Cloud Contamination Research Articles

A joint learning Im-BiLSTM model for incomplete time-series Sentinel-2A data imputation and crop classification

Multi-temporal deep learning approaches can make full use of crop growth patterns and phenological characteristics, resulting in excellent crop classification performance in large areas. However, obtaining complete time-series remote sensing images during the growing season is challenging due to cloud contamination. Hence, given time-series multispectral data, it is important to impute missing data and accurately classify crops. A novel Imputation-BiLSTM model (Im-BiLSTM) was developed based on Bidirectional Long Short-term Memory network (BiLSTM) to jointly perform missing data imputation and crop classification. The Im-BiLSTM model regards missing data as network variables, which are efficiently updated during backpropagation. Im-BiLSTM treats the interaction between imputation and classification tasks, reducing the error and uncertainty caused by the separation operation of imputation to classification. Furthermore, we improved the interpretability of the Im-BiLSTM model by evaluating the importance of input features and visualizing hidden state units. In Shawan County, Xinjiang, China, we acquired a total of 10 Sentinel-2A images from April to October 2016, of which 3 images lost partial data due to cloud cover. The Im-BiLSTM model was applied to incomplete time-series data containing 10 time-steps for pixel-level crop classification, and the BiLSTM model was constructed based on cloud-free images for comparison. The performance of the proposed model was tested in four different cases of images missing and missing rates. The results showed that the classification of the Im-BiLSTM model outperformed the BiLSTM model, the overall accuracy was improved by a maximum of 4.2%, and the F1-scores of spring corn and tomato was improved by a maximum of 16.1% and 21.4%, respectively. Therefore, the Im-BiLSTM model can effectively improve classification performance by jointly imputing missing data. The imputation results (the coefficient of determination values range 0.4 ∼ 0.9) indicated that the bands with the larger contribution to the classification had higher imputation accuracy. Feature importance evaluation showed that the Im-BiLSTM model captured the key phenological periods and features from time-series input. The visualization of hidden units demonstrated that the Im-BiLSTM model accumulated useful information over time, and the learned high-level features made the crops more separable than the original inputs.

A novel big data mining framework for reconstructing large-scale daily MAIAC AOD data across China from 2000 to 2020

ABSTRACT Satellite-based aerosol optical depth (AOD) retrieval products are essential in air pollution and climate change research. Unfortunately, cloud contaminations and unfavorable surface conditions result in a considerable proportion of missing AOD data. Numerous studies have been conducted to reconstruct large-scale AOD data gaps by utilizing adjacent spatiotemporal information or modeling AOD data via various external geographical data. However, the erratic variation of AOD and the inconvenience of external data weaken the accuracy and efficiency of reconstruction. To address these issues, a novel big data based iterative variation mining framework (IVMF), utilizing multi-spatiotemporal information on AOD variations, is proposed to reconstruct large-scale AOD data over China from 2000 to 2020. Simulated and real-data experiments are carried out to validate the effectiveness and robustness of the IVMF. Results show that the spatial patterns of the reconstructed AOD are consistent with those of the original AOD in the simulated experiments. The final reconstructed AOD data strongly correlate with in-situ AOD measurements in the real-data experiments (correlation coefficient of 0.91). After reconstruction, the average daily AOD coverage increases from 30.42% to 96.69% (a 218% increase). Besides, results reveal that central China exhibits severe AOD levels, while northwest China presents low AOD levels. Overall, the proposed IVMF can largely resolve the missing AOD data problem with outstanding accuracy and efficiency, and has great potential to be generalized to other regions and remote sensing products.

Development of the GLASS 250-m leaf area index product (version 6) from MODIS data using the bidirectional LSTM deep learning model

