Abstract:The change of Arctic sea ice has recently attracted much attention among climate researchers due to the climate effect of “Arctic Amplification”. Sea ice concentration, which is the main parameter of passive microwave remote sensing of sea ice, can characterize the sea ice conditions, which can be used to guide the polar navigation and study the sea ice change in different scales. The sea ice area and sea ice extent can also be calculated by using the sea ice concentration, which is of great significance for the forecast of polar sea ice conditions and the study of climate change. This work discusses how to use the high resolution channels of FY-3B/MWRI (FY-3B/Microwave Radiometer Imager) to retrieve the sea ice concentrations in the Arctic. Based on the ASI (ARTIST [Arctic Radiation and Turbulence Interaction Study] Sea Ice) algorithm, the Arctic sea ice concentration is calculated in this study by improving the tie points of the algorithm. According to the cross calibration of brightness temperatures between the FY-3B/MWRI and the Aqua/AMSR-E (Advanced Microwave Scanning Radiometer-EOS), the differences between the two brightness temperature data are between ±4 K, which will result in a maximum bright temperature difference of 8 K. Accordingly, this study first sets the variation range of the tie points for FY-3B/MWRI centered on the original values of the ASI algorithm to 11.7±8 K for sea ice and 47.0±8 K for open water, separately, with a step length of 1 K. After the combination, 289 series of point value combinations are obtained. Then, the sea ice concentrations corresponding to each set of tie points are compared with those of the AMSR-E L3 product. Meanwhile, the tie points corresponding to the smallest deviation of the two data sets are selected. According to the tie points determined above, this study calculated the Arctic sea ice concentrations based on the FY-3B/MWRI brightness temperatures, hereinafter referred to as the Retrieved Sea Ice Concentration (RSIC). The RSICs in this study are compared with the MWRI Level 2 sea ice concentration product (hereinafter referred to as MWRI). First, the sea ice concentrations obtained from the Aqua/MODIS (Moderate Resolution Imaging Spectroradiometer) reflectivity data from July to September 2011 are used to verify the two data sets. The results show that the bias of RSIC is comparable to that of the MWRI product. However, the standard deviation and root mean square error are significantly reduced. Meanwhile, the accuracy of RSIC is much higher than that of the MWRI product in the areas with a sea ice concentration lower than 95%. Then, the two sea ice concentration data sets are compared with the sea ice concentration product from the University of Bremen (SIC_UB). The bias and the standard deviation between RSIC and SIC_UB are 3.3% and 10.6%, which are lower than the values between the MWRI and the SIC_UB products: 5.9% and 16.4%, respectively. Finally, the time series of the daily averaged sea ice concentration, sea ice area, and sea ice extent from the RSIC, MWRI, SIC_UB, and NSIDC/AMSR-E (National Snow and Ice Data Center/Advanced Microwave Scanning Radiometer-E) sea ice concentration products are compared. The results show that the values of RSIC are significantly lower than those of the MWRI product in three statistical methods and much closer to the AMSR-E and SIC_UB products. In this study, the sea ice concentrations in the Arctic region are retrieved based on the brightness temperatures from the FY-3B/MWRI high frequency channels. The sea ice concentrations in this study have a higher spatial resolution and a better accuracy compared with the FY-3B/MWRI L2 sea ice concentration product, which is conducive to the long-time series study of climate change in the Arctic.
