upload-test-data

README.md

@@ -3,7 +3,8 @@ license: cc-by-4.0
 language:
 - en
 - la
-pretty_name: "Beetles as Sentinel Taxa: Predicting drought conditions from NEON specimen imagery"
+pretty_name: 'Beetles as Sentinel Taxa: Predicting drought conditions from NEON specimen
+  imagery'
 task_categories:
 - image-feature-extraction
 dataset_info:
@@ -40,13 +41,25 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 14289671472.3
+    num_bytes: 14087525685.19
     num_examples: 22370
   - name: validation
-    num_bytes: 1742043556.572
+    num_bytes: 1559423695.07
     num_examples: 2486
-  download_size: 15078112217
-  dataset_size: 16031715028.872
+  - name: seen
+    num_bytes: 2923805033.104
+    num_examples: 7028
+  - name: seen_challenge
+    num_bytes: 3306626373.775
+    num_examples: 6725
+  - name: unseen
+    num_bytes: 1343141646.385
+    num_examples: 2795
+  - name: unseen_challenge
+    num_bytes: 1434282817.54
+    num_examples: 3106
+  download_size: 20021233279
+  dataset_size: 24654805251.064
 tags:
 - biology
 - image
@@ -69,7 +82,20 @@ configs:
     path: data/train-*
   - split: validation
     path: data/validation-*
-description: "Images of pinned carabid beetle specimens collected by the National Ecological Observatory Network (NEON) from ecological sites across the U.S., along with associated metadata and drought severity indices (Standardized Precipitation Evapotranspiration Index (SPEI)). It was developed to for the second HDR ML Challenge, and is intended to support the development of machine learning models to predict environmental conditions—specifically drought status—from organismal traits captured in specimen imagery."
+  - split: seen
+    path: data/seen-*
+  - split: seen_challenge
+    path: data/seen_challenge-*
+  - split: unseen
+    path: data/unseen-*
+  - split: unseen_challenge
+    path: data/unseen_challenge-*
+description: Images of pinned carabid beetle specimens collected by the National Ecological
+  Observatory Network (NEON) from ecological sites across the U.S., along with associated
+  metadata and drought severity indices (Standardized Precipitation Evapotranspiration
+  Index (SPEI)). It was developed to for the second HDR ML Challenge, and is intended
+  to support the development of machine learning models to predict environmental conditions—specifically
+  drought status—from organismal traits captured in specimen imagery.
 ---
 
 
@@ -115,15 +141,19 @@ This dataset contains images of pinned carabid beetle specimens collected by the
     - Wei-Lun Chao, The Ohio State University, Columbus, OH, USA
     - Eric R. Sokol, National Ecological Observatory Network (NEON), Boulder, CO, USA
     - Sydne Record, University of Maine, Orono, ME, USA
-- **Homepage:** https://www.nsfhdr.org/mlchallenge-y2
-- **Repository:** https://github.com/Imageomics/HDR-SMood-Challenge
+- **Homepage:** [nsfhdr.org/mlchallenge-y2](https://www.nsfhdr.org/mlchallenge-y2)
+- **Challenge Repository:** [Imageomics/HDR-SMood-Challenge](https://github.com/Imageomics/HDR-SMood-Challenge)
+- **Sample Repository:** [Imageomics/HDR-SMood-Challenge-sample](https://github.com/Imageomics/HDR-SMood-Challenge-sample)
+- **Data Generation Repository:** [Imageomics/CarabidImaging](https://github.com/Imageomics/CarabidImaging)
 <!-- - **Paper:**  TBD -->
 
 This dataset contains high-resolution images and metadata for pinned specimens of carabid beetles collected across U.S. ecosystems by the [National Ecological Observatory Network (NEON)](https://www.neonscience.org/). Each image is linked to specimen-level metadata (e.g., collection date, site, taxonomic identification) and environmental context, including drought severity indicators ([Standardized Precipitation Evapotranspiration Index (SPEI)](https://spei.csic.es/)) calculated over multiple timescales using remote sensing data. This dataset is designed to support research on the relationship between ecological traits and climate stress, and is intended for training and evaluating machine learning models that predict environmental conditions-particularly drought status—from biological imagery. It was created for the Imageomics portion of the [second HDR ML Challenge](https://www.nsfhdr.org/mlchallenge-y2), which emphasizes model generalization across sites with different climates, land cover types, and beetle species pools. Not all available information is provided as part of the challenge, but it will be added at the end of the challenge to allow for broader use beyond the challenge. This includes smaller beetles, which were entirely excluded from the challenge, but will be added to this dataset later.
 
