Joint Source-Environment Adaptation of Data-Driven Underwater Acoustic Source Ranging Based on Model Uncertainty

TL;DR

The study addresses the challenge of adapting data-driven underwater acoustic source ranging models to unseen environments without labeled target data. It combines uncertainty-informed test-time adaptation (MUMI and PU) with two algorithms: SHOT, which refines features while keeping the classifier fixed, and JSEA, which jointly leverages source-power information to generate pseudo-labels for uncertain samples. Empirical results on synthetic and real SWellEx-96 data show that JSEA, especially when paired with a CNN classifier, yields substantial gains in MAE and PCL over mismatched baselines and SHOT, highlighting the value of separating source power from environmental effects. The proposed approach offers a practical pathway to robust UWA localization in diverse, noisy, and unknown environments with no additional labeled data from the target domain.

Abstract

Adapting pre-trained deep learning models to new and unknown environments remains a major challenge in underwater acoustic localization. We show that although the performance of pre-trained models suffers from mismatch between the training and test data, they generally exhibit a higher uncertainty in environments where there is more mismatch. Additionally, in the presence of environmental mismatch, spurious peaks can appear in the output of classification-based localization approaches, which inspires us to define and use a method to quantify the "implied uncertainty" based on the number of model output peaks. Leveraging this notion of implied uncertainty, we partition the test samples into sets with more certain and less certain samples, and implement a method to adapt the model to new environments by using the certain samples to improve the labeling for uncertain samples, which helps to adapt the model. Thus, using this efficient method for model uncertainty quantification, we showcase an innovative approach to adapt a pre-trained model to unseen underwater environments at test time. This eliminates the need for labeled data from the target environment or the original training data. This adaptation is enhanced by integrating an independent estimate based on the received signal energy. We validate the approach extensively using real experimental data, as well as synthetic data consisting of model-generated signals with real ocean noise. The results demonstrate significant improvements in model prediction accuracy, underscoring the potential of the method to enhance underwater acoustic localization in diverse, noisy, and unknown environments.

Figures (19)

• Figure 1: The CNN classifier includes a feature extraction part followed by a linear classifier that will be frozen during adaptation.
• Figure 2: An example of a metric-inspired smoothed label $\mathbf{y}_i$ corresponding to $d_i = 2.9$ km or $d^q_i = 20$. Here, we have used the absolute error as the metric, hence, we have used a truncated exponential as the label. The quantization bins are $100$ m wide.
• Figure 3: MUMI uncertainty of a regression dropout network reveals the mismatch between the training and test environments. These are using the SWellEx-96-inspired synthetic data, consisting of a shallow source with a monotone signal at $f=109$ Hz and all of the results are averaged over $100$ independent realizations of an additive white Gaussian noise (AWGN) channel. Top row shows that the uncertainty is different for each []source rangesample and the bottom row shows that, on average, both the uncertainty and the root-mean-squared-error (RMSE) increase with the amount of mismatch.
• Figure 4: Variation of the average peakwise uncertainty with the amount of environmental mismatch in (a) depth and (b) SSP.
• Figure 5: An example of a predicted label in a test environment with a different SSP than the training environment. The quantization bins are $100$ m wide. Observe that the pre-trained model prediction (the label with maximum probability) is far from the ground-truth. However, there is another peak in the output closer to the ground-truth. []After JSEA-based adaptation, the output shows one clear peak at the true label, showing how JSEA corrects and adapts the network to give correct labels in the presence of environmental mismatch.