Cage and raft aquaculture (CRA) is vital for the coastal economy and provides high-quality aquatic products. Accurately monitoring large-scale CRA lays the foundation for predicting CRA product yield and mitigating environmental impacts. This study, focusing on the challenges of detecting large-scale. CRA from freely downloaded, multispectral remote sensing imagery due to the complexity of both CRA and marine environment, proposed an attention-fused deep learning model for accurately retrieving large-scale CRA in China's offshore sea using open-source Sentinel-2 (S2) satellite data. We first downloaded the cloud-free preprocessed S2 images in selected study areas. Manual labeling of cage, raft, and background areas was performed using high-resolution remote sensing images, with labeled images clipped into 32×32 patches. To enhance the perception ability of the feature, the convolutional block attention module was integrated into the well-performing UNet++ by incorporating both channel and spatial attention in each convolutional block of the encoder as well as the Level 1 convolutional blocks of the decoder. Using the sample dataset in 2021, the proposed AF-UNet++ was trained, compared to four mainstream convolutional neural networks, and then adopted to map CRA in both 2021 and 2018 in the study areas, as well as four additional sites. Experimental results demonstrate: 1) our model has the highest OA, F1, and m intersection over union (IoU), with IoU for cage 4.15% higher than other models. 2) Visual comparison illustrates that AF-UNet++ best excels in extracting CRA. 3) Extraction results both in 2021 and 2018 confirm the proposed model can effectively monitor large-scale CRA and has the spatio-temporal stability.
Information on the plastic-mulched farmland (PMF) spatial distribution is essential for supporting precision agriculture and assessing its impact on climate and environment. However, accurate mapping of PMF remains challenging due to its small size relative to satellite data, variable shapes, seasonal dynamics, and short lifespan of visible PMF. In this paper, we propose a method called Spatially Aware - Classifying Raw Irregular Time series (SA-CRIT) which maximizes the use of both temporal and spatial information for 2.5-m PMF mapping from Sentinel-2 time series without compositing. The proposed method extends the single-pixel, Transformer-based CRIT model to incorporate spatial patch modeling for patch-based irregular time series classification and integrates 10-m to 2.5-m super-resolution (SR) reconstruction to better address the challenges posed by small field sizes. Our method was applied to the globally representative PMF dataset, constructed across seven selected sites, with label data generated using Sentinel-2 time series with the aid of high-resolution PlanetScope imagery. SA-CRIT can achieve an overall accuracy (OA) of 92.16% and a mean intersection over union (mIoU) of 78.95% for 2.5-m resolution classifications across the seven study sites, outperforming its baseline CRIT model by 3.01% in OA and 4.23% in mIoU. Compared to the results from SA-CRIT when applied to 10-m Sentinel-2 time series, the 2.5-m SA-CRIT enhanced OA and mIoU by 1.40% and 2.73% respectively, with larger gains (4.12% in OA, 4.95% in mIoU) in boundary regions. A comprehensive spatial transferability assessment of SA-CRIT demonstrates its strong cross-regional adaptability. The dataset and code will be publicly available.