Leaf area index (LAI) is a terrestrial essential climate variable that is required in a variety of ecosystem and climate models. The Global LAnd Surface Satellite (GLASS) LAI product has been widely used, but its current version (V5) from Moderate Resolution Imaging Spectroradiometer (MODIS) data has several limitations, such as frequent temporal fluctuation, large data gaps, high dependence on the quality of surface reflectance, and low computational efficiency. To address these issues, this paper presents a deep learning model to generate a new version of the LAI product (V6) at 250-m resolution from MODIS data from 2000 onward. Unlike most existing algorithms that estimate one LAI value at one time for each pixel, this model estimates LAI for 2 years simultaneously. Three widely used LAI products (MODIS C6, GLASS V5, and PROBA-V V1) are used to generate global representative time-series LAI training samples using K-means clustering analysis and least difference criteria. We explore four machine learning models, the general regression neural network (GRNN), long short-term memory (LSTM), gated recurrent unit (GRU), and Bidirectional LSTM (Bi-LSTM), and identify Bi-LSTM as the best model for product generation. This new product is directly validated using 79 high-resolution LAI reference maps from three in situ observation networks. The results show that GLASS V6 LAI achieves higher accuracy, with a root mean square (RMSE) of 0.92 at 250 m and 0.86 at 500 m, while the RMSE is 0.98 for PROBA-V at 300 m, 1.08 for GLASS V5, and 0.95 for MODIS C6 both at 500 m. Spatial and temporal consistency analyses also demonstrate that the GLASS V6 LAI product is more spatiotemporally continuous and has higher quality in terms of presenting more realistic temporal LAI dynamics when the surface reflectance is absent for a long period owing to persistent cloud/aerosol contaminations. The results indicate that the new Bi-LSTM deep learning model runs significantly faster than the GLASS V5 algorithm, avoids the reconstruction of surface reflectance data, and is resistant to the noises (cloud and snow contamination) or missing values contained in surface reflectance than other methods, as the Bi-LSTM can effectively extract information across the entire time series of surface reflectance rather than a single time point. To our knowledge, this is the first global time-series LAI product at the 250-m spatial resolution that is freely available to the public (www.geodata.cn and www.glass.umd.edu).

Long‐Term and Fine‐Scale Surface Urban Heat Island Dynamics Revealed by Landsat Data Since the 1980s: A Comparison of Four Megacities in China

AbstractLong‐term and fine‐scale monitoring of surface urban heat island (SUHI) is critical for the design of heat mitigation strategies. Landsat series offer long‐term (since the 1980s) and fine‐scale land surface temperature (LST) observations for such SUHI analysis. However, Landsat data are characterized by a long revisit period (16 days) and are seriously impacted by cloud contamination and stripe gaps, making the Landsat‐derived long‐term SUHI trends across cities incomparable. To address this issue, here we applied the Prophet model to reconstruct temporally consistent long‐term (∼1985–2019) clear‐sky Landsat LSTs over four megacities in China (Beijing, Chongqing, Shanghai, and Shenzhen). The results show that the mean absolute error of the SUHI intensity (SUHII) estimated using the reconstructed LSTs ranges from 0.2 to 0.6°C. The reliability of the reconstructed LSTs is further evidenced by the general consistency between the reconstructed Landsat and original Moderate Resolution Imaging Spectroradiometer LSTs. Our analysis further demonstrates that the overall SUHII has been increasing slowly over Beijing but has been increasing rapidly in the other three megacities since the 1980s. Local SUHII trends also varied between urban regions. Urban core is characterized by an initial increase and then a slight decrease in the local SUHII, while a continuously increasing trend is observed for the urban fringe. For the new urban development, the interannual SUHII trend lies somewhere between these trends. Overall, our study provides a practical approach to reconstruct long‐term clear‐sky Landsat LSTs, and delivers a better understanding of long‐term and fine‐scale SUHI variations across cities.

Investigation of Potential Prevailing Wind Impact on Land Surface Temperature at Gas Flaring Sites in the Niger Delta, Nigeria.

This research examines the effects of South prevailing wind on Land Surface Temperature (LST) retrieved from Earth Observation (EO) Satellites at 11 gas flaring sites in Rivers State, Niger Delta region, Nigeria. 7 Landsat 5 Thematic Mapper (TM) and 18 Landsat 7 Enhanced Thematic Mapper Plus (ETM+) from 17/01/1986 to 08/03/2013 with < 5 % cloud contamination were considered. All sites are located within a single Landsat scene (Path 188, Row 057). The atmospherically corrected reflectance was used for the classification of 4 land cover (LC) types at each site. The emissivity (ε) for each site is estimated by using standard values for determined LC from Look Up Table (LUT). The surface-leaving radiance (Lλ) is computed from the atmospherically corrected thermal band 6 (High gain) and the emissivity (ε) values. The Planck equation was inverted using Landsat calibration constants to derive LST. Geospatial analysis of LST results using ArcGIS show 6 ranges of LST values for all sites. For both sensors LST retrieved for the flare stack sources are the highest values compared to other locations within the sites. Wind directions and wind speeds for Landsat data acquisitions dates and the South prevailing wind were applied to the LST for assessing their effects on it. The results show that for Eleme I and II, and Onne, the p-values results showed that no statistically significant relationships between δLST values in different directions (δLSTNE, δLSTNS and δLSTNW) existed. For Obigbo site, the wind direction (South) for data acquisition date combined with the South prevailing wind to generate a noticeable impact on the LST towards the North-East and the North-West directions. For Alua, Bonny, Chokocho, Rukpokwu, Umurolu and Sara sites, the p-value obtained is statistically significant for all the 3 (δLSTNE, δLSTNS and δLSTNW) relationships; therefore, producing a circle flare δLST footprint. For Umudioga site, only δLSTN versus δLSTW is statistically significant, causing a noticeable effect on the flare δLST in the North-West direction. Based on these results, it can be concluded that the volume and rate of burning gas, and the speed of the South wind at the time of satellite overpass are major factors that determine the influence of the South prevailing wind on the LST retrieved at the flaring sites in the Niger Delta.

