Semantic segmentation of high-resolution remote sensing images (HRSIs) is a challenging task because objects in HRSIs usually have great scale variance and appearance variance. Although deep convolutional neural networks (DCNNs) have been widely applied in the semantic segmentation of HRSIs, they have inherent limitations in capturing global context. Attention mechanisms and transformer can effectively model long-range dependencies, but they often result in high computational costs when being applied to process HRSIs. In this article, an encoder-decoder network (MSGCNet) is proposed to fully and efficiently model multiscale context and long-range dependencies of HRSIs. Specifically, the multiscale interaction (MSI) module employs an efficient cross-attention to facilitate interaction among multiscale features of the encoder, which bridges the semantic gap between high- and low-level features and introduces more scale information to the network. In order to efficiently model long-range dependencies in both spatial and channel dimensions, the transformer-based decoder block (TBDB) implements window-based efficient multihead self-attention (W-EMSA) and enables interactions cross windows. Furthermore, to further integrate the global context generated by TBDB, the scale-aware fusion (SAF) module is proposed to deeply supervise the decoder, which iteratively fuses hierarchical features through spatial attention. As demonstrated by both quantitative and qualitative experimental results on two publicly available datasets, the proposed MSGCNet exhibits superior performance compared to currently popular methods. The code will be available athttp://github.com/JingxiangZhou/MSGCNet.

Most of the existing object detection methods have complicated hand-designed components, such as non-maximum suppression procedures and manual resizing of anchor boxes. Based on DETR, this paper not only eliminates the need for manual component adjustment, but also solves three problems of poor remote sensing image for directional object capture, slow DETR convergence, and the same attention allocated by different layers of Decoder. First, the D-Angle module is used to align the rotating object region while accelerating the convergence using the a priori angle. Then the overall computation of the model is reduced by using Adaptive Proposal Selection(APS) in the cascade structure. Finally, the Adaptive Query Selection(AQS) module is applied so that Decoder in different layers get different attention weights to optimize the layer-by-layer fine-tuning process. In this paper, the effectiveness of the proposed method is verified using two public datasets, DOTA and HRSC2016.

Over the past few years, various strategies have been proposed to improve the multi-scale information capture capability of networks, such as encoder-decoder framework, convolution layers with different kernel sizes in parallel and multiple branches framework. However, many methods only rely on one of the strategies, which limits their performance when processing remote sensing images with large scale variance. To address this issue and enable the fast and effective extraction of multi-scale semantic information, this manuscript introduces a novel cross-scale feature propagation network (CFPNet). Specifically, the multi-scale convolution (MSC) module aims to capture fine-grained multi-scale context with different receptive fields, and the attention up-sample (AUS) module embeds the semantic information of high-level features into low-level features while maintaining spatial details. Besides, the feature semantic enhancement (FSE) module is proposed to aggregate the multi-layer features of the decoder to enhance the final feature representation. The experimental results on the BLU and GID datasets demonstrate the effectiveness and efficiency of our CFPNet.

Like many other countries, China is still facing severe air pollution issues after extensive efforts. The difficulties in deriving near-surface concentrations from satellite measurements restrict the application of remote sensing of large-scale surface air quality. Aiming at providing daily accurate near-surface ail pollution estimates (PM2.5, PM10, O3, NO2, SO2, and CO), we propose a robust estimation framework called learning air pollutants from satellite observations (LAPSO). The principle of LAPSO is to derive a nonlinear relationship between surface pollutant concentrations of interest and satellite observations with the aid of meteorological reanalyzes based on deep learning techniques. The LAPSO framework is superior to other algorithms due to its robust retrieval performance, independence from chemical transport models (CTMs), lower hardware requirements, and a user-friendly interface. The retrieval results of LAPSO were in good agreement with ground-level measurements according to extensive cross-validation at 1628 sites ($\text{R}^{2}>$0.8 in polluted areas and uncertainty$\ll 5~\mu \text{g}/\text{m}^{3}$for most pollutants) in China. The framework also showed a strong capability to capture the temporal variability of different air pollutants. By comparing with the estimation results from different satellite platforms, TROPOspheric monitoring instrument (TROPOMI) onboard the Sentinel-5P demonstrated marginally better performance for estimating PM2.5. Although the selection of satellite observations did not significantly affect the results of O3 estimation, the number and spatial sampling density of in situ sites imposed large impacts on O3 estimation performance. The success of LAPSO for estimating near-surface concentrations from satellite remote sensing at an enhanced spatiotemporal resolution is expected to serve the continuous and dynamical monitoring of regional and global air pollution.

In contrast to the threats to urban human health, impacts of air pollutants on the ecosystem photosynthesis seem to be less concerned. The existence of aerosols could promote photosynthesis by increasing the ratio of diffuse to direct solar radiation; on the contrary, ozone (O3) could inhibit photosynthesis, as it is detrimental to leaf stomata. However, it is unknown whether these two opposite impacts worldwide cancel each other out. In the current mainstream methods, earth system models may show conflicts within situexperimental results due to their relatively coarse resolution. In virtue of satellite remote sensing and a global eddy covariance (EC) network, we studied ten years of data to explore the impacts of aerosol and O3on photosynthesis by fitting an explainable machine learning model. The impacts of aerosol on gross primary productivity (GPP) were positive in many cases, yet very weak. By means of the nitrogen dioxide (NO2) to formaldehyde (HCHO) ratio, O3was seen with positive impacts on photosynthesis under the NOx-sensitive regime, but the apparent positive impacts correlated with the plant phenology. Under the volatile organic compound (VOC)-sensitive regime, the impacts of O3on GPP were not obvious, which was likely due to the prioritized depletion of O3by NO2and VOCs. The impacts of air pollutants depended on many factors and results varied case by case, but the overall net impacts were negative.

