The code and data access for the paper - EarthObsNet: A Comprehensive Dataset for Data-Driven Earth Observation Image Synthesis
EarthObsNet contains ready-to-analysis image data for several research activities including:
a) Earth observation image synthesis using physical driving forces and previous Earth observation images and
b) Independent semantic segmentation study and downstream studies of a).
-
The code for those three approaches can be found in code -> mergeInput. In particular, SA blends neighboring input with localized information through spatial aggregation, SS blends through spatial-sequence processing, whereas S0 only works on localized information (no blending).
-
The code for the independent localized approach (same as S0 but with a larger dataset) can be found in code -> local. The reason the comparison of S0, SA, and SS works with a part of the dataset can be found in this previous work.
-
The code for semantic segmentation (both for independent investigation and for downstream exploration of the results of a)) can be found in code -> binary.
The dataset contains 34 SAR images and a few meteorological and surface input layers with the same spatial scope as those images. The data was collected during the 2019 Central US Floods. Data from different sources are pre-processed, stacked, aligned at the pixel-level, and sliced into non-overlapping 256-by-256-pixel patches.
The dataset is organized following the 34 SAR images' IDs (identified by the four-digit string of each zip file). Each ID will have a core_xxxx.zip file, such as core_086E.zip and core_344F.zip. The core file is about 0.8 GB each and contains all the data needed for both activities a) and b) with the default experiment setting as introduced in our previous works. As default, neighboring image patches within 5 hops will be considered and three data options to aggregate those neighboring information - two-step averaging (inside the aggInput_avgavg folder), two-step sum (inside the aggInput folder), and two-sep maximizing (inside the aggInput_maxmax folder) are provided. We highly recommend readers check our previous work for more technical details regarding the neighboring image patch selection and data aggregation.
In addition to the core dataset for each image ID, our dataset allows users to generate custom sub-datasets.
Suppose you want to customize neighboring patches (i.e., less or more than 5-hop) and apply a new data aggregation approach for those neighboring patches. In that case, you can do so using the complete dataset with any specific ID without core_ in the file name. Note that the complete dataset for each image ID is large and may contain 2 to 3 separate zip files. For instance, AC44.z01, AC44.z02, and AC44.zip are different parts of the complete dataset of the AC44 image. Inside each zip file, you will find image patches (inside the slicing folder) and the corresponding DEM files (inside the _slicing_DEM_folder).
CSV files documenting the train-test split for each research activity described above are in the train-test-split folder.
Specifically, the train and test files for the semantic segmentation and the independent localized approach are the train_cor.csv and test_cor.csv, respectively. Here, _cor refers to files in the core folders we discussed in the previous section. The four files starting with robustTest can be used for independent tests to investigate the generalization of models. Image patches and the neighboring information used in S0, SA, and SS are listed in train.csv/trainOutside.csv and test.csv/testOutside.csv, where items in the _Outside.csv contain the neighboring information.
-
Band sequence for image patches (those inside both the slicing folder of the complete dataset and the Tr_p10 folder of the core dataset for each image ID):
Band 1 - target SAR image.
Band 2 - HAND.
Band 3 - land cover.
Band 4 - cumulative precipitation over 24 h (0-1 day's rainfall before the target time).
Band 5 - cumulative precipitation over 25-48 h (1-2 day's rainfall before the target time).
Band 6 - cumulative precipitation over 49-72 h (2-3 day's rainfall before the target time).
Band 7 - cumulative precipitation over 73-120 h (3-5 day's rainfall before the target time).
Band 8 - cumulative precipitation over 121-168 h (5-7 day's rainfall before the target time).
Band 9 - soil moisture obtained during the 3-7 days before the target date.
Band 10 - soil moisture obtained during the first 3 days before the target date.
Band 11 - SAR image captured 12 days before the target date. -
The files inside the labels folder of each core zip file are the flood extent images that can be used as labels for semantic segmentation tests.
-
The files inside the slope folder of each core zip file can be used as independent variables for semantic segmentation tests.
-
The Tr_p10 folder of each core zip file means each image patch inside has more than 10 % of the total area being pre-classified as open water as pre-classified and indicated by the land cover layer. This is because hillslope pixels that are too far away from the river channel or lake area may not change significantly during rainfall and flooding processes.
If needed, users can generate customized sub-datasets with different thresholds for open water areas using the complete dataset for each image ID.
Feel free to send us feedback by filing an issue.
This project is licensed under the MIT License - see the LICENSE file for details.
This project is developed by the University of Iowa Hydroinformatics Lab (UIHI Lab):
https://hydroinformatics.uiowa.edu/
-
Li, Z., & Demir, I. (2024). Better localized predictions with Out-of-Scope information and Explainable AI: One-Shot SAR backscatter nowcast framework with data from neighboring region. ISPRS Journal of Photogrammetry and Remote Sensing, 207, 92–103. https://doi.org/10.1016/J.ISPRSJPRS.2023.11.021
-
Li, Z., Xiang, Z., Demiray, B. Z., Sit, M., & Demir, I. (2023). MA-SARNet: A one-shot nowcasting framework for SAR image prediction with physical driving forces. ISPRS Journal of Photogrammetry and Remote Sensing, 205, 176–190. https://doi.org/10.1016/J.ISPRSJPRS.2023.10.002