Water Quality Chl-a Inversion Based on Spatio-Temporal Fusion and Convolutional Neural Network

The combination of remote sensing technology and traditional field sampling provides a convenient way to monitor inland water. However, limited by the resolution of remote sensing images and cloud contamination, the current water quality inversion products do not provide both high temporal resolution and high spatial resolution. By using the spatio-temporal fusion (STF) method, high spatial resolution and temporal fusion images were generated with Landsat, Sentinel-2, and GaoFen-2 data. Then, a Chl-a inversion model was designed based on a convolutional neural network (CNN) with the structure of 4-(136-236-340)-1-1. Finally, the results of the Chl-a concentrations were corrected using a pixel correction algorithm. The images generated from STF can maintain the spectral characteristics of the low-resolution images with the R2 between 0.7 and 0.9. The Chl-a inversion results based on the spatio-temporal fused images and CNN were verified with measured data (R2 = 0.803), and then the results were improved (R2 = 0.879) after further combining them with the pixel correction algorithm. The correlation R2 between the Chl-a results of GF2-like and Sentinel-2 were both greater than 0.8. The differences in the spatial distribution of Chl-a concentrations in the BYD lake gradually increased from July to August. Remote sensing water quality inversion based on STF and CNN can effectively achieve high frequency in time and fine resolution in space, which provide a stronger scientific basis for rapid diagnosis of eutrophication in inland lakes.

A global seamless 1 km resolution daily land surface temperature dataset (2003–2020)

Abstract. Land surface temperature (LST) is one of the most important and widely used parameters for studying land surface processes. Moderate Resolution Imaging Spectroradiometer (MODIS) LST products (e.g., MOD11A1 and MYD11A1) can provide this information with moderate spatiotemporal resolution with global coverage. However, the applications of these data are hampered because of missing values caused by factors such as cloud contamination, indicating the necessity to produce a seamless global MODIS-like LST dataset, which is still not available. In this study, we used a spatiotemporal gap-filling framework to generate a seamless global 1 km daily (mid-daytime and mid-nighttime) MODIS-like LST dataset from 2003 to 2020 based on standard MODIS LST products. The method includes two steps: (1) data pre-processing and (2) spatiotemporal fitting. In the data pre-processing, we filtered pixels with low data quality and filled gaps using the observed LST at another three time points of the same day. In the spatiotemporal fitting, first we fitted the temporal trend (overall mean) of observations based on the day of year (independent variable) in each pixel using the smoothing spline function. Then we spatiotemporally interpolated residuals between observations and overall mean values for each day. Finally, we estimated missing values of LST by adding the overall mean and interpolated residuals. The results show that the missing values in the original MODIS LST were effectively and efficiently filled with reduced computational cost, and there is no obvious block effect caused by large areas of missing values, especially near the boundary of tiles, which might exist in other seamless LST datasets. The cross-validation with different missing rates at the global scale indicates that the gap-filled LST data have high accuracies with the average root mean squared error (RMSE) of 1.88 and 1.33∘, respectively, for mid-daytime (13:30) and mid-nighttime (01:30). The seamless global daily (mid-daytime and mid-nighttime) LST dataset at a 1 km spatial resolution is of great use in global studies of urban systems, climate research and modeling, and terrestrial ecosystem studies. The data are available at Iowa State University's DataShare at https://doi.org/10.25380/iastate.c.5078492 (T. Zhang et al., 2021).

Can we detect more ephemeral floods with higher density harmonized Landsat Sentinel 2 data compared to Landsat 8 alone?

