.3262\nEarlyStopping counter: 1 out of 4\nEpoch: 24/80 Training Loss: 0.3286 Validation Loss: 0.2941\nValidation loss decreased (0.303265 --> 0.294091). Saving model ...\nEpoch: 25/80 Training Loss: 0.3281 Validation Loss: 0.3032\nEarlyStopping counter: 1 out of 4\nEpoch: 26/80 Training Loss: 0.3294 Validation Loss: 0.3060\nEarlyStopping counter: 2 out of 4\nEpoch: 27/80 Training Loss: 0.3234 Validation Loss: 0.2952\nEarlyStopping counter: 3 out of 4\nEpoch: 28/80 Training Loss: 0.3141 Validation Loss: 0.2961\nEarlyStopping counter: 4 out of 4\nEarly stopping\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"
![](\"https://github.com/ChrisNick92/Geospatial-Data/blob/main/Lab6/Images/deeplab.png?raw=true\")
\n\n\nThe atrous conv is the main operation of the model. It uses dilated convolution to capture the contextual information at multiple scales and it is really effective for high resolution images.","metadata":{}},{"cell_type":"code","source":"model = smp.DeepLabV3Plus(encoder_name = \"resnet50\",\n encoder_weights = \"imagenet\",\n decoder_atrous_rates = (12,18,24),\n encoder_output_stride =16,\n classes = 5).to(device)","metadata":{"execution":{"iopub.status.busy":"2023-03-18T12:30:58.133247Z","iopub.execute_input":"2023-03-18T12:30:58.133616Z","iopub.status.idle":"2023-03-18T12:30:58.917049Z","shell.execute_reply.started":"2023-03-18T12:30:58.133583Z","shell.execute_reply":"2023-03-18T12:30:58.916049Z"},"trusted":true},"execution_count":26,"outputs":[]},{"cell_type":"code","source":"# Training starts!\ntraining_loop(model, train_dloader, val_dloader, epochs, lr, loss_fn, mod_epochs =1,\n regularization = \"L2\", reg_lambda = 1e-6, early_stopping = True,\n patience = 5, verbose = True, model_title = \"DeepLabV3+ with Resnet encoder\", save = True,\n stopping_criterion = \"loss\")","metadata":{"execution":{"iopub.status.busy":"2023-03-18T12:30:58.918896Z","iopub.execute_input":"2023-03-18T12:30:58.919337Z","iopub.status.idle":"2023-03-18T14:47:06.129419Z","shell.execute_reply.started":"2023-03-18T12:30:58.919298Z","shell.execute_reply":"2023-03-18T14:47:06.128186Z"},"trusted":true},"execution_count":27,"outputs":[{"name":"stdout","text":"Training of DeepLabV3+ with Resnet encoder starts!\nUsing loss as stopping criterion.\n\nEpoch: 1/80 Training Loss: 0.5107 Validation Loss: 0.3652\nValidation loss decreased (inf --> 0.365216). Saving model ...\nEpoch: 2/80 Training Loss: 0.3606 Validation Loss: 0.3042\nValidation loss decreased (0.365216 --> 0.304198). Saving model ...\nEpoch: 3/80 Training Loss: 0.3239 Validation Loss: 0.2822\nValidation loss decreased (0.304198 --> 0.282246). Saving model ...\nEpoch: 4/80 Training Loss: 0.3000 Validation Loss: 0.2720\nValidation loss decreased (0.282246 --> 0.271957). Saving model ...\nEpoch: 5/80 Training Loss: 0.2788 Validation Loss: 0.2571\nValidation loss decreased (0.271957 --> 0.257109). Saving model ...\nEpoch: 6/80 Training Loss: 0.2715 Validation Loss: 0.2503\nValidation loss decreased (0.257109 --> 0.250283). Saving model ...\nEpoch: 7/80 Training Loss: 0.2560 Validation Loss: 0.2662\nEarlyStopping counter: 1 out of 5\nEpoch: 8/80 Training Loss: 0.2522 Validation Loss: 0.2435\nValidation loss decreased (0.250283 --> 0.243539). Saving model ...\nEpoch: 9/80 Training Loss: 0.2512 Validation Loss: 0.2519\nEarlyStopping counter: 1 out of 5\nEpoch: 10/80 Training Loss: 0.2440 Validation Loss: 0.2437\nEarlyStopping counter: 2 out of 5\nEpoch: 11/80 Training Loss: 0.2378 Validation Loss: 0.2358\nValidation loss decreased (0.243539 --> 0.235804). Saving model ...\nEpoch: 12/80 Training Loss: 0.2271 Validation Loss: 0.2384\nEarlyStopping counter: 1 out of 5\nEpoch: 13/80 Training Loss: 0.2297 Validation Loss: 0.2469\nEarlyStopping counter: 2 out of 5\nEpoch: 14/80 Training Loss: 0.2228 Validation Loss: 0.2415\nEarlyStopping counter: 3 out of 5\nEpoch: 15/80 Training Loss: 0.2190 Validation Loss: 0.2283\nValidation loss decreased (0.235804 --> 0.228275). Saving model ...\nEpoch: 16/80 Training Loss: 0.2181 Validation Loss: 0.2452\nEarlyStopping counter: 1 out of 5\nEpoch: 17/80 Training Loss: 0.2170 Validation Loss: 0.2281\nValidation loss decreased (0.228275 --> 0.228136). Saving model ...\nEpoch: 18/80 Training Loss: 0.2171 Validation Loss: 0.2249\nValidation loss decreased (0.228136 --> 0.224857). Saving model ...\nEpoch: 19/80 Training Loss: 0.2138 Validation Loss: 0.2339\nEarlyStopping counter: 1 out of 5\nEpoch: 20/80 Training Loss: 0.2028 Validation Loss: 0.2362\nEarlyStopping counter: 2 out of 5\nEpoch: 21/80 Training Loss: 0.2047 Validation Loss: 0.2203\nValidation loss decreased (0.224857 --> 0.220251). Saving model ...\nEpoch: 22/80 Training Loss: 0.1999 Validation Loss: 0.2329\nEarlyStopping counter: 1 out of 5\nEpoch: 23/80 Training Loss: 0.2015 Validation Loss: 0.2371\nEarlyStopping counter: 2 out of 5\nEpoch: 24/80 Training Loss: 0.1992 Validation Loss: 0.2180\nValidation loss decreased (0.220251 --> 0.218037). Saving model ...\nEpoch: 25/80 Training Loss: 0.1951 Validation Loss: 0.2207\nEarlyStopping counter: 1 out of 5\nEpoch: 26/80 Training Loss: 0.1947 Validation Loss: 0.2368\nEarlyStopping counter: 2 out of 5\nEpoch: 27/80 Training Loss: 0.2024 Validation Loss: 0.2307\nEarlyStopping counter: 3 out of 5\nEpoch: 28/80 Training Loss: 0.1903 Validation Loss: 0.2235\nEarlyStopping counter: 4 out of 5\nEpoch: 29/80 Training Loss: 0.1936 Validation Loss: 0.2210\nEarlyStopping counter: 5 out of 5\nEarly stopping\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"