ISS Machine Learning Models

This repository contains machine learning models and tracking tools developed for the International Space Station (ISS), specifically focusing on predicting its geographical position and identifying its hemispheric location.

Model Summary

The Hemisphere Classifier is a supervised classification model implemented using Logistic Regression to determine the spatial quadrant of the International Space Station (ISS). Using the station's instantaneous coordinates (latitude and longitude) as input, the model predicts whether the ISS is currently in the Northern Hemisphere (latitude $\ge 0$, labeled as 1) or the Southern Hemisphere (latitude $< 0$, labeled as 0). Prior to training, raw coordinates are standardized using a standard scaler to ensure equal feature weighting. The model learns a linear decision boundary of the form $z = w_{lon} \cdot \text{longitude_scaled} + w_{lat} \cdot \text{latitude_scaled} + b$, and applies the sigmoid function to determine the probability of a northern position. Highly interpretable, this approach isolates distinct weights for the features and provides an efficient, real-time classifier with options for customizable training-test splits and evaluation metrics (accuracy, confusion matrix, precision, and recall).

To predict continuous spatial trajectories, the project implements a Location Predictor designed to forecast the ISS's latitude and longitude approximately one hour into the future (~60 minutes). Because traditional linear or polynomial regression models fail to capture periodic, wave-like trajectories, this system employs a custom Sinusoidal Regressor fitted using scipy's non-linear least squares curve-fitting library. The regressor maps time-based features (day of the year, hour, and minute) to target coordinates by optimizing parameters for amplitude, frequency, phase shift, and offset. This approach leverages the station's physical orbital dynamics, which follow an approximate 92.9-minute cycle. To prevent coordinate wrap-around issues at the longitudinal ($\pm 180^\circ$) and latitudinal ($\pm 90^\circ$) boundaries, input coordinates are preprocessed into their corresponding sine and cosine values before prediction.

Both models process a unified high-frequency ISS dataset comprising telemetry logged at 10-second intervals. In the Location Predictor, a temporal matching algorithm assigns each entry its future target position within a five-minute tolerance window, using a chronological (non-shuffled) train-test split to prevent temporal data leakage and evaluate the predictions using Mean Absolute Error (MAE). Conversely, the Hemisphere Classifier uses a stratified random split to validate classification metrics. Together, these models establish a robust predictive suite—one identifying discrete geographical hemispheres and the other predicting continuous future coordinates—providing a comprehensive mathematical representation of the station's physical orbit.