Spatiotemporal quantification of surface water and flooding is essential given that floods are among the largest natural hazards. Effective disaster response management requires near real-time information on flood extent. Satellite remote sensing is the only way of monitoring these dynamics across vast areas and over time. Previous water and flood mapping efforts have relied on optical time series, despite cloud contamination. This reliance on optical data is due to the availability of systematically acquired and easily accessible optical data globally for over 40 years. Prior research used either MODIS or Landsat data, trading either high temporal density but lower spatial resolution or lower cadence but higher spatial resolution. Both MODIS and Landsat pose limitations as Landsat can miss ephemeral floods, whereas MODIS misses small floods and inaccurately delineates flood edges. Leveraging high temporal frequency of 3–4 days of the existing Landsat-8 (L8) and two Sentinel-2 (S2) satellites combined, in this research, we assessed whether the increased temporal frequency of the three sensors improves our ability to detect surface water and flooding extent compared to a single sensor (L8 alone). Our study area was Australia’s Murray-Darling Basin, one of the world’s largest dryland basins that experiences ephemeral floods. We applied machine learning to NASA’s Harmonized Landsat Sentinel-2 (HLS) Surface Reflectance Product, which combines L8 and S2 observations, to map surface water and flooding dynamics. Our overall accuracy, estimated from a stratified random sample, was 99%. Our user’s and producer’s accuracy for the water class was 80% (±3.6%, standard error) and 76% (±5.8%). We focused on 2019, one of the most recent years when all three HLS sensors operated at full capacity. Our results show that water area (permanent and flooding) identified with the HLS was greater than that identified by L8, and some short-lived flooding events were detected only by the HLS. Comparison with high resolution (3 m) PlanetScope data identified extensive mixed pixels at the 30 m HLS resolution, highlighting the need for improved spatial resolution in future work. The HLS has been able to detect floods in cases when one sensor (L8) alone was not, despite 2019 being one of the driest years in the area, with few flooding events. The dense optical time-series offered by the HLS data is thus critical for capturing temporally dynamic phenomena (i.e., ephemeral floods in drylands), highlighting the importance of harmonized data such as the HLS.

TROPOspheric Monitoring Instrument observations of total column water vapour: Algorithm and validation

In this paper, we present the total column water vapour (TCWV) retrieval for the TROPOspheric Monitoring Instrument (TROPOMI) observations in the visible blue spectral band. The TROPOMI TCWV algorithm is being optimized and validated in the framework of the Sentinel 5 Precursor Product Algorithm Laboratory (S5P-PAL) project from the European Space Agency (ESA). The retrieval was first developed to retrieve TCWV from the Global Ozone Monitoring Experiment 2 (GOME-2). We have optimized the settings of the retrieval to adapt it for TROPOMI observations. The TROPOMI TCWV algorithm follows the typical two step approach, using spectral fit retrieval of slant columns, and conversion of the slant columns to vertical columns using air mass factors (AMFs). An iterative optimization algorithm is developed to dynamically find the optimal a priori water vapour profile for the AMF calculation. Further optimizations on the spectral retrieval, air mass factor calculations as well as a new surface albedo retrieval approach are implemented.The TCWV retrieval algorithm is applied to TROPOMI observations from May 2018 to May 2021. The results are validated by comparing them to ERA5 reanalysis data, GOME-2, MODerate resolution Imaging Spectroradiometer (MODIS) and Special Sensor Microwave Imager Sounder (SSMIS) satellite observations. TCWV derived from TROPOMI observations agree well with the other data sets with Pearson correlation coefficient (R) ranging from 0.96 to 0.99. The mean bias between TROPOMI and ERA5 data is −1.24 kg m−2 for measurements over land and 0.73 kg m−2 for measurements over water. The comparison to MODIS observations show similar results with small dry bias of 1.51,kg m−2 for measurements over land and a small wet bias of 1.25 kg m−2 for measurements over water. Slightly larger dry bias of 1.98 kg m−2 for measurements over land and 1.74 kg m−2 for measurements over water are found when compared to GOME-2 obserations. Compared to SSMIS data over water, TROPOMI observations are bias low by 3.25 kg m−2. The small discrepancies found between TROPOMI and reference data sets are related to the differences in measurement technique, measurement time, surface albedo issue, as well as cloud and aerosol contamination. This study demonstrates that the algorithm can provide stable and consistent results on a global scale and can be applied to generate operational TCWV products from TROPOMI and the forthcoming Copernicus missions Sentinel-4 and Sentinel-5. We have also demonstrated the capability of retrieving fine scale water vapour structures in a case study over the Amazon. This indicates that the TROPOMI data set is also useful for local and regional climate studies.

