Autonomous spacecraft navigation in cislunar space requires filtering algorithms that remain accurate and robust under nonlinear CRTBP dynamics, realistic measurement noise, and intermittent disturbances. Classical Cubature Kalman Filters (CKF) provide efficient nonlinear estimation but are sensitive to bias, impulsive solar radiation pressure (SRP) events, and burst-like measurement anomalies. Heuristic adaptive variants such as ARCKF improve short-horizon performance but rely solely on instantaneous innovation magnitude, often leading to over‑correction and degraded behavior under complex disturbance patterns. This work introduces a TinyML‑enhanced adaptive filtering framework, ML‑ARCKF‑MLP, which augments ARCKF with a lightweight multilayer perceptron operating entirely on‑board. The MLP processes a 32‑step residual window to infer adaptive noise scaling, enabling temporally aware covariance adjustment with negligible computational overhead. Monte Carlo evaluations across nominal, bias, SRP‑spike, ISR‑burst, and combined disturbance scenarios show that ML‑ARCKF‑MLP consistently improves RMS and final position accuracy relative to CKF and ARCKF. In particular, the learned residual encoding provides strong robustness to bursty and compound disturbances while maintaining runtime comparable to classical filters. Results demonstrate that TinyML‑based noise adaptation is a practical and effective mechanism for enhancing on‑board cislunar spacecraft navigation under realistic environmental and measurement degradation.
Short-term forecasting of coastal environmental conditions is critical for supporting oyster-reef management, early detection of stress events, and informed decision-making in dynamic estuarine systems. While deep learning models have gained traction in environmental prediction, lightweight machine learning approaches remain compelling due to their low computational cost, robustness to noisy sensor data, and interpretability. This paper presents a systematic baseline study for multivariate forecasting using a 48-hour input window and a 24-hour prediction horizon across seven key oyster-reef environmental variables. Random Forest (RF) is evaluated as the primary model and compared against four widely used benchmarks—ARIMA, Gradient Boosting (GB), LSTM, and GRU—under identical windowing, time-ordered splits, and per-feature evaluation. Results show that RF delivers strong and stable performance across most variables, outperforming ARIMA and matching or exceeding GB and recurrent neural networks. The study provides a reproducible, low-complexity baseline that supports future research in coastal forecasting, environmental risk assessment, and real-time autonomous monitoring systems.