We present the implementation of four FPGA-accelerated convolutional neural network (CNN) models for onboard cloud detection in resource-constrained CubeSat missions, leveraging Xilinx’s Vitis AI (VAI) framework and Deep Learning Processing Unit (DPU), a programmable engine with pre-implemented, parameterizable IP cores optimized for deep neural networks, on a Zynq UltraScale+ MPSoC. This study explores both pixel-wise (Pixel-Net and Kernel-Net) and image-wise (U-Net and Patch-Net) models to benchmark trade-offs in accuracy, latency, and model complexity. Applying channel pruning, we achieved substantial reductions in model parameters (up to 98.6%) and floating-point operations (up to 90.7%) with minimal accuracy loss. Furthermore, the VAI tool was used to quantize the models to 8-bit precision, ensuring optimized hardware performance with negligible impact on accuracy. All models retained high accuracy post-FPGA integration, with a cumulative maximum accuracy drop of only 0.6% after quantization and pruning. The image-wise Patch-Net and U-Net models demonstrated strong real-time inference capabilities, achieving frame rates per second of 57.14 and 37.45, respectively, with power consumption of around 2.5 W, surpassing state-of-the-art onboard cloud detection solutions. Our approach underscores the potential of DPU-based hardware accelerators to expand the processing capabilities of small satellites, enabling efficient and flexible onboard CNN-based applications.

The exacerbation of wildfires, attributed to the effects of climate change, presents substantial risks to ecological systems, infrastructure, and human well-being. In the context of the Sustainable Development Goals (SDGs), particularly those related to climate action, prioritizing the assessment and management of the occurrence and intensity of extensive wildfires is of utmost importance. In recent times, there has been a significant increase in the frequency and severity of widespread wildfires worldwide, affecting several locations, including Australia, Italy, and the United States of America. The presence of complex phenomena marked by limited predictability leads to significant negative impacts on biodiversity and human lives. The utilization of satellite-derived data with neural networks, such as convolutional neural networks (CNNs), is a potentially advantageous approach for augmenting the monitoring capabilities of wildfires. This research examines the generalization capability of four neural network models, namely the fully connected (FC), one-dimensional (1D) CNN, two-dimensional (2D) CNN, and three-dimensional (3D) CNN model. Each model’s performance, as measured by accuracy, recall, and F1 scores, is assessed through K-fold cross-validation. Subsequently, T-statistics and p-values are computed based on these metrics to conduct a statistical comparison among the different models, allowing us to quantify the degree of similarity or dissimilarity between them. By using training data from Australia and Sicily, the performances of the trained model are evaluated on the test dataset from Oregon. The results are promising, with cross-validation on the training dataset producing mean precision, recall, and F1 scores ranging between approximately 0.97 and 0.98. Especially, the fully connected model has superior generalization capabilities, whilst the 3D CNN offers more refined and less distorted classifications. However, certain issues, such as false fire detection and confusion between smoke and shadows, persist. The aforementioned methodologies offer significant perspectives on the capabilities of neural network technologies in supporting the detection and management of wildfires. These approaches address the crucial matter of domain transferability and the associated dependability of predictions in new regions. This study makes a valuable contribution to the ongoing efforts in climate change by assisting in monitoring and managing wildfires.

This paper investigates the optimization properties of hybrid quantum-classical quanvolutional neural networks (QuanvNNs), i.e., hybrid architectures merging quanvolutional layers with classical dense layers. We explore several optimization strategies by assessing 9 optimizers across various configurations of layers and kernel size. Through rigorous experimentation, it was found that optimizers such as Adabelief, RMSProp, and Adam outperform others, showcasing high accuracy scores. Conversely, Fromage, SM3, and Adagrad exhibit notably lower performance, highlighting the difficulty of effectively navigating the loss landscape. These findings emphasize the critical role of optimizer selection in a gradient-based variational optimization strategy. Additionally, quanvolutional layer parameters exhibit minimal fluctuations during the training phase, while classical parameters undergo variations on an order of magnitude higher. This disparity suggests a distinct behavior in the optimization dynamics between quantum and classical components, potentially influencing the efficacy of each layer type and offering valuable insights into the underlying mechanisms governing quanvolutional filters. We also evaluate QuanvNN robustness through comparisons against classical CNNs in scenarios involving few-shot learning and input perturbations. Our experimental findings underscore not only enhanced convergence capabilities of QuanvNNs, but also emphasize their superior efficacy in handling severely perturbed inputs. Finally, a randomized QuanvNN variant is tested, yielding comparable performance to optimized models as suggested by previous experiments on parameters variation. All these findings underscore the potential of QuanvNNs in offering robustness and optimization potential compared to their classical counterparts, with implications for quantum-enhanced machine learning.