High-Resolution Seamless Daily Sea Surface Temperature Based on Satellite Data Fusion and Machine Learning over Kuroshio Extension

Sea SurfaceTemperature (SST) is a critical parameter for monitoring the marine environment and understanding various ocean phenomena. While SST can be regularly retrieved from satellite data, it often suffers from missing data due to various reasons including cloud contamination. In this study, we proposed a novel two-step data fusion framework for generating high-resolution seamless daily SST from multi-satellite data sources. The proposed approach consists of (1) SST reconstruction based on Data Interpolate Convolutional AutoEncoder (DINCAE) using the SSTs derived from two satellite sensors (i.e., Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Microwave Scanning Radiometer 2(AMSR2)), and (2) SST improvement through data fusion using random forest for consistency with in situ measurements with two schemes (i.e., scheme 1 using the reconstructed MODIS SST variables and scheme 2 using both MODIS and AMSR2 SST variables). The proposed approach was evaluated over the Kuroshio Extension in the Northwest Pacific, where a highly dynamic SST pattern can be found, from 2015 to 2019. The results showed that the reconstructed MODIS and AMSR2 SSTs through DINCAE yielded very good performance with Root Mean Square Errors (RMSEs) of 0.85 and 0.60 °C and Mean Absolute Errors (MAEs) of 0.59 and 0.45 °C, respectively. The results from the second step showed that scheme 2 and scheme 1 produced RMSEs of 0.75 and 0.98 °C and MAEs of 0.53 and 0.68 °C, respectively, compared to the in situ measurements, which proved the superiority of scheme 2 using multi-satellite data sources. Scheme 2 also showed comparable or even better performance than two operational SST products with similar spatial resolution. In particular, scheme 2 was good at simulating features with fine resolution (~50 km). The proposed approach yielded promising results over the study area, producing seamless daily SST products with high quality and high feature resolution.

Augmented Sample-Based Real-Time Spatiotemporal Spectral Unmixing

Recently, the method of spatiotemporal spectral unmixing (STSU ) was developed to fully explore multi-scale temporal information (e.g., MODIS –Landsat image pairs) for spectral unmixing of coarse time series (e.g., MODIS data). To further enhance the application for timely monitoring, the real-time STSU( RSTSU) method was developed for real-time data. In RSTSU, we usually choose a spatially complete MODIS–Landsat image pair as auxiliary data. Due to cloud contamination, the temporal distance between the required effective auxiliary data and the real-time data to be unmixed can be large, causing great land cover changes and uncertainty in the extracted unchanged pixels (i.e., training samples). In this article, to extract more reliable training samples, we propose choosing the auxiliary MODIS–Landsat data temporally closest to the prediction time. To deal with the cloud contamination in the auxiliary data, we propose an augmented sample-based RSTSU( ARSTSU) method. ARSTSU selects and augments the training samples extracted from the valid (i.e., non-cloud) area to synthesize more training samples, and then trains an effective learning model to predict the proportions. ARSTSU was validated using two MODIS data sets in the experiments. ARSTSU expands the applicability of RSTSU by solving the problem of cloud contamination in temporal neighbors in actual situations.

Estimate of Cloudy-Sky Surface Emissivity From Passive Microwave Satellite Data Using Machine Learning

The derivation of microwave land surface emissivity (MLSE) under various weather conditions from the microwave radiometer plays a crucial role in acquiring land surface and atmospheric parameters. Nevertheless, currently, most existing studies mainly focus on the clear-sky scenarios owing to a lack of cloudy-sky land surface temperature (LST) and uncertainties in simulating the scattering and emission properties of atmospheric hydrometeors. Under this background, with satellite observations and the random forest (RF) model, this study proposes a method to estimate the MLSE under cloudy skies. First, clear-sky MLSEs with satisfactory accuracy are retrieved by using the brightness temperatures (BTs) from the Advanced Microwave Scanning Radiometer-Earth sensor, LSTs from the Moderate Resolution Imaging Spectroradiometer, and atmospheric profiles from the ERA5 reanalysis. Then, the relation among the clear-sky MLSE and related impact factors is built with the RF and extended to the cloudy-sky environment for generating all-weather MLSEs with a 0.25°. The results show that the input datasets present a considerable impact on the calculation of instantaneous MLSE, and a 5.73 K bias of ERA5 LST may generate a 0.014-0.021 error in the MLSE from 6.9 to 89 GHz horizontal polarization, while the impacts of BT and profile uncertainties on the MLSE are smaller. The retrieved clear-sky MLSE is coincident with the existing MLSE for the spatiotemporal variations, and there is an average difference range from -0.035 to 0.035 in January 2008. Meanwhile, the constructed RF model can successfully apply to cloudy-sky status and recover the MLSE image gaps affected by cloud contamination.

