Operationalized Model Documentation - conrad-blucher-institute/semaphore GitHub Wiki
Model Documentation
Each model deployed on Semaphore comes with unique requirements and specifications tailored to optimize its performance and accuracy in forecasting.
South Bird Island Water Temperature (Cold Stunning Models)
21 models prediction water temperature for South Bird Island at varying leadtimes.
| Model | Prediction Span |
|---|---|
| Bird-Island_Water-Temperature_3hr | 3 hours |
| Bird-Island_Water-Temperature_6hr | 6 hours |
| Bird-Island_Water-Temperature_12hr | 12 hours |
| Bird-Island_Water-Temperature_18hr | 18 hours |
| Bird-Island_Water-Temperature_24hr | 24 hours |
| Bird-Island_Water-Temperature_30hr | 30 hours |
| Bird-Island_Water-Temperature_36hr | 36 hours |
| Bird-Island_Water-Temperature_42hr | 42 hours |
| Bird-Island_Water-Temperature_48hr | 48 hours |
| Bird-Island_Water-Temperature_54hr | 54 hours |
| Bird-Island_Water-Temperature_60hr | 60 hours |
| Bird-Island_Water-Temperature_66hr | 66 hours |
| Bird-Island_Water-Temperature_72hr | 72 hours |
| Bird-Island_Water-Temperature_78hr | 78 hours |
| Bird-Island_Water-Temperature_84hr | 84 hours |
| Bird-Island_Water-Temperature_90hr | 90 hours |
| Bird-Island_Water-Temperature_96hr | 96 hours |
| Bird-Island_Water-Temperature_102hr | 102 hours |
| Bird-Island_Water-Temperature_108hr | 108 hours |
| Bird-Island_Water-Temperature_114hr | 114 hours |
| Bird-Island_Water-Temperature_120hr | 120 hours |
Special Requirements:
- The models are based of NDFD data which has a 30 second delay when called to prevent semaphore from spamming NDFD.
- The models are have a 10 minuet offset to allow time for NDFD data to be updated before being called.
Documentation
Models Purpose and Goal
The purpose of the model is to predict water temperature in the Laguna Madre. The modelers know from past measurements that the water temperature is very homogeneous across the Laguna Madre and so they only need to predict it for a single location. One location for which we can get good measurements and predictions is South Bird Island Basin in Padre Island National Seashore, and so that this the location for which we want to generate predictions.
These water temperature predictions are being made with the primary goal of warning stakeholders of cold stunning events. Cold stunning is a condition in which sea turtles become very weak and inactive from exposure to cold temperatures. It generally occurs when water temperatures fall below 50˚F (10˚C) where sea turtles are present. Cold stunned turtles become lethargic and are eventually unable to swim. By predicting ahead of time when these cold stunning events will occur, it is possible for the modelers stakeholders to prepare for these events and reduce fatalities.
Models Lead times
6h, 12h, 18h, 24h, 36h, 48h, 60h, 72h, 84h, 96h, 120h
Training Data
The models are trained the models with past measurements from 2011 to current day from Lighthouse stations collecting air and water temperature on an hourly basis. The measurements are taken from a South Bird Island station and any missing data is gap filled using other lighthouse stations in the area, starting with another South Bird Island sensor, then Packery Channel, finally using linear interpolation if there are still missing values.
Inputs
All models use the same inputs but over a different time range as described below:
- Air temperature measurement:
- Units: Celsius
- Source: South Bird Island Lighthouse Station (station number)
- Interval: Hourly
- Range: 24 hours of past data (including the now time)
- Location: South Bird Island
- Water temperature measurement:
- Units: Celsius
- Source: South Bird Island Lighthouse Station (station number)
- Interval: Hourly
- Range: 24 hours of past data (including the now time))
- Location: South Bird Island
- Air temperature predictions:
- Units: Celsius
- Source: NDFD
- Interval: Hourly (Might change)
- Range: Leadtime + 6h of predictions (including the now time))
- Location: South Bird Island
Outputs
All models generate a single output value representing the forecasted water temperature in Celsius degrees.
Magnolia
Together, the magnolia and transform models predict water levels 48 hours at Magnolia Beach, Texas.
| Model | Prediction Span | Description |
|---|---|---|
| magnolia_12 | 7 hours | Predicts the water levels for the next 7 hours |
| magnolia_transform_12 | 12 hours | Uses the previous 7 hours predicted by magnolia_12 to forecast the next 5 hours, making a total of 12 hours |
| magnolia_24 | 19 hours | Predicts the water levels for the next 19 hours |
| magnolia_transform_24 | 24 hours | Uses the previous 19 hours predicted by magnolia_24 to forecast the next 5 hours, making a total of 24 hours |
| magnolia_48 | 43 hours | Predicts the water levels for the next 43 hours |
| magnolia_transform_48 | 48 hours | Uses the previous 43 hours predicted by magnolia_48 to forecast the next 5 hours, making a total of 48 hours |
Special Requirements:
The transform model must have at least a 1-minute offset after its corresponding magnolia model. For instance, if magnolia_12 runs at 1:00, magnolia_transform_12 should run at 1:01 to ensure it captures the latest data point generated by magnolia_12.
