This dataset was used in **Improved Difference the Images for Change Detection Classifiers in SAR Imagery Using Deep Learning** paper. The more in depth description of the dataset is available in the paper. The paper is undergoing peer-review, however a preprint version of the paper is available on arXiv: [https://arxiv.org/abs/2303.17835](https://arxiv.org/abs/2303.17835) # License The dataset is created by combining multiple open data sources. The licenses for the data sources are listed in the acknowledgements section. By using this dataset, you must comply to the licenses of the data sources. # Documentation The dataset uses compressed TFRecords to store the dataset samples. [https://www.tensorflow.org/tutorials/load_data/tfrecord](https://www.tensorflow.org/tutorials/load_data/tfrecord) Example code for parsing and using the dataset can be found from: [https://github.com/janne-alatalo/sar-change-detection](https://github.com/janne-alatalo/sar-change-detection) One sample is constructed from the following features: `image_location`: Geographic centroid point of the sample in Well-known text representation in `EPSG:4326` coordinate system. The string is generated with `ogr.Geometry.ExportToWkt()` method which uses the `POINT(LATITUDE LONGITUDE)` format. `dem_rast`: Digital Elevation map from the location that is projected to same projection window with the SAR images (same 512x512 resulution with same pixel geographical alignment) `forestmask`: Binary mask (`1 == forest` or `0 == non-forest`) that shows the forested areas in the scene. The pixels have same alignment as the SAR images and DEM. This data was not used in the final results in the paper, however during the development phase an experiment was conducted where the loss function weighted forested area pixels more than non-forest pixels. The result was that the custom loss function was not much different from vanilla MSE, thus the result was not reported in the paper. The data is acquired from Finnish Forest Centre: [https://aineistot.metsaan.fi/avoinmetsatieto/Metsamaski/Maakunta/](https://aineistot.metsaan.fi/avoinmetsatieto/Metsamaski/Maakunta/) `target_image`: The "newest" image in the sample (I_t). This is the target in the neural network training. Both bands are in the same image, meaning that the image shape is (2, 512, 512) with band order (VV, VH). `target_image_data_take_id`: The data take id of the `target_image`. Identifies the data take from which the target image is cropped from. `target_image_start_time`: The data take start time. Identifies the data take from which the target image is cropped from and tells when the image is acquired. The timestamp is created with Python standard library `datetime.isoformat()` method. [https://docs.python.org/3/library/datetime.html#datetime.datetime.isoformat](https://docs.python.org/3/library/datetime.html#datetime.datetime.isoformat) `target_image_incidence_angle`: Incidence angle at the centroid point of the target SAR image. `target_image_platform_heading`: Satellite fly direction during the acquisition. `target_image_precipitations`: Rain amounts from the acquisition day and three previous days from the location. `target_image_temperature`: Mean temperature from the acquisition day from the location. `target_image_snow_depth`: Snow depth from the acquisition day at the location. `input_image_stack`: The input images that are stacked in the band dimension (8, 512, 512). ``` I_{t - 4} VV band is input_image_stack[0, :, :] I_{t - 4} VH band is input_image_stack[1, :, :] ... I_{t - 1} VV band is input_image_stack[6, :, :] I_{t - 1} VH band is input_image_stack[7, :, :] ``` All of the `target_image_*` features have matching `input_image_*s` features, except that the `input_image_*` features are arrays with length 4. The input images are ordered from oldest to newest: `input_image_start_time[0] < input_image_start_time[-1]`. Furthermore, the validatation dataset samples also include the statistical simulated changes that are described in the paper. The dataset includes the following additional features: `simulated_change_image`: target_image that includes the simulated statistical change. `simulated_change_mask`: 512 x 512 image that indicates the simulated change pixels with non 0 values. NOTE that the mask value can be anything between 0 and num_simulated_changes. 1 indicates the pixels associated with the first simulated change, 2 with the second and so on... `num_simulated_changes`: Indicates the number of simulated changes in the simulated_change_image. `img_superpixel_map` and `img_superpixel_sets`: Helper features for the simulated change algorithm implementations. The features are a result from that SAR images have different resolutions in different directions, however due to technical decision the SAR images in the training dataset have same resolutions in both width and height directions. Therefore, the images in the dataset have many neighboring pixels with exact same value. These features keep track of these pixels so that the simulated changes can be made to look realistic. # Acknowledgements This dataset contains modified Copernicus Sentinel data 2020-2021. The data is governed by the Legal Notice on the use of Copernicus Sentinel Data and Service Information [https://sentinels.copernicus.eu/documents/247904/690755/Sentinel_Data_Legal_Notice](https://sentinels.copernicus.eu/documents/247904/690755/Sentinel_Data_Legal_Notice) This dataset contains weather data from Finnish Meteorological Institute, licensed under Creative Commons Attribution 4.0 International license (CC BY 4.0) [https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/) This dataset contains geographical data (digital elevation map) from National Land Survey of Finland, licensed under Creative Commons Attribution 4.0 International license (CC BY 4.0) [https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/) This dataset contains geographical data (forest mask) from Finnish Forest Centre, licensed under Creative Commons Attribution 4.0 International license (CC BY 4.0) [https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/)

2023