 ### Supported Tasks and Leaderboards
 
-Leaderboard is available on the [Codabench Challenge page](https://www.codabench.org/competitions/9854#/results-tab).
+Development Phase leaderboard is available on the [Codabench Challenge page](https://www.codabench.org/competitions/9854#/results-tab).
+
+The Final Challenge Leaderboard is hosted on the [HDR SMood Challenge Website](https://www.nsfhdr.org/html/mlchallenge-y2/winners.html).
 
 ## Dataset Structure
 
@@ -141,10 +171,22 @@ Leaderboard is available on the [Codabench Challenge page](https://www.codabench
         ...
         scalebar_<scalebarid k>.png
     data/
+        seen-00000-of-00006.parquet
+        ...
+        seen-00005-of-00006.parquet
+        seen_challenge-00000-of-00007.parquet
+        ...
+        seen_challenge-00006-of-00007.parquet
         train-00000-of-00029.parquet
         train-00001-of-00029.parquet
         ...
         train-00028-of-00029.parquet
+        unseen-00000-of-00003.parquet
+        ...
+        unseen-00002-of-00003.parquet
+        unseen_challenge-00000-of-00003.parquet
+        ...
+        unseen_challenge-00002-of-00003.parquet
         validation-00000-of-00004.parquet
         ...
         validation-00003-of-00004.parquet
@@ -153,7 +195,11 @@ Leaderboard is available on the [Codabench Challenge page](https://www.codabench
         <img_id 2>.png
         ...
         <img_id n>.png
+    seen_domain.csv
+    seen_domain_challenge.csv
     train.csv
+    unseen_domain.csv
+    unseen_domain_challenge.csv
     val.csv
 ```
 
@@ -163,13 +209,23 @@ Each record in `train.csv` or `val.csv` corresponds to a single pinned carabid b
 
 Metadata include an identifier for a collection event (`eventID`), the date of the collection event (`collectDate`), an anonymized identifier of the domain (`domainID`) and site (`siteID`) where the collection event took place, taxonomic information (`scientificName`), unique beetle image identifier (`public_id`), a link to the beetle image file (`relative_img_loc`), a link to the color palette image (`colorpicker_path`), a link to the scale image (`scalebar_path`), and Standardized Precipication Evapotranspiratoin Index (SPEI) values that correspond with the location and time that the specimen was collected. The SPEI values were calculated for the 30 day (`SPEI_30d`), 1 year (`SPEI_1y`), and 2 year (`SPEI_2y`) time windows preceding the time of collection at each location for each beetle specimen in the dataset. 
 
-The `train.csv` and `val.csv` files are to be used for training models for submission and reflect the information that will be given during testing sans `siteID`, `collectDate`, and the target variables `SPEI_30d`, `SPEI_1y`, and `SPEI_2y`. Please see the [challenge sample repository](https://github.com/Imageomics/HDR-SMood-Challenge-sample) for an example of how these were used in training the baseline submission.
+The `train.csv` and `val.csv` files are to be used for training models for submission and reflect the information that will be given during testing sans `siteID`, `collectDate`, and the target variables `SPEI_30d`, `SPEI_1y`, and `SPEI_2y`. Please see the [challenge sample repository](https://github.com/Imageomics/HDR-SMood-Challenge-sample) for an example of how these were used in training the baseline submission. Now that the challenge has closed, the testings files `seen_domain.csv`, `seen_domain_challenge.csv`, `unseen_domain.csv`, `unseen_domain_challenge.csv` have been included with the same columns as `train.csv` and `val.csv`.
+
+The `data/` folder contains the dataset in parquet format, where `train` prefix indicates it corresponds to the images and metadata in `train.csv`, while `validation` corresponds to `val.csv`. Similarly:
+
+`seen` -> `seen_domain.csv`
+
+`unseen` -> `unseen_domain.csv`
+
+`seen_challenge` -> `seen_domain_challenge.csv`
+
+`unseen_challenge` -> `unseen_domain_challenge.csv`
 
