Kang Sun

Hyperspectral imagery enables precise land cover characterization via continuous spectral measurements, but severe inter-band redundancy necessitates unsupervised band selection (UBS) as a critical preprocessing step. Existing subspace partitioning-based UBS methods suffer from high computational complexity, tight coupling between partitioning and selection criteria, and a lack of rigorous physical prior support, while the counterintuitive superior performance of simple uniform sampling still lacks theoretical interpretation. To address these gaps, this paper proposes a universal UBS enhancement framework grounded in the inherent Markov property of hyperspectral bands. We first design the spectral Markov index (SMI) to quantitatively characterize the spectral Markov property, and theoretically prove that continuous uniform partitioning is the globally optimal strategy under this prior. The proposed two-stage framework fully decouples subspace partitioning and selection criteria, enabling non-intrusive performance enhancement for conventional UBS methods. Systematic experiments on five benchmark datasets validate that the proposed framework achieves an 86.67% performance improvement probability, and confirm that uniform spectral coverage is the dominant determinant of UBS performance.

Dissolved Oxygen (DO) is an important water quality indicator, reflecting aquatic ecological health and influencing biological and biogeochemical cycles. Remote sensing of DO is exigent yet intricate due to its weak correlation with water optical properties. This study addresses this gap by integrating deep learning (DL) techniques with Chlorophyll-a (Chl-a), temperature, and temporal features. We employed a novel indirect approach to first estimate Chl-a and temperature using in-situ time series water quality data from 94 monitoring stations across the Hong Kong coast, covering five years (2019–2023), Sentinel-2 remote sensing data and normalized cyclic month features using machine learning algorithm eXtreme Gradient Boosting (XGB) and then utilized these predicted Chl-a and temperature values, along with normalized cyclic month features, to estimate DO leveraging DL technique two-dimensional convolutional neural network (2D CNN). XGB model achieved its accuracy in estimation of Chl-a (r = 0.78, RMSE = 3.16 μg/L, MAE = 2.31 μg/L) and temperature (r = 0.96, RMSE = 1.14 °C, MAE = 0.89 °C). 2D CNN yielded a promising DO estimation accuracy (r = 0.78, RMSE = 0.67 mg/L, MAE = 0.52 mg/L). This indirect DO estimation approach achieved 50% greater accuracy than direct methods and also showed 10% improvement in r, 8%–17% reduction in RMSE-MAE values over models trained using Chl-a and temperature only, highlighting the unique potential of XGB-2D CNN approach, incorporating cyclic month features. Inclusion of cyclic month features alongside Chl-a and temperature enabled the model to capture both seasonal variations and long-term trends, significantly enhancing the predictive performance of DO estimations. Consequently, monthly/seasonal mean products of Chl-a, temperature, and DO were generated for spatio-temporal analysis for practical use.

Hyperspectral imagery(HSI) clustering aims to assign pixel-wise data with large amount of spectral bands into different groups, where each group indicates one of land-cover objects existed in HSI. Without available label information in clustering task, the clustering performance heavily depends on the reliability of unsupervised feature learned from HSI. Nevertheless,when HSI data are corrupted with noise,the conventional feature learning methods often failed. To address this problem, in this paper, a dual graph-based robust unsupervised feature extraction framework for HSI is proposed to realize reliable clustering. Firstly, low-rank reconstruction and projected learning are incorporated into the proposed framework to improve the data quality and obtain their robust structures. Then, a novel learning schemes is designed to learn two reliable graphs from the above robust structures respectively. We show that the scheme can reveals the latent similarity relationships while removing the noise influence. Meanwhile, the two reliable graphs are also integrated into a comprehensive graph with consistent constraint. At last, a joint learning framework is proposed, in which the data quality improvement, reliable graphs and consistent graph are learned iteratively to benefit from each other. After that, the normalized cut technique is applied to the learned consistent graph to obtain the final unsupervised feature. Several experiments are conducted on the two public HIS datasets to show advantage of our proposed method against the existing methods.