Abstract:Cloud pixel detection is a crucial pre-process in numerous remote sensing applications, such as the aerosol parameter retrieval, land use change, anomaly detection/classification, crop monitoring, and marine ecological survey. On the one hand, cloud pixel misidentification (mistaken as surface or aerosol) in optical images has substantial negative effects on the above-mentioned traditional applications, due to the significant influence of cloud layers on shortwave radiation. On the other hand, traditional threshold methods are largely constrained in applicability by the high spatiotemporal heterogeneity of aerosols and clouds as well as the diversification of satellite sensors and spectral channels. Therefore, a reliable cloud detection method with wide applicability is in demand, especially for studies over multiple surfaces (i.e., land, ocean, and cryosphere). Thus, this study proposed an innovative method based on the radiative transfer simulation and machine learning, namely CRMC (Combine Reflectance simulation and Machine learning for Cloud detection), to detect cloud pixels in optical images produced from the MERSI Ⅱ sensor onboard the FY-3D satellite. The biggest advantage of this method is its compatibility with different sensors, generating cloud pixel samples through the physical process simulation, and applying the machine learning technique to learn sample features, thereby excluding the influence of anthropogenic factors. Specifically, to address the mismatch between the MODIS cloud detection algorithm and MERSI-based aerosol inversion, the CRMC method sets different Inherent Optical Properties (IOPs) of surface and atmospheric objects, considering binomial reflection characteristics of underlying surfaces and various parameters of aerosols and clouds. In addition, the method outputs cloud probabilities in pixels and allows custom thresholds to control the strictness level of cloud pixel detection. The CRMC method mainly includes three steps: (1) Defining 11 typical underlying reflectance parameters from MODIS binomial reflection products using a cluster analysis approach; (2) Inputting the typical underlying reflectance and aerosol and cloud parameters with random inherent optical properties into the SBDART radiation transmission model to obtain a simulated reflectance dataset for training the shallow neural network; (3) Calculating the cloud probability of the target image with the trained shallow neural network and selecting a suitable threshold according to the actual need to complete the cloud detection. Compared with the CALIPSO Vertical Feature Mask (VFM), results of the CRMC method show a maximum total accuracy of 79.6% (78.5% and 81.2% on land and sea, respectively). Under the condition of cloud probability threshold=0.2 (hit rates for cloud and cloud-free pixels are the same in these cases), the CRMC outperforms MODIS cloud mask products (MYD35) over land, especially on broad-leaved forest, farmland, urban and bare soil. However, the accuracy of the CRMC over sea is lower than that of MYD35. In the view of the surface uniformity, the cloud detection over sea can enrich the brightness temperature information to optimize the corresponding performance. To sum up, while greatly improving the applicability to different optical sensors by not relying on special spectral range, the CRMC method can achieve a good cloud pixel identification effect for FY-3D MERSI Ⅱ images, with a similar hit rate compared with MODIS products. Also, the CRMC has certain anti-disturbance ability to haze.
Keywords:cloud detection;Radiative Transfer Simulation;neural network;FY-3D;MERSI II
Abstract:FengYun-3C (FY-3C) is the first satellite of the second-generation polar-orbiting operational meteorological satellite in China. As one of the key payloads onboard FY-3C, the MicroWave Humidity and Temperature Sounder (MWHTS) is a cross-track microwave sounder and has 15 channels ranging from 89.0 GHz to 191.0 GHz, with eight (channels 2—9) located near 118.75 GHz along an oxygen absorption line, five (channels 11—15) close to the 183.31 GHz water vapor absorption line, and the remaining two window channels (1 and 10) centered at 89.0 and 150.0 GHz. This instrument’s measurement allows for probing the atmospheric temperature and moisture under clear and cloudy conditions. The MWHTS attracted worldwide attention because of its special configuration. FY-3C MWHTS radiance data have already been assimilated into operational numerical weather prediction models in the European Centre for Medium-Range Weather Forecasts, UK Met Office, and China Meteorological Administration. The calibration accuracy and stability of MWHTS can directly affect the data assimilation effects in NWP. This research establishes a quality control model and observed brightness temperature quality score for MWHTS to filter out the poor quality data during the calibration processing. The five and a half years historical raw data from MWHTS are analyzed. The telemetry parameters from the raw data considered in this study include the blackbody target temperature, instrument temperature, instrument component temperature, counts of the blackbody target and cold space, scan angles, and scan periods. These telemetry parameters thresholds were set accordingly for quality control. Then, based on the radiometer calibration transfer function and observation mechanism of MWHTS, five key parameters (instrument temperature, blackbody target temperature, blackbody view counts, cold space view counts, and scan periods) were selected to score the MWHTS calibration data quality. The sensitivity analysis of each parameter to the differences between the observations and radiance transfer simulations were carried out. The results show that the scan period has the most significant influence on the O-B results, and the instrument temperature has the least effect. The effect proportion was used as the weight to score the observed brightness temperature in centesimal system. The results show that the quality control scheme of each parameter can eliminate abnormal data, and the quality scoring system characterizes the MWHTS calibration quality, and the data application is ensured. The quality control model is established for FY-3C MWHTS to meet the application requirements of onboard microwave observation data. The threshold of the quality control mode depends on the various characteristics of the telemetry data in orbit. This model has been used in the operational calibration algorithm of FY-3C MWHTS, and the score results are included in the MWHTS L1 data to global real-time releases. The MWHTS observed brightness temperature quality score can indicate the data quality throughout the operational in-orbit radiometer calibration. The higher the score, the better the data quality. Accordingly, users can choose the score threshold for data availability according to the application requirements. The quality scoring system is based on only five key telemetry parameters, and more parameters will be analyzed to improve this system in the future.
