Preprocessing
\n\nAll images were rescaled to RGB image size and concatenated into 20 channel images.
\n\nANNs
\n\nI used two UNET-like ANNs with 160x160 input and binary cross-entropy objective. The first one (7c) learned to predict simultaneously: Buildings, Misc Manmade structures, Road, Track, Trees, Crops, Waterway + Standing water as one class. The second ANN (2c) learned to predict simultaneously: Vehicle Large, Vehicle Small.
\n\nTraining
\n\nFor ANN 7c I made corresponding autoencoder and trained it on random patches from all images. Then I transferred encoder part weights to ANN 7c, fixed the weights, and trained the net on random rotational patches. Then I unfixed the weights and trained the net for some additional time on random rotational patches. I started train ANN 2c on random rotational patches that contain vehicle with probability ~0.5. Then I trained ANN 2c on random rotational patches for additional time.
\n\nPredicting
\n\nTo construct prediction on entire image I used only the central part of ANN predictions. The coverage was 4x (varied offsets, rotations). To predict Waterway and Standing water I used ANN predictions and CCCI index from awesome kernel by Vladimir Osin. To separate Waterway from Standing water I used not just area, but a bit more complex score which includes shape properties (Waterway more likely elongated).
", "messages": [ { "id": "166122", "postDate": "03/08/2017 14:30:22", "content": "Preprocessing
\n\nAll images were rescaled to RGB image size and concatenated into 20 channel images.
\n\nANNs
\n\nI used two UNET-like ANNs with 160x160 input and binary cross-entropy objective. The first one (7c) learned to predict simultaneously: Buildings, Misc Manmade structures, Road, Track, Trees, Crops, Waterway + Standing water as one class. The second ANN (2c) learned to predict simultaneously: Vehicle Large, Vehicle Small.
\n\nTraining
\n\nFor ANN 7c I made corresponding autoencoder and trained it on random patches from all images. Then I transferred encoder part weights to ANN 7c, fixed the weights, and trained the net on random rotational patches. Then I unfixed the weights and trained the net for some additional time on random rotational patches. I started train ANN 2c on random rotational patches that contain vehicle with probability ~0.5. Then I trained ANN 2c on random rotational patches for additional time.
\n\nPredicting
\n\nTo construct prediction on entire image I used only the central part of ANN predictions. The coverage was 4x (varied offsets, rotations). To predict Waterway and Standing water I used ANN predictions and CCCI index from awesome kernel by Vladimir Osin. To separate Waterway from Standing water I used not just area, but a bit more complex score which includes shape properties (Waterway more likely elongated).
", "rawMarkdown": "**Preprocessing**\n\nAll images were rescaled to RGB image size and concatenated into 20 channel images.\n\n**ANNs**\n\nI used two UNET-like ANNs with 160x160 input and binary cross-entropy objective. The first one (7c) learned to predict simultaneously: Buildings, Misc Manmade structures, Road, Track, Trees, Crops, Waterway + Standing water as one class. The second ANN (2c) learned to predict simultaneously: Vehicle Large, Vehicle Small.\n\n**Training**\n\nFor ANN 7c I made corresponding autoencoder and trained it on random patches from all images. Then I transferred encoder part weights to ANN 7c, fixed the weights, and trained the net on random rotational patches. Then I unfixed the weights and trained the net for some additional time on random rotational patches. I started train ANN 2c on random rotational patches that contain vehicle with probability ~0.5. Then I trained ANN 2c on random rotational patches for additional time.\n\n**Predicting**\n\nTo construct prediction on entire image I used only the central part of ANN predictions. The coverage was 4x (varied offsets, rotations). To predict Waterway and Standing water I used ANN predictions and CCCI index from awesome kernel by Vladimir Osin. To separate Waterway from Standing water I used not just area, but a bit more complex score which includes shape properties (Waterway more likely elongated).", "votes": null }, { "id": "166341", "postDate": "03/09/2017 08:45:18", "content": "Pls share the code for learning. Thank you very much.
", "rawMarkdown": "Pls share the code for learning. Thank you very much.", "votes": null }, { "id": "166352", "postDate": "03/09/2017 09:16:22", "content": "can you give example of what layers a UNet-like ANN has?
", "rawMarkdown": "can you give example of what layers a UNet-like ANN has?", "votes": null }, { "id": "732111", "postDate": "01/29/2020 13:59:09", "content": "Can you please explain more clearly how you trained your model.
", "rawMarkdown": "Can you please explain more clearly how you trained your model.", "votes": null }, { "id": "2146872", "postDate": "02/16/2023 08:15:47", "content": "From which Satellaite you are collecting data, if I need the images of a particular area for a custom task how can I access them?
", "rawMarkdown": "From which Satellaite you are collecting data, if I need the images of a particular area for a custom task how can I access them?", "votes": null } ], "comments": [ { "id": 2146872, "author_name": "adityasharma2406", "author_url": "", "post_date": "02/16/2023 08:15:47", "content": "From which Satellaite you are collecting data, if I need the images of a particular area for a custom task how can I access them?
", "votes": null, "replies": [] }, { "id": 166341, "author_name": "samihaq", "author_url": "", "post_date": "03/09/2017 08:45:18", "content": "Pls share the code for learning. Thank you very much.
", "votes": null, "replies": [] }, { "id": 166352, "author_name": "cubbus", "author_url": "", "post_date": "03/09/2017 09:16:22", "content": "can you give example of what layers a UNet-like ANN has?
", "votes": null, "replies": [] }, { "id": 732111, "author_name": "shristiagrawal", "author_url": "", "post_date": "01/29/2020 13:59:09", "content": "Can you please explain more clearly how you trained your model.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "166122": "**Preprocessing**\n\nAll images were rescaled to RGB image size and concatenated into 20 channel images.\n\n**ANNs**\n\nI used two UNET-like ANNs with 160x160 input and binary cross-entropy objective. The first one (7c) learned to predict simultaneously: Buildings, Misc Manmade structures, Road, Track, Trees, Crops, Waterway + Standing water as one class. The second ANN (2c) learned to predict simultaneously: Vehicle Large, Vehicle Small.\n\n**Training**\n\nFor ANN 7c I made corresponding autoencoder and trained it on random patches from all images. Then I transferred encoder part weights to ANN 7c, fixed the weights, and trained the net on random rotational patches. Then I unfixed the weights and trained the net for some additional time on random rotational patches. I started train ANN 2c on random rotational patches that contain vehicle with probability ~0.5. Then I trained ANN 2c on random rotational patches for additional time.\n\n**Predicting**\n\nTo construct prediction on entire image I used only the central part of ANN predictions. The coverage was 4x (varied offsets, rotations). To predict Waterway and Standing water I used ANN predictions and CCCI index from awesome kernel by Vladimir Osin. To separate Waterway from Standing water I used not just area, but a bit more complex score which includes shape properties (Waterway more likely elongated).", "166341": "Pls share the code for learning. Thank you very much.", "166352": "can you give example of what layers a UNet-like ANN has?", "732111": "Can you please explain more clearly how you trained your model.", "2146872": "From which Satellaite you are collecting data, if I need the images of a particular area for a custom task how can I access them?" }, "source": "meta" }