The transform models retrieve information from Semaphore's database to generate predictions.
Documentation:
05/30/2024: We expanded the range of magnolia_transform_12. Initially, this model was missing two inputs. To address this issue, we consulted with the model creator to determine the best course of action, and it was decided to increase the range of inputs to resolve the issue.
06/07/2024: We have decided to switch the input datum to D2W in preparation for visualization. The initial data that will be compared with the Magnolia predictions now exclusively use D2W datums instead of MLLW.
08/02/2024: Requests from lighthouse were unsuccessful. This affected predictions for Magnolia. On 08/02/2024 we fixed the lighthouse issue by changing the request url to use https instead of http.
09/06/2024: Discovered that the magnolia models trained on data that was alternating in descending and ascending order. Semaphore only sends data in ascending order. The magnolia models have now been retrained and the h5 files have been replaced.
Surge
These models predict the water level surges for the next 72 hours at Aransas Pass, North Jetty, Port Isabel, and Rockport.
| Location | Model | Prediction Span |
|---|---|---|
| Aransas Pass | ar_mlp_surge_12h | 12 hours |
| ar_mlp_surge_24h | 24 hours | |
| ar_mlp_surge_48h | 48 hours | |
| ar_mlp_surge_72h | 72 hours | |
| North Jetty | nj_mlp_surge_12h | 12 hours |
| nj_mlp_surge_24h | 24 hours | |
| nj_mlp_surge_48h | 48 hours | |
| nj_mlp_surge_72h | 72 hours | |
| Port Isabel | pi_mlp_surge_12h | 12 hours |
| pi_mlp_surge_24h | 24 hours | |
| pi_mlp_surge_48h | 48 hours | |
| pi_mlp_surge_72h | 72 hours | |
| Rockport | rp_S2S_surge_12h | 12 hours |
| rp_S2S_surge_24h | 24 hours | |
| rp_S2S_surge_48h | 48 hours | |
| rp_S2S_surge_72h | 72 hours |
Special Requirements:
The 48-hour and 72-hour models require interpolation to function properly on Semaphore
Documentation:
5/30/2024: All 48-hour and 72-hour models will be temporarily deactivated because they require post-processing that is only possible with an upgrade to dspec version 2.0. Once their dspec version is upgraded, they will be reactivated.
Virginia Key
Predicts 96 hours of water level data at Virginia Key, Florida.
| Model | Prediction Span |
|---|---|
| 12hr_VirginiaKey_wl | 12 hours |
| 24hr_VirginiaKey_wl | 24 hours |
| 48hr_VirginiaKey_wl | 48 hours |
| 72hr_VirginiaKey_wl | 72 hours |
| 96hr_VirginiaKey_wl | 96 hours |
Inundation
These models predict inundation at Aransas for each month.
| Month | 12-hour Lead Time | 24-hour Lead Time | 48-hour Lead Time |
|---|---|---|---|
| January | jan_inundation_12h | jan_inundation_24h | jan_inundation_48h |
| February | feb_inundation_12h | feb_inundation_24h | feb_inundation_48h |
| March | mar_inundation_12h | mar_inundation_24h | mar_inundation_48h |
| April | apr_inundation_12h | apr_inundation_24h | apr_inundation_48h |
| May | may_inundation_12h | may_inundation_24h | may_inundation_48h |
| June | jun_inundation_12h | jun_inundation_24h | jun_inundation_48h |
| July | jul_inundation_12h | jul_inundation_24h | jul_inundation_48h |
| August | aug_inundation_12h | aug_inundation_24h | aug_inundation_48h |
| September | sep_inundation_12h | sep_inundation_24h | sep_inundation_48h |
| October | oct_inundation_12h | oct_inundation_24h | oct_inundation_48h |
| November | nov_inundation_12h | nov_inundation_24h | nov_inundation_48h |
| December | dec_inundation_12h | dec_inundation_24h | dec_inundation_48h |
MLP-OP
This single model can predict 2 days of water temperature data at South Bird Island.
- MLP-OP
This single model can predict 2 days of water temperature data at South Bird Island.
Documentation:
06/05/2024: Deactivated temporarily due to the temporary deactivation of the Packery Channel station.
06/02/2024: Reactivated as Packery Channel has been reinstated
08/02/2024: Requests from lighthouse were unsuccessful. This affected predictions for MLP-OP. On 08/02/2024 we fixed the lighthouse issue by changing the request url to use https instead of http.
09/20/2024: Updated the dspec version to 2.0. Allowing postprocess calls and interpolation.
Thermal Refuge
This single model predicts the current water temperature in the Laguna Madre canals using historical data from South Bird Island.
Model
- ThermalRefuge
Documentation:
08/02/2024: Requests from lighthouse were unsuccessful. This affected predictions for Thermal Refuge. On 08/02/2024 we fixed the lighthouse issue by changing the request url to use https instead of http.