Keywords:Fengyun-3C;Microwave Humidity and Temperature Sounder (MWHTS);quality control;observed brightness temperature score;telemetry parameters
Abstract:The chlorophyll-a (Chla) concentration that refers to the content of Chla contained in per unit volume water, is a key indicator describing the eutrophication degree of lake waters. Accurate quantification of Chla concentration is of great significance for water environment assessment and water quality monitoring and has become a focus on the study of watercolor remote sensing. Orbita hyperspectral (OHS) satellite is a new generation of hyperspectral satellites launched by Zhuhai Orbita Aerospace Technology Co., Ltd. in 2018, which covers spectral range of 400~1000 nm and 32 spectral channels with both high spectral and high spatial resolution (2.5 nm and 10 m, respectively), showing great potential for inland water quality monitoring application. However, this satellite has a short operating period from launch, and the applicability of the generated images needs to be further investigated and validated.Dianchi Lake, a typical eutrophic plateau lake, was used as the study area for Chla concentration retrieval based on the OHS hyperspectral imagery. We collected in-situ spectra and Chla concentration from two cruise surveys in Dianchi Lake and acquired the satellite-ground synchronization data within one day of the OHS satellite overpass. Data from two field campaigns including 72 sampling sites were used for model calibration and validation, and ground data matched with satellite overpass including 10 sampling sites was used to further validate the retrieval results after the calibrated model was applied to the OHS imagery. We first utilized all 72 in-situ spectra to explore the relationship between all possible combinations of band ratio and Chla concentrations to seek the optimal band ratio model. Immediately after we used the spectral response function of the OHS imagery to resample the in-situ spectra to the band configuration of the OHS imagery, the OHS-based band ratio model was calibrated using 48 field-measured data according to the optimal band ratio combination of in-situ spectra, and the remaining 24 data were used to evaluate model accuracy. We further validated the retrieval results using the Chla concentrations at 10 sampling points synchronized with the OHS image after the OHS-based band ratio model was applied to the OHS image, and the spatial pattern of Chla concentration in Dianchi Lake was revealed.The band ratio Rrs(716)/Rrs(595) had the highest correlation with Chla concentration in terms of the in-situ spectra with R2=0.819, and the corresponding OHS-based band ratio model (B17/B9) was suitable for remote sensing retrieval of Chla concentration in Dianchi Lake with R2 of 0.804, the root-mean-square error (RMSE) of 6.99 μg/L and the mean absolute percentage error (MAPE) of 6.32%. The retrieval results of the OHS-based band ratio model applied to the OHS image and the spatial pattern of Chla concentration were reasonable with acceptable errors (RMSE=12.47 μg/L, MAPE=22.53%). The spatial pattern of Chla concentration in Dianchi Lake showed a decreasing trend from the lakeshore to the center of the lake on April 2, 2019, the northeast and southeast decrease fitted a power function, whereas the northwest decrease described a linear function. The pixel reflectance of the nearshore waters may be higher than that of the normal waters due to the land adjacency effect, which may lead to a high concentration of retrieved Chla along the coast. In the OHS imagery of Dianchi Lake, four nearshore water pixels could be easily influenced by the land adjacency effect, so these four pixels needed to be masked to eliminate the influence. In addition, compared with the existing Chla concentration retrieval algorithms, the band ratio model (B17/B9) proposed in this study improved the retrieval accuracy of Chla concentration.In conclusion, the OHS-based band ratio model works efficiently and reliably for retrieving Chla concentration in Dianchi Lake. OHS hyperspectral data show great potential in terms of accurate retrieval of Chla concentration for inland waters, providing a new means for remote sensing monitoring of Chla concentration. However, whether the OHS-based band ratio model developed in this study applies to other water bodies with different optical properties still needs to be further investigated and tested. In future studies, the performance of the model will be further examined by collecting more field data in different lakes.
Abstract:Passive microwave remote sensing is an important method for observing the sea ice concentration (SIC). The ASI (ARTIST Sea Ice algorithm) can obtain the highest spatial resolution of 6.25 km2 in the current SIC products by using the dual-polarized brightness temperature of AMSR-E at 89 GHz. However, additional auxiliary low-frequency band data are still needed for weather filter. In this work, a new SIC retrieving algorithm based on single-frequency multi-incident angle brightness temperature data has been proposed and studied, which can be applied on an 89 GHz synthetic aperture radiometer. Fully utilizing the multi-incident angle brightness temperature with a synthetic aperture radiometer can separate the information of sea ice and seawater, improve the precision of SIC retrieving, and achieve high spatial resolution. The first step involves creating a simulation system for space-borne observation of sea–ice radiant brightness temperature and using the measured data of the 89 GHz channel from the FY-3C/MWHS to carry out the sensitivity analysis between brightness temperature and angle at 89 GHz. The second step consists of developing an SIC retrieving algorithm based on the angle brightness temperature difference at 89 GHz and completing the preliminary retrieving verification by combining the SIC product of ECMWF and the brightness temperature simulation system. The results show that the SIC retrieval can be realized by using the incident angle brightness temperature difference. Moreover, the minimum root mean square weighted average post-processing by using the combination of multi-incidence angle can minimize the retrieving error of SIC. When the input Gaussian white noise is 2 K, 5% SIC error can be obtained. The final results show that the SIC retrieval with the combination of multi-incidence angle can fully expand the distinction between sea water and sea ice.
Abstract:Based on the satellite-based active-passive sensor data, this study uses the Constrained Spectral Radiance Matching (CSRM) algorithm to construct 3D structural models of deep convective clouds in eight western Pacific typhoon events, and accordingly analyses the deep convective clouds in different stages of typhoon development in terms of The horizontal distribution characteristics and microphysical characteristics of deep convective clouds in different stages of typhoon development are analysed. The significance of using CSRM algorithm for atmospheric 3D structure construction is to transfer the vertical distribution information from active remote sensing to the detection range of passive remote sensing, increase the number of contours available for statistical analysis of the eight typhoon events, and further ensure the validity of statistical analysis of deep convective clouds. The results of statistical analysis show that: (1) The proportion of deep convective clouds in the enhanced typhoons studied in this paper is 61.4% on average, and the standard deviation along all directions is 12.4% on average. In comparison, the proportion and standard deviation of typhoons in the weakening period decreased, which were 25.3% and 8.4%, respectively. Compared with typhoons, the average proportion of deep convective clouds in tropical storms is 45.2%, and the difference is greater in different directions, with a standard deviation of 23.1%. (2) The ice cloud effective radius of deep convective clouds in tropical cyclones is proportional to the height. The ice water number concentration first increases and then decreases with the increase in height. The vertical distribution of Ice Water Content (IWC) of deep convective clouds at different development stages of the same tropical cyclone is quite different. When a tropical cyclone develops into a typhoon, the IWC high-value area of its deep convective clouds gradually concentrates toward the top of the cloud layer, and the maximum value of IWC significantly increases as the typhoon transforms from an intensified period to a weakened period.
Abstract:Satellite-borne passive terahertz remote sensing is currently most promising method for ice cloud sounding due to a number of potential advantages that complement existing visible and infrared techniques. Since the wavelength of terahertz radiation is comparable to the size of ice crystals, observed brightness temperature are well correlated to ice mass. The purpose of this article is to present the terahertz radiation and scattering characteristics of ice and graupel particles which constitute ice clouds and to understand how ice cloud parameters affect the terahertz brightness temperature, which guide the design of terahertz ice cloud sounding instruments. In this study, the terahertz brightness temperature and Jacobian matrix are simulated by applying the Discrete-Ordinate Tangent Linear Radiative Transfer (DOTLRT) model to the spatially and microphysically detailed output of ice clouds predicted by the Weather Research and Forecasting (WRF) model and Final (FNL) analysis data. The DOTLRT model uses the classical Mie scattering formula to calculate the scattering characteristics of liquid, rain, ice, snow, and graupel particles in clouds. The validity of the simulated 183 GHz brightness temperature is verified by comparison with the collocated observation from the Advanced Technology Microwave Sounder (ATMS). The simulation shows that the terahertz brightness temperature of the ice clouds is affected by ice and graupel particles with different properties, e.g., for a cloud with IWP of 300 g/m2 and GWP of 300 g/m2, the brightness temperature depression due to ice and graupel particles at 183.31+2.0 GHz are 2.28 K and 12.26 K respectively, and at 183.31+7.0 GHz are 6.30 K and 62.37 K respectively. Then, the impacts of the observation geometry and the parameters of the ice and graupel particles, including mass equivalent spherical diameter (Dme), Ice Water Path (IWP), and Graupel Water Path (GWP), on the terahertz brightness temperature were quantitatively analyzed. Finally, the optimal sounding frequency bands (183 GHz, 243 GHz, 325 GHz, 448 GHz, 664 GHz, and 874 GHz) and observation angle (53°±5°) of the terahertz ice cloud sounding instrument were derived based on the sensitivity analysis of the terahertz brightness temperature, Jacobian matrix, and trace gas absorption. The calculated terahertz brightness temperature spectrums indicate that the ice and graupel particles need to be considered in the terahertz ice cloud remote sensing. The terahertz radiation and scattering characteristics of the ice and graupel particles studied in this work can provide technical support for the development of the future terahertz ice cloud sounding instrument.
Keywords:Terahertz;ice cloud remote sensing;the radiation and scattering characteristics;brightness temperature simulation;Ice Water Path;mass equivalent spherical diameter;ice particles;graupel particles
Abstract:Aerosols play an important role in determining the Earth's radiation budget and its impact on climate change. Aerosol optical depth (AOD) is a crucial fundamental parameter for meteorological observation and a basic optical property of aerosol derived from satellites. Over land, the aerosol contribution in satellite signals is small compared with the surface, making it difficult to separate the aerosol path radiance from satellite measurements, particularly over the urban area. In the past several decades, numerous different AOD retrieval algorithms have been proposed by using different satellite sensors, but most of them do not consider surface anisotropy.The main purpose of this work is to improve the accuracy of aerosol retrievals and reduce the uncertainty of the operational MODIS AOD products over mixed surfaces. On this basis, a new generic high-performance aerosol retrieval algorithm is presented and explained. The new method is developed by coupling the non-Lambertian atmospheric radiative transfer model and semiempirical linear kernel-driven BRDF model. First, an a priori surface BRDF shape parameter database is constructed using the daily MODIS BRDF/Albedo product by using penalized least square regression based on a 3D discrete cosine transform (DCT-PLS) method. Then, the estimation of surface reflectance, including bidirectional reflectance, directional to hemispheric reflectance, hemispheric to directional reflectance, and bi-hemispheric reflectance (also called white-sky albedo, WSA), is based on this database and kernel-driven BRDF model. The presented method is tested on the Landsat 8 OLI images around the Beijing area, which features highly heterogeneous surfaces and severe air pollution problems. AOD retrievals with 500 m resolution can be successfully obtained over dark and bright surfaces.An accuracy assessment of the new algorithm, WSA-derived and HARLS AOD retrievals against AERONET AOD, from the four selected stations indicated the superiority of new algorithm, which is reflected in the high PWE and low RMSE. The comparison results show that the new algorithm is in good agreement with ground-based AOD (R=0.911) compared with the WSA-derived and HARLS AOD retrievals. Furthermore, the new algorithm and MODIS aerosol algorithms have similar spatial patterns of AOD. The new algorithm significantly improves the accuracy of aerosol retrievals, which is verified by AERONET AOD data, especially over brighter surfaces, because surface anisotropy is considered in this algorithm. The new algorithm can provide a detailed AOD spatial distribution over mixed surfaces and shows high ability in capturing fine-scale features. The new algorithm and MAIAC AOD retrievals have a similar spread of uncertainty envelopes. However, the new algorithm AOD retrievals have a higher correlation and smaller RMSE than the MAIAC retrievals, and the number of collections with AERONET for the new algorithm is almost 1.5 times those for MAIAC.This new AOD retrieval algorithm can provide a possibility for high-precision urban aerosol remote sensing monitoring and solve other pressing issues, such as long-term trend analysis of urban aerosols and air quality conditions, especially in heavily polluted areas. Based on the collocated observations, the new algorithm achieved satisfactory retrieval accuracy. However, several issues remain to be solved in the future. First, the retrieval errors of the MODIS BRDF kernel parameters are also a major source of uncertainty. Second, more analyses of the aerosol models and model selection are required. Third, the application in other regions and sensors is required in further work to evaluate the applicability of new algorithm.
Keywords:aerosol optical depth;surface anisotropy;kernel-driven BRDF model;MODIS;Landsat 8 OLI
Abstract:The distribution of altitudinal natural zones is closely related to the natural characteristics and service functions of mountain ecosystems. Quantitative identification of its elevation range can efficiently capture the results of the interaction between climate change and vegetation. Digital extraction and extensive analysis in such a critical elevation range is often restricted by the suitability of the experimental model and the spatial-temporal continuity of the data. In this study, temperate deserts, montane steppes, montane coniferous forests, alpine meadows and alpine cushion vegetation were considered as the stable state of the ecosystem. The land surface temperature was derived from the Landsat 8 Thermal Infrared Sensor data using the mono-window algorithm. Latent class analysis was conducted to test whether the remotely sensed indicator exhibited a multi-peak mode and to evaluate the number of states in an actual ecosystem. The potential energy of two adjacent states of the altitudinal natural zone was estimated by the potential energy analysis model, and the ecological transition range and demarcation elevation on the northern slope of Bogda Mountain were identified. The results showed that the frequency distribution of LST was bimodal, indicating the presence of distinct alternative modes. Moreover, the number of system modes changed from 1 to 2, and the AIC rapidly decreased. Both states revealed the existence of different ecosystem states. Second, potential energy analysis was used to identify the ecological transition area and demarcation elevation between adjacent altitudinal natural zones along the elevation gradient, temperate desert-montane steppes (transition range: 1062—1093 m, demarcation elevation: 1066 m), montane steppes-montane coniferous forests (transition range: 1689—1764 m, demarcation elevation: 1707 m), montane coniferous forests-alpine meadows (transition range: 2690—2744 m, demarcation elevation: 2714 m), and alpine meadows-alpine cushion vegetation (transition range: 3251—3263 m, demarcation elevation: 3257 m). Finally, the reliability of the results was verified by field survey data in July 2018, the overall RMSE was 17.19 m. In this study, the remotely sensed indicator and the principle of ecological stable states in combination with the potential energy analysis method provide a reference basis for quickly and accurately extracting the ecological transition range and demarcation elevation of the altitudinal natural zones.