Deep-Learning-Based Spatio-Temporal-Spectral Integrated Fusion of Heterogeneous Remote Sensing Images

It is a challenging task to integrate the spatial, temporal, and spectral information of multi-source remote sensing images, especially in the case of heterogeneous images. To this end, for the first time, this paper proposes a heterogeneous integrated framework based on a novel deep residual cycle generative adversarial network (GAN). The proposed network consists of a forward fusion part and a backward degeneration feedback part. The forward part generates the desired fusion result from the various observations; the backward degeneration feedback part considers the imaging degradation process and regenerates the observations inversely from the fusion result. The heterogeneous integrated fusion framework supported by the proposed network can simultaneously merge the complementary spatial, temporal, and spectral information of multi-source heterogeneous observations to achieve heterogeneous spatio-spectral fusion, spatio-temporal fusion, and heterogeneous spatio-temporal-spectral fusion. Furthermore, the proposed heterogeneous integrated fusion framework can be leveraged to relieve the two bottlenecks of land-cover change and thick cloud cover. Thus, the inapparent and unobserved variation trends of surface features, which are caused by the low-resolution imaging and cloud contamination, can be detected and reconstructed well. Images from many different remote sensing satellites, i.e., Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat 8, Sentinel-1, and Sentinel-2, were utilized in the experiments conducted in this study, and both the qualitative and quantitative evaluations confirmed the effectiveness of the proposed image fusion method.

Reconstruction of Optical Image Time Series With Unequal Lengths SAR Based on Improved Sequence–Sequence Model

Optical remote sensing time series are optimal for understanding and monitoring biochemical changes of key phenological parameters, which is essential for the assessment of vegetation health. However, due to cloud contamination, optical images often lack several days’ to months’ worth of information. Therefore, reconstructing optical time series based on the synthetic aperture radar (SAR) is necessary, which has the advantage of the production of continuous images under all weather conditions. In this study, an improved sequence-to-sequence (Seq2Seq) model was proposed, which integrated teacher forcing (TF) with the attention mechanism to handle input and output time series with unequal lengths. The proposed model could be used to reconstruct optical full-band time series using SAR time series data and reduce sample requirements based on the modification of the loss calculation method. To explore the spatiotemporal scalability of the proposed model, the test samples were divided into three categories: same time but different spatial domain, different time but same spatial domain, and both different temporal and spatial domains. The major findings can be summarized as follows: 1) 72.7% of the generated Landsat 8 sequence values had an absolute error of less than 0.05; 2) The mean absolute error of all bands was less than 0.0812; 3) The mean squared errors of all samples were lower than 0.015 regardless of the type of the test sample; and 4) TF was introduced to the model to improve the accuracy of the generated Landsat 8 sequence, yielding an increase of more than 5.97%. We concluded that the proposed model had good robustness both temporally and spatially. It performed well in the reconstruction of optical image time series, provided a basis for the use of time series pairs with unequal lengths, and can be applied to cloud removal and vegetation index reconstruction.

All-Weather Near-Surface Air Temperature Estimation Based on Satellite Data Over the Tibetan Plateau

Near-surface air temperature (NSAT) plays an important role in land surface and atmosphere interactions. It is widely used in many fields, such as hydrology, climatology, and environment. Although remote sensing-based approaches for NSAT estimation have been proposed by the scientific communities, many of them are limited in cloudy areas and, thus, are not able to provide all-weather NSAT estimates. To satisfy NSAT-related applications for all-weather conditions over the Tibetan Plateau (TP), this study develops a novel model for estimating daily 1-km all-weather NSAT (AW-NSAT) based on machine learning techniques. The input variables for the AW-NSAT model include land surface temperature (LST) from a newly released satellite all-weather LST dataset (i.e., TRIMS LST), as well as other parameters. The model is trained with in situ NSAT, and the results show that the random forest model with all-weather LST as the main input yields the best performance. Validation with independent in situ NSAT shows that the AW-NSAT estimate has good accuracy: an overall root mean square error of 2.43 and an R2 of 0.93. Intercomparison with an existing NSAT dataset based on MODIS LST shows that AW-NSAT has similar accuracy. Nevertheless, AW-NSAT has an evident spatial seamless characteristic, indicating that the developed model has good ability to overcome cloud contamination by introducing TRIMS LST. The developed model provides the possibility for generating the AW-NSAT for the whole TP. Furthermore, the proposed model can also be extended to other areas and then support NSAT's subsequent applications.