High-levels of surface ozone (O3) pollution threaten human and environmental health. Chongqing, a mountainous municipality located in southwest China, is exposed to serious O3 pollution and requires more studies. Due to its complex terrain and always foggy weather, it is difficult to maintain many in-situ sites in Chongqing, and Chemical Transportation Model (CTM) simulations are also challenged. The recently launched (in 2017) Sentinel-5p satellite provides O3 columns with advanced spatiotemporal resolution. Without the dependence on CTMs, we linked O3 columns and surface monitoring data from 2019 to 2021 in virtue of a deep forest machine-learning model. Compared with another widely used machine-learning model and previous studies, our results showed great advantages in estimating surface O3 on a daily scale. Validated against in-situ sites in Chongqing, averaged R2 of cross-validations reached 0.9 while the root mean squared error (RMSE) and mean bias error (MBE) were 13.57 and 0.37 μg/m3. We found out that the model performance is associated with relative height difference between training sites and the test site. The model performed stably when the height difference was lower than 200 m, but obvious performance degradation was seen when the height difference exceeding 400 m.

Continuously monitoring surface ozone (O3) spatial distribution and forecasting its variations are beneficial to improving air quality and ensuring public health in China, although achieving this goal faces challenges from currently available observations and retrieval techniques. Hence, we introduce a coupled surface O3estimation framework (LESO) to address these challenges by integrating ground-level observing networks and satellite remote sensing. LESO features easy-to-use deep learning algorithms, independence on chemical transportation models (CTMs), and consistent performance using data from different satellites. LESO includes a Deep Forest 21 (DF21) model to interpolate O3concentration by learning spatial patterns and a Long Short-Term Memory (LSTM) model to forecast O3concentration by learning data from the past. We used sites of city-levelin-situnetworks as the control sites to manifest short-distance O3transportation. Satellite-based observations of O3precursor indicators were incorporated to capture O3photochemical reactions. DF21 explained a larger fraction of O3variability (90 %) with a mean bias error of smaller than 1 μg/m3. We also investigated the impact of the number of training sites on the DF21 performance, which suggested that five training sites could ensure a good DF21 performance for the most areas (R2> 0.85 and bias < 2 μg/m3). The forecasted O3concentration via LSTM showed a good and stable agreement (R2≈ 0.85 and bias < 5 μg/m3) with ground-based measurements for 8-hour, 24-hour, 28-hour, and 72-hour time periods, respectively. Overall, LESO aims to bring convenient functionality and reliable surface O3estimates for broad users.

Continuous spatial knowledge is required to control the regional ozone pollution. Measurements from ground-level sites are beneficial to this goal, but their number is limited due to the huge expenses of site establishment, operation, and maintenance. Remote sensing seems a promising data source, but its application is challenged by bad weather conditions. Always covered by thick clouds, Chongqing, a populated industrial city in west China, is facing serious ozone pollution, but relevant studies here are relatively insufficient. Another alternative is estimating ozone by models. Well-performed models degrade in Chongqing partially due to the very complex terrain. Modeled hourly ozone does not agree with ground-level measurements. Therefore, an optimization approach is proposed to improve model estimates for such regions. This approach integrates the ground-level information (e.g., measured ozone and meteorology) through the employment of ResNet (Residual Network). ResNet overcomes the notorious vanishing gradient issue in classic neural networks, and the ability of learning complex systems is largely boosted. Ozone distribution is like a gray image that varies every second, which is not the case usually learned by ResNet. A color-image alike data structure is raised to address this “nonstill image” problem; according to the Taylor Expansion, polynomials can describe a complex system, and the errors are acceptable. To facilitate the usage in business operations, this approach is designed to be robust, inexpensive, and easy to use. The scheme of control site selection is discussed in detail. In cross-validations, this approach performs well, averaged$R^{2}$is higher than 0.9 and the error is less than$5 ~\mu \text {g/m}^{3}$.

The prediction of PM2.5 concentration is a canonical predictive challenge due to the distribution of $PM_{2.5}$ appears serious spatiotemporal variability at multiple scales. Currently, using satellite-based remote sensing data to estimate ground-level PM2.5 is a promising method for providing spatiotemporal continuous information of PM2.5. In this letter, we proposed a deep neural network (DNN)-based PM2.5 prediction model to capture the spatiotemporal variability of ground-level PM2.5 using the remote sensing aerosol optical depth (AOD) data from the Himawari-8 satellite along with the conventional meteorological observation variables (denoted as PM25-DNN). The PM25-DNN model was trained and tested using the data from Beijing-Tianjin-Hebei region of China in 2017, and we compared the prediction performance between the PM25-DNN and the current state-of-the-art methods in this field. The results show that the PM25-DNN outperforms the other models with the cross-validated coefficient of determination ($\text{R}^{2}$ ), root-mean-square error (RMSE), mean prediction error (MPE), and relative prediction error (RPE) were 0.84, $19.9~\mu \text{g}/\text{m}^{3}$ , $11.89~\mu \text{g}/\text{m}^{3}$ , and 41.21%, respectively. Then, the trained PM25-DNN model was applied to estimate the hourly gridded PM2.5 with 1-km spatial resolution. Our results indicate that the DNN architecture can capture the essential spatiotemporal distribution associated with PM2.5 only using AOD data and conventional meteorological observational variables without more handcrafted features. The proposed PM25-DNN model can greatly improve the accuracy of PM2.5 estimation, and it provides a new perspective for PM2.5 monitoring using end-to-end deep learning method.