Keywords:Vegetation altitudinal natural zones;ecosystem states;critical transitions;potential energy analysis;Bogda Mountain
Abstract:This study aims to understand the potential effects of climate change on vegetation in arid and semi-arid areas of Central Asia and analyze the temporal responses of different typical vegetation to climate factors. The MODIS-NDVI dataset and CHELSA climate data were used for regression analysis and time-series correlation analysis, which extracted the spatial pattern and variation trend of NDVI, precipitation and temperature in the period of 2000 to 2013. Meanwhile, joint land use/cover type and terrestrial ecoregion to analyze the correlations of NDVI to climate factors for different vegetation types and different ecoregions could help explore the mechanism of vegetation response to climate change in Central Asia. The results showed that: (1) The NDVI values of Central Asia were generally low and exhibited a decreasing trend from high altitude to low altitude. The spatial distribution of precipitation decreased from the east to the west. The spatial distribution of temperature decreased from the west to the east in Central Asia. From 2000 to 2013, the spatial variation trend of NDVI, precipitation and temperature were the significant spatial differences in Central Asia. (2) There was a significant positive correlation between NDVI and precipitation in most areas. The negative correlation pixels between NDVI and precipitation were mainly distributed in the relatively humid areas in central and southern Central Asia and along the rivers and lakes. Meanwhile, a positive correlation between NDVI and temperature was found in eastern and northern Central Asia. The low correlation between NDVI and temperature mainly distributed in the arid and semi-arid regions, only a few areas were showed negative correlations. In addition, the correlation of NDVI between precipitation and temperature was significant in different vegetation types. (3) There was a lag time of 32 days and a cumulative time of 16-96 days between NDVI and precipitation in most parts of Central Asia. The correlation coefficient between NDVI and temperature of the current period was relatively high and the high correlation coefficients of the cumulative time were 16-48 days. The response of NDVI of different typical vegetation to precipitation is different, and the response of NDVI to temperature is generally similar with other vegetation types except for bare land. (4) The responses of NDVI to precipitation/temperature were different in different ecoregions. Among them, the longest lag time and cumulation time of precipitation in the forest were 96d. There were 34.16% pixels of alpine meadow showed the 96d lag times for the correlation of NDVI response to precipitation, the lag time and cumulation time of NDVI response to temperature were 16 d and 32 d.
Keywords:NDVI;climate response;Central Asia;time-series correlation analysis;temporal response mechanism;Arid and semi-arid areas
Abstract:With the development of modern remote sensing observation technology, there are several versions of public elevation datasets. However, due to the shielding effect or absorption of electromagnetic waves by dense vegetation in forest areas, the elevation obtained by remote sensing techniques is inevitably influenced by the systematic deviation. In this paper, the radar-derived DEM products (i.e., SRTM1 and TanDEM-X 90) and optical-sensor based AW3D30 dataset were chosen to analyze the characteristics of elevation deviations of multi-source DEM products in forest areas. The Maryland and Sonoma County of the United States were selected as the study regions. Using the 3DEP as the reference of bare-ground elevation, the vegetation-induced penetration rates of the three elevation products in the areas with different terrain conditions and covered by various forest types were obtained. The results showed that the overall penetration rate of SRTM1 obtained through C-band radar detection is the largest among the three datasets, followed by TanDEM-X 90 and AW3D 30 with comparable penetration rates. In terms of forest types, the penetration rates in coniferous forests are greater than that in broad-leaved forest. Moreover, the forest penetration rate decreases with the increase of canopy coverage, especially for TanDEM-X 90 and AW3D 30. In addition to the influence of detection wavelength, the main reason is that both TanDEM-X 90 and AW3D 30 are sampled from high-resolution data, and the high-resolution raw data before resampling make it easier to detect the open area between trees. In addition, the complex terrain environment will decrease the vertical accuracy of DEM datasets, and result in relatively lower penetration rates. Based on the results, the vegetation-induced penetration rate was synthetically affected by multiple factors, including sensor’s working waveband, forest type, canopy coverage, and terrain slope. The canopy height should be removed for the elevation datasets considering different penetration power of vegetation to obtain more accurate surface information. This research is helpful in improving the accuracy of canopy height removal for elevation products, and it provides a scientific reference basis for users when selecting the appropriate elevation products in the applications.
Abstract:Vegetation plays an important role in the earth’s ecosystem, and it has always been a focus of quantitative remote sensing research. The spatial heterogeneity of the vegetation canopy results in the complicated Bidirectional Reflectance Factor (BRF) distribution which brings great difficulties to the high-resolution vegetation remote sensing research. The 3D radiation transmission model can accurately describes the interaction between heterogeneous vegetation canopy and solar radiation, which is important for modeling and application of vegetation quantitative remote sensing using high-resolution optical data.In this work, the probabilistic plants with statistical properties similar to actual plants are used to construct a 3D vegetation canopy. Combining Monte Carlo ray tracing technology and canopy porosity calculation, a 3D canopy radiation transmission model is constructed by rationally designing the random transmission process of optical particles in the canopy. Using probabilistic plants to build the canopy not only accurately describes the canopy’s heterogeneity but also takes into account the non-uniform spatial distribution of leaves. The model considers light to be a particle with the dual properties of frequency and energy. The Monte Carlo method is used to simulate the transmission behavior of light particles in the canopy, and the canopy porosity is calculated to improve the model’s calculation stability.Taking the corn canopy with a typical ridge planting structure as an example, the bidirectional reflectance function of the corn canopy in different growth periods is simulated and compared with the multi-angle observation data. The comparison results illustrate that the model has good simulation accuracy for the optical BRF of corn canopy in different growing periods. The RMSE between the model simulation value and the measured value is 0.0085 at the red band, and the R2 is 0.96. At the near-infrared band, the RMSE and R2 are 0.013 and 0.96, respectively. For the homogeneous canopy, the comparison between the model developed in this work and the SAIL model shows that the model of this paper perfectly describes the transport path and energy transfer of the light particles.The model proposed in this paper can accurately simulate the BRF of the canopy with the centrosymmetric statistical characteristics, and it offers great convenience in constructing inhomogeneous canopy scenes as well as high simulation stability. The model provides an effective simulation tool for studying canopy BRF and supporting verification of key canopy parameters such as vegetation biomass and leaf area index.
Keywords:probability plants;three-dimensional canopy;Monte Carlo;canopy porosity;bidirectional reflectance function
Abstract:Some existing target detection algorithms are insufficient for feature extraction in remote sensing images. They cannot solve the difficult problem of large target scale differences in remote sensing images, especially in detecting small targets, resulting in low average detection accuracy. In response to these problems, this paper uses the Faster Region Convolutional Neural Network algorithm as the basic algorithm. Furthermore, it combines the target characteristics in the remote sensing images to improve the basic algorithm. Finally, this paper proposes a new remote sensing image target detection algorithm. First, we use the Residual Network with more powerful feature extraction capabilities to replace the Visual Geometry Group network in the original algorithm. It can solve the shortcomings of the original algorithm’s insufficient feature extraction of the remote sensing images. The deep residual network adopts the identity mapping method, which not only ensures that the performance of the network will not degrade as the network deepens but also extracts deeper features. Second, we add a feature pyramid network to the algorithm to fully integrate feature maps of different scales. The feature map obtained in this way has high-level semantic and low-level detail information. Accordingly, it can take category and location information into account. This approach can greatly solve the difficult problem of large target scale differences in remote sensing images and improve the detection accuracy of small targets to a certain extent. In addition, we use the focal loss function to replace the cross entropy loss function in the original algorithm to solve the problem of the weight of the hard and easy samples to the total loss. Finally, given the problem that the used data set contains a small number of images, we use data augmentation to expand the dataset. This paper carries out two sets of comparative experiments to verify the effect of this algorithm. The first set of experiments is the ablation experiments on the NWPU VHR-10 dataset and RSOD-Dataset of the improved modules proposed in this paper. The second set of experiments is the comparison experiments of the algorithm in this paper and the other comparison algorithms on the NWPU VHR-10 dataset. The results of the first set of ablation experiments show that the various improved modules proposed in this paper can help improve the accuracy of target detection in remote sensing images. For the NWPU VHR-10 dataset, after adding the feature pyramid network, focal loss function, and data augmentation strategy, the algorithm in this paper improves mean Average Precision by 2.6%, 4.8%, and 0.8%, respectively. Furthermore, on the RSOD dataset, the algorithm in this paper improves the mean Average Precision by 0.6%, 1.6%, and 0.9%, respectively. Accordingly, the target detection accuracy rates of the algorithm in this paper can reach 93.4% and 93.0% on the NWPU VHR-10 dataset and RSOD-Dataset, respectively. The results of the second set of comparative experiments show that the target detection accuracy of the proposed algorithm is better than the comparison algorithm, further proving that the proposed algorithm has good performance in remote sensing image target detection. Finally, compared with BOW, COPD, RICNN, original Faster R-CNN, ODDP, and Mask R-CNN, the algorithm in this paper improves the mean Average Precision by 68.8%, 12.7%, 20.8%, 10.6%, 6.7%, and 9.5%, respectively. The remote sensing image target detection algorithm proposed in this paper can better solve the difficult problem of large differences in target scale in remote sensing images. It can improve the target detection accuracy of remote sensing images, especially the detection accuracy of small targets.
Abstract:Rapid and accurate identification of water bodies from remote sensing images is of great significance to water resources management and flood disaster monitoring. At present, traditional methods for identifying water bodies from satellite images still have shortcomings, and sometimes the results are not accurate enough to meet the practical needs. Recently, the Convolutional Neural Network (CNN) methods have emerged and been rapidly developed, providing a new idea for a identifying water bodies from satellite images. In this work, the Densely Connected Deep Convolutional Neural Network (DenseNet) is used to identify water bodies in the Hongze Lake area, together with the ResNet, VGG, HRNet networks, and the traditional method of Normalized Difference Water Index (NDWI). We have added the upsampling process and the skip connection structure to its classical structure to improve the performance of the DenseNet network. These methods are applied to the GF-1 satellite images of the Hongze Lake area to identify the water bodies in different seasons. Experiments are conducted to determine the optimal parameters of DenseNet, ResNet, VGG, and HRNet networks for water body identification. Moreover, the OSTU method is used to determine the optimal threshold of NDWI to reduce the uncertainty of threshold determination. Several indices of precision (P), recall (R), F1 score, and misclassification rate (MRate) are used to evaluate the performance of these methods. The main conclusions we have reached are as follows: (1) All the CNN models of ResNet, VGG, HRNet, and DenseNet have significantly outperformed the traditional NDWI method; for example, the precision (P) of water identification by using the NDWI method is only 0.779 compared with ground truth; however, it is highly improved to >0.922 by utilizing the CNN models. (2) The modified DenseNet model has effectively alleviated the problems of gradient explosion and disappearance, and the water body identification result is much better than the other CNN models, e.g., with the best P (0.960) and MRate (0.041). The training efficiency of the modified DenseNet model also appears far better than that of the other CNN models with the shorted training time, and the lowest loss function. (3) The modified DenseNet model shows also a better capability in identifying the fine features of water bodies, even if their shapes and water colors change largely in different seasons. These results have indicated that the CNN models are good tools for identifying water bodies from satellite images, and the modified DenseNet model appears to be the most promising one among them.
Keywords:satellite images;Water Identification;normalized difference water index;convolutional neural network
Abstract:With the progress of deep learning, researchers are increasingly paying attention to its application in hyperspectral image classification. Many experiments are conducted to achieve a trade-off between accuracy and efficiency to improve the feature extraction performance of neural networks toward small training sample sets.This work has proposed a high-speed and high-precision neural network structure based on spatial spectral information. A cascaded neural network for spectral spatial information extraction is constructed by combining the idea of DenseNet and adopting dilated convolutions instead of 3D convolutions as the main calculation method. The whole network structure is divided into four components: spectral information extraction, spectral compression, fusion of spatial and spectral information, and voting solution.Three convolutional layers are built in the spectral information extraction component. In each layer, 1×1×7 convolution kernels are used to extract spectral information and maintain the independence of spatial information. The number of kernels is set to 60. In light of the DenseNet idea, the network outputs of the first and second layers are dimensionally split in spectrum and inputted into the third layer. The outputs of the first, second, and third layers are also dimensionally split and inputted into the spectral compression component.In the spectral compression component, a 1×1×7 convolution kernel is used with a step size set to three. The spectral dimension is compressed, and the number of parameters of the deeper network is lessened by reducing the size of the feature map.In the spatial and spectral information fusion component, the goal is to fuse spatial information for the first time with 3×3 receptive fields and integrate the spectral information of the data. Separable convolutions are adopted instead of traditional 3D convolutions, and the 3×3×K convolution kernel is decomposed into a 3×3×1 convolution and a 1×1×K convolution. The value of K is equal to the spectral dimension of the input feature map. Then, 40 9×9×1 feature maps are outputted.Voting means that if the output of most pixels is the same value, then the average value of all values will also be pulled near this certain value. In the voting solution component using parameter-free global average pooling, the 9×9×1 feature maps are voted to obtain 1×1×1 output values. These 40 output values are spliced into the fully connected layer, and the classification results our outputted through Softmax.A series of experiments were carried out on the Indian Pains and Pavia University and Kennedy Space Center datasets. In