The problem of real-time optimal guidance is extremely important for successful autonomous missions. In this paper, the last phases of autonomous lunar landing trajectories are addressed. The proposed guidance is based on the Particle Swarm Optimization, and the differential flatness approach, which is a subclass of the inverse dynamics technique. The trajectory is approximated by polynomials and the control policy is obtained in an analytical closed form solution, where boundary and dynamical constraints are a priori satisfied. Although this procedure leads to sub-optimal solutions, it results in beng fast and thus potentially suitable to be used for real-time purposes. Moreover, the presence of craters on the lunar terrain is considered; therefore, hazard detection and avoidance are also carried out. The proposed guidance is tested by Monte Carlo simulations to evaluate its performances and a robust procedure, made up of safe additional maneuvers, is introduced to counteract optimization failures and achieve soft landing. Finally, the whole procedure is tested through an experimental facility, consisting of a robotic manipulator, equipped with a camera, and a simulated lunar terrain. The results show the efficiency and reliability of the proposed guidance and its possible use for real-time sub-optimal trajectory generation within laboratory applications.

Accurate fuel mapping is crucial for effectively determining wildfire risk and implementing management strategies. The primary challenge in fuel type mapping lies in the need to develop accurate and efficient methods for identifying and categorizing the various combustible materials present in an area, often on a large scale. In response to this need, this paper presents a comprehensive approach that combines remote sensing data and Convolutional Neural Network (CNN) to discriminate between fire behavior fuel models. In particular, a CNN-based classification approach that leverages Sentinel-2 imagery is exploited to accurately classify fuel types into seven preliminary main classes (broadleaf, conifers, shrubs, grass, bare soil, urban areas, and water bodies). To further refine the fuel mapping results, subclasses were generated from the seven principles by using biomass and bioclimatic maps. These additional maps provide complementary information about vegetation density and climatic conditions, respectively. By incorporating this information, we align our fuel type classification with the widely used Standard Scott and Burgan (2005) fuel classification system. The results are highly promising, showcasing excellent CNN training performance with all three metrics—accuracy, recall, and F1 score—achieving an impressive 0.99%. Notably, the network exhibits exceptional accuracy in a test case conducted in the southern region of Sardinia, successfully identifying Burnable classes in previously unseen pixels: broadleaf at 0.99%, conifer at 0.79%, shrub at 0.76%, and grass at 0.84%. The proposed approach presents a valuable tool for enhancing fire management, contributing to more effective wildfire prevention and mitigation efforts. Thus, this tool could be leveraged by fire management agencies, policymakers, and researchers to improve the determination of wildfire risk and management.

This paper introduces an innovative approach to explore the capabilities of Quantum Annealing (QA) for trajectory optimization in dynamic systems. The proposed method involves transforming trajectory optimization problems into equivalent binary optimization problems using Quadratic Unconstrained Binary Optimization (QUBO) representation. The procedure is general and adaptable, making it applicable to a wide range of optimal control problems that entail the satisfaction of dynamic, boundary, and path constraints. Specifically, the trajectory is discretized and approximated using polynomials. In contrast to the conventional approach of determining the polynomial degree (n) solely based on the number of boundary conditions, a specific factor is introduced in our method to augment the polynomial degree. As a result, the ultimate polynomial degree is calculated as a composite of two components: n = l + ($m-1$), where m denotes the count of boundary conditions and l signifies the number of independent variables. By leveraging inverse dynamics, the control required to follow the approximated trajectory can be determined as a linear function of independent variables l. As a result, the optimization function, which is represented by the integral of the square of the control, can be formulated as a QUBO problem and the QA is employed to find the optimal binary solutions.