-The `data/` folder contains the dataset in parquet format, where `train` prefix indicates it corresponds to the images and metadata in `train.csv`, while `validation` corresponds to `val.csv`. These are rendered by the dataset viewer.
+These are rendered by the dataset viewer.
 
 ### Data Fields
 
-Both `train.csv` and `val.csv` have the following columns.
+`train.csv`, `val.csv`, `seen_domain.csv`, `seen_domain_challenge.csv`, `unseen_domain.csv`, `unseen_domain_challenge.csv` have the following columns.
 | fieldName | description | dataType | relatedTerms |
 |---|---|---|---|
 | eventID | An (anonymized) identifier for the set of information associated with the event, which includes information about the place and time of the event | string | [DWC_v2009-04-24:eventID](http://rs.tdwg.org/dwc/terms/history/index.htm#eventID-2009-04-24)
@@ -201,13 +257,33 @@ All domains were split into two sets: one for out-of-distribution (OOD) testing
 #### Training v. Testing (ID)
 All domains contain up to three sites where collection events of beetles took place. Each collection event is defined by when it took place (`collectDate`) and where (`siteID`). All ID domains were split based on their sites into training and testing (ID). If a domain only contained one site, then all events from that site were placed in training. If a domain contained more than one, then one site was held out for testing (ID) and all others were placed in training.
 
-#### Training v. Validation
-During our preliminary experiments we extract a validation set out of the training set for early stopping in training. This split is what is made publicly available in the `train.csv` and `val.csv` files for training and validation respectively.
+##### Loading Development Phase Data
+Development files for training and validation can be loaded with:
 
-#### Initial phase v. Challenge phase
+```
+training = load_dataset("imageomics/sentinel-beetles", split="train")
+validation = load_dataset("imageomics/sentinel-beetles", split="validation")
+```
 
+#### Initial phase (Private Validation) v. Challenge phase (Final Challenge)
 The testing sets currently contain two subsets: an ID set (images from sites whose domain was seen in training), and an OOD set which has images from domains unique to the test set. Both the ID and OOD sets were split in half for the two phases of our competition. The initial (development) phase of the competition will evaluate all submitted models on one half only. These will be represented by the private files `seen_domain.csv` and `unseen_domain.csv` for ID and OOD, respectively. Around the end of the competition, we will also evaluate all submitted models on the other half, which is contained in the private files `seen_domain_challenge.csv` and `seen_domain_challenge.csv`.
 
+##### Loading Intial Phase Data
+Files used for validation during the intial phase of the challenge for seen domains and unseen domains from those available in the development files can be loaded with: 
+
+```
+seen_initial = load_dataset("imageomics/sentinel-beetles", split="seen")
+unseen_initial = load_dataset("imageomics/sentinel-beetles", split="unseen")
+```
+
+##### Loading Challenge Phase Data
+Files used for final evaluation during the challenge phase for seen domains and unseen domains from those available in the development files can be loaded with:
+
+```
+seen_challenge = load_dataset("imageomics/sentinel-beetles", split="seen_challenge")
+unseen_challenge = load_dataset("imageomics/sentinel-beetles", split="unseen_challenge")
+```
+
 The most metadata will be provided for the training dataset; `siteID`, `collectDate` and the target variables values will be redacted for the challenge dataset. Only images and unredacted metadata will be provided. The challenge will be to recover the target variable values for given collections of beetle images from given site-date combinations. All data will be released at the end of the challenge.
 
 ## Dataset Creation