train_data.zip for training and validating the model
test_data.csv for which you will need to make predictions on
To learn more about how to format your submission, please visit the submission rules section.
Differences Between Training and Testing Data
Each dataset is made of several unique events (close encounters betwen two objects) which are indexed by a unique number in the event_id column.
The training set has 162634 rows and 13154 unique events (giving on average about 12 rows/CDMs per close encounter).
The testing set has 24484 rows and 2167 unique events (giving on average about 11 rows/CDMs per close encounter).
Important: Note that the testing set and the training set have not been randomly sampled from the database. In other words, while they come from the same database, with the same collection process and the same features, they have been hand picked in order to over-represent high risk events and to create an interesting predictive model. This is a characterstic of this competition where high risk events are scarce, but represent the true final target of a useful predictive model.
In particular, the testing data differs in two major ways compared to the training set:
It only contains events for which the latest CDM is within 1 day ( time_to_tca < 1) of the time to closest approach (TCA). This is because, in some cases, the latest available CDM is days away from the (known) time to closest approach. It would be wrong to assume that the computed risk 7 days before the actual time to closest approach can be a good approximation to the risk at TCA. Furthermore, predicting the risk many days prior the time to closest approach is not of great interest to us. On the other hand, the training set is unfiltered and you will find many cases where the latest available CDMs is days away from the TCA. We have chosen to keep these collision events in the training set because they may still be useful when it comes to predicting events from the test set.
There are no CDMs to learn from which are within 2 days of the TCA. In other words, the data available closest to the TCA will be at least 2 days away. This is because, as mentioned in the challenge section, a potential avoidance manoeuvre is planned at least 2 days prior to closest approach. Similarly to the above, the training set will contain all cases, including events where no data is available at least 2 days prior to closest approach (i.e. events with all their CDMs being within 2 days of TCA are still present in the dataset).
The dataset is represented as a table, where each row correspond to a single CDM, and each CDM contains 103 recorded characteristics/features. There are thus 103 columns, which we describe below. The dataset is made of several unique collision/close approach events, which are identified in the event_id column. In turn, each collision event is made of several CDMs recorded over time. Therefore, a single collision event can be thought of as a times series of CDMs. From these CDMs, for every collision event, we are interested in predicting the final risk which is computed in the last CDM of the time series (i.e. the risk value in the last row of each collision event).
For the column description, we first describe columns which have unique names and then the columns whose name difference only depends on whether they are referring to the target object (if the column name starts with a t) or the chaser object (if the column name starts with a c). Here, target refers to the ESA satellites while chaser refers to the space debris/object we want to avoid. describe the column names shared for both the chaser and the target, we replace t and c with the placeholder x. For instance, c_sigma_r and t_sigma_r both correspond to the description of x_sigma_r.
Note that all the columns are numerical except for c_object_type.
Uniquely Named Columns
risk:self-computed value at the epoch of each CDM [base 10 log]. In the test set, this value is to be predicted, at the time of closest approach for each event_id. Note that, as mentioned above, in the test set, we do not know the actual data contained in CDMs that are within 2 days to closest approach, since they happen in the "future".
event_id: unique id per collision event
time_to_tca: Time interval between CDM creation and time-of-closest approach [days]
mission_id: identifier of mission that will be affected
max_risk_estimate: maximum collision probability obtained by scaling combined covariance
max_risk_scaling: scaling factor used to compute maximum collision probability
miss_distance: relative position between chaser & target at tca [m]
relative_speed: relative speed between chaser & target at tca [m/s]
relative_position_n: relative position between chaser & target: normal (cross-track) [m]
relative_position_r: relative position between chaser & target: radial [m]
relative_position_t: relative position between chaser & target: transverse (along-track) [m]
relative_velocity_n: relative velocity between chaser & target: normal (cross-track) [m/s]
relative_velocity_r: relative velocity between chaser & target: radial [m/s]
relative_velocity_t: relative velocity between chaser & target: transverse (along-track) [m/s]
c_object_type: object type which is at collision risk with satellite
geocentric_latitude: Latitude of conjunction point [deg]