The Bayesian inference has a two-step recursion structure, i.e., prediction and updating, which can be viewed as a dynamic reasoning process. Based on this elegant structure, various multi-target tracking (MTT) algorithms have been invented and successfully applied in many areas. On the other hand, Bayesian inference MTT algorithms are model-based methods that rely on models’ accuracy and first-order Markov assumption. In recent years, the MTT algorithms based on deep learning have received much attention due to their model-free property and the ability to learn from data, although they have issues such as over-fitting, generalization, etc. In this work, we propose a Transformer-based multi-target tracker whose architecture mimics the Bayesian inference, referred to as the Bayesian inference-based Transformer (BAIT) for MTT. To deal with the model mismatch issues, BAIT uses neural networks instead of the pre-assumed motion and observation models while retaining the excellent architecture of Bayesian inference. BAIT can recursively complete accurate predictions and updates via Transformer by refining the estimation of target states in a Bayesian inference-like manner. Thus, BAIT can be viewed as a combination of model-based and data-based methods. The simulation results show that, because of combining the advantages of Bayesian architecture with intelligent data association structure, BAIT is competitive in simple scenarios and achieves superior performance when the data association task becomes complicated.

Deep learning (DL) approaches, which do not rely on models and can learn complex relationships within data, garner increasing attention in the model-free multi-target tracking (MTT) domain. However, the study of applying the DL method to multi-sensor fusion-based MTT is relatively less. In this paper, we propose a Transformer-based distributed multi-sensor MTT approach, which adopts a hybrid fusion structure with both feature-level and decision-level fusion. First, for each local sensor, the high-dimensional feature information is extracted from the measurements based on a Transformer-based tracking module, which enables continuous tracking of multiple targets and provides the predicted target states and corresponding uncertainties. Then, the outputs of local sensors are fused using the covariance interception (CI) fusion rule. Finally, to further improve the fusion performance, the decision-level information is fed into a fusion decoder with the feature-level information to obtain the predicted target state and uncertainties after deep fusion. In this way, we realize a deep utilization of different sensors’ information and achieve a feature-level decision-level hybrid multi-sensor fusion, namely, Transformer-based multi-sensor hybrid fusion (TMSHF). Simulation results show that the proposed fusion method outperforms the CI algorithm in various tracking scenarios.

The swarm targets have shown great potential for both military and civilian applications, driving a high demand for reliable trajectory generation and accurate activity recognition. In this paper, we propose a trajectory generation method and establish an end-to-end deep learning model for dynamic continuous activity recognition of swarm targets. First, we devise an activity transition model of the drone swarm based on a continuous-time Markov chain (CTMC). Subsequently, the minimum snap trajectory generation algorithm is employed to generate the trajectories. After that, to recognize the dynamic continuous activity of targets, we develop an end-to-end neural network model to extract spatial and temporal features for swarm targets detected by radar across multiple frames. Finally, we demonstrate the effectiveness and robustness of our proposed method through simulation results.

Although the 5% mission goal for NASA’s standard atmospheric correction (AC) algorithm (i.e., the near-infrared (NIR) algorithm) for oligotrophic oceans has been met, this algorithm applies only to blue bands and is highly sensitive to contamination from cloud straylight and sunglint. Here, we developed an AC algorithm for clear waters based on deep learning (namely, DLAC). The algorithm was trained using 3.6 million pairs of MODIS-Aqua high-quality Rrs from the NIR algorithm and Rayleigh-corrected reflectances selected across the global oceans and from all seasons. Validations usingin situdata and a chlorophyll (Chl) constraint-based approach showed that the uncertainties in the Rrsretrievals for DLAC are lower than those for the NIR algorithm, especially for the green and red bands. More importantly, the DLAC algorithm is more tolerant to cloud adjacency effects and moderate sunglint. As a result, the number of valid observations increased by ~50%, and the coverage of monthly global Level-3 Rrscomposites increased by up to 20%. More spatially and temporally consistent patterns were also found for the Level-3 Rrsand Chl products, and large changes in their magnitudes (up to 20% for Rrsand 30% for Chl) were detected in some oceanic regions. With these improvements in the quality and quantity of data, our DLAC algorithm may be valuable as another option for processing global data.

The operational Land Imager (OLI) aboard Landsat 8 was launched in February 2013 to continue the Landsat's mission of monitoring earth resources at relatively high spatial resolution. Compared to Landsat heritage sensors, OLI has an additional 443-nm band (termed coastal/aerosol (CA) band), which extends Landsat's potential for mapping/monitoring water quality in coastal/inland waters. In addition, OLI's pushbroom design allows for longer integration time and, as a result, higher signal-to-noise ratio (SNR). Using a series of radiative transfer simulations, we provide insights into the radiometric sensitivity of OLI when studying coastal/inland waters. This will address how the changes in water constituents manifest at the sensor level and whether the changes are resolvable (focal plane) relative to OLI's overall noise1.