A Combined Approach for Monitoring Monthly Surface Water/Ice Dynamics of Lesser Slave Lake Via Earth Observation Data

Surface water/ice dynamic monitoring is crucial for many purposes, such as water resource management, agriculture, climate change, drought, and flood forecasting. New advances in remote sensing satellite data have made it possible to monitor the surface water/ice dynamics both spatially and temporally. However, there are many challenges when using these data, such as the availability of valid imagery, cloud contamination issues for Landsat-8, and sensitivity of Sentinel-1 C-band to wind speed, topography, and others. A combined methodology using Landsat-8 and Sentinel-1 Synthetic Aperture Radar (SAR) data was proposed to create monthly change maps at 30 m spatial resolution for the Lesser Slave Lake in Alberta, Canada, for the period 2017-2020. The potentials of multi-spectral indices for Landsat-8, such as the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Modified NDWI (MNDWI) as well as the Sentinel-1 SAR backscattering coefficients (VV-VH) and Normalized Difference Polarized Index (NDPI) for separating water/ice from the land were investigated. The results obtained from satellite data with historical discharge and water level measurements for the lake were compared. Furthermore, the results show that the MNDWI and VH are the most effective indices for creating the change maps. The overall accuracies achieved for MNDWI and VH are 92.10% and 68.86% for cold months and 99.88% and 98.49% for warm months, respectively.

Multitask Deep Learning Framework for Spatiotemporal Fusion of NDVI

High spatial and temporal resolution normalized difference vegetation index (NDVI) time series are indispensable for monitoring land surfaces dynamics in spatiotemporally heterogeneous areas. Spatiotemporal fusion (STF) is one of the most common methods used for producing such data. These methods require the use of one or two pairs of fine images (with fine spatial but rough temporal resolution, such as Landsat images) and coarse images [with fine temporal but rough spatial resolution, such as Moderate Resolution Imaging Spectroradiometer (MODIS)]. A coarse image at the prediction date is also required to predict the corresponding missing fine image in the time series. Recently, the proposed deep learning (DL)-based STF methods have achieved promising fusion performance but are challenged in areas with frequent cloud contamination and landcover change prediction, while they also suffer from unstable fusion performance. Moreover, current STF methods lack a quality assessment process for the fusion results. To address these limitations, in this article, we propose a multitask DL framework for STF of NDVI time series, which integrates two types of DL-based STF methods. Four experiments in two study sites were conducted to test the effectiveness of the proposed method, and the results indicate that it achieves accurate and stable fusion capable of predicting landcover changes even when image pairs obtained at long intervals are used. In addition, a fusion uncertainty estimation method is proposed, which has the potential to be used as a quality assessment metric.

Removing Cloud Cover Interference from Sentinel-2 imagery in Google Earth Engine by fusing Sentinel-1 SAR data with a CNNmodel

ABSTRACT Google Earth Engine (GEE) provides a convenient platform that enables the use of a diverse array of applications based on optical satellite images covering long time periods and large areas. However, cloud contamination often causes optical satellite images to lose land surface information, which can greatly affect their spatial and temporal availability in areas with frequent cloud cover. Reconstructing the missing values is very important when researchers attempt to improve the potential for using optical satellite images. Recently, convolutional neural network (CNN)-based methods have shown great potential in cloud removal. Unfortunately, current CNN-based cloud removal methods can only be applied locally, which would result in inefficient data downloading and storage problems. To address these problems, this paper proposes a method that integrates GEE data and a multi-level feature connected CNN (DeepGEE-S2CR) to remove clouds in Sentinel-2 imagery using Sentinel-1 Synthetic Aperture Radar imagery as auxiliary data. Specifically, the training process of the multi-level feature connected CNN was conducted locally, so that cloud removal could then be performed on the client of GEE using online data with the trained CNN model. An experiment using a set of globally distributed Sentinel-2 images was conducted to validate the proposed method; the results showed the process could generate results comparable to a classic deep Sentinel-2 cloud removal method (DSen2-CR) that should be run locally (the average root mean square errors of DeepGEE-S2CR and DSen2-CR were 0.045 and 0.042, respectively). The proposed method is especially convenient for removing clouds in Sentinel-2 imagery covering a large area.

Seamless and automated rapeseed mapping for large cloudy regions using time-series optical satellite imagery

Accurate field-level mapping of rapeseed spatial distributions using remote sensing technology supports agricultural planning, food and edible oil security, and ecological system management. The phenological and spectral methods for rapeseed mapping mainly rely on optical images collected during the flowering period. However, these images are often occluded by clouds. Moreover, the scarcity of well-represented training samples with various phenology patterns constrains large-scale supervised classification and extraction of rapeseed. To solve these problems, we propose a novel Seamless and Automated Rapeseed Mapping (SARM) method using time-series optical satellite imagery for large cloudy regions to generate seamless rapeseed maps at field scale without any manually collected samples. First, a Winter Rapeseed Index (WRI) was designed to automatically extract rapeseed in cloud-free areas by enhancing the phenological separability of rapeseed. Second, a multiple random forest-based fusion mapping framework was developed using WRI-based training samples, and was applied for the cloudy regions to map rapeseed seamlessly across the whole study area. The proposed SARM method was validated under different years, sensors, and cropping patterns at four study sites both in China and the USA. Extensive experimental results and analysis suggest that the SARM method outperforms other traditional methods on all four study sites with an average overall accuracy varying from 88.58% to 98.78%. These results also reveal that our proposed approach is capable of expanding the time windows for rapeseed mapping from the significant flowering period to the insignificant flowering period. More importantly, the spatial incompleteness of rapeseed maps caused by cloud contamination was mitigated to some degree. The SARM method robustly delivers seamless and automated rapeseed mapping by using time-series optical satellite imagery, thus offering new insights for crop investigation, especially in large cloudy regions.

Intercomparison of CO measurements from TROPOMI, ACE-FTS, and a high-Arctic ground-based Fourier transform spectrometer

Abstract. The TROPOspheric Monitoring Instrument (TROPOMI) provides a daily, spatially resolved (initially 7×7 km2, upgraded to 7×5.6 km2 in August 2019) global dataset of CO columns; however, due to the relative sparseness of reliable ground-based data sources, it can be challenging to characterize the validity and accuracy of satellite data products in remote regions such as the high Arctic. In these regions, satellite intercomparisons can supplement model- and ground-based validation efforts and serve to verify previously observed differences. In this paper, we compare the CO products from TROPOMI, the Atmospheric Chemistry Experiment (ACE) Fourier transform spectrometer (FTS), and a high-Arctic ground-based FTS located at the Polar Environment Atmospheric Research Laboratory (PEARL) in Eureka, Nunavut (80.05∘ N, 86.42∘ W). A global comparison of TROPOMI reference profiles scaled by the retrieved total column with ACE-FTS CO partial columns for the period from 28 November 2017 to 31 May 2020 displays excellent agreement between the two datasets (R=0.93) and a small relative bias of -0.83±0.26% (bias ± standard error of the mean). Additional comparisons were performed within five latitude bands: the north polar region (60 to 90∘ N), northern mid-latitudes (20 to 60∘ N), the equatorial region (20∘ S to 20∘ N), southern mid-latitudes (60 to 20∘ S), and the south polar region (90 to 60∘ S). Latitudinal comparisons of the TROPOMI and ACE-FTS CO datasets show strong correlations ranging from R=0.93 (southern mid-latitudes) to R=0.86 (equatorial region) between the CO products but display a dependence of the mean differences on latitude. Positive mean biases of 7.93±0.61 % and 7.21±0.52 % were found in the northern and southern polar regions, respectively, while a negative bias of -9.41±0.55% was observed in the equatorial region. To investigate whether these differences are introduced by cloud contamination, which is reflected in the TROPOMI averaging kernel shape, the latitudinal comparisons were repeated for cloud-covered pixels and clear-sky pixels only, as well as for the unsmoothed and smoothed cases. Clear-sky pixels were found to be biased higher with poorer correlations on average than clear+cloudy scenes and cloud-covered scenes only. Furthermore, the latitudinal dependence on the biases was observed in both the smoothed and unsmoothed cases. To provide additional context to the global comparisons of TROPOMI with ACE-FTS in the Arctic, both satellite datasets were compared against measurements from the ground-based PEARL-FTS. Comparisons of TROPOMI with smoothed PEARL-FTS total columns in the period of 3 March 2018 to 27 March 2020 display a strong correlation (R=0.88); however, a positive mean bias of 14.7±0.16 % was also found. A partial column comparison of ACE-FTS with the PEARL-FTS in the period from 25 February 2007 to 18 March 2020 shows good agreement (R=0.79) and a mean positive bias of 7.89±0.21 % in the ACE-FTS product relative to the ground-based FTS. The magnitude and sign of the mean relative differences are consistent across all intercomparisons in this work, as well as with recent ground-based validation efforts, suggesting that the current TROPOMI CO product exhibits a positive bias in the high-Arctic region. However, the observed bias is within the TROPOMI mission accuracy requirement of ±15 %, providing further confirmation that the data quality in these remote high-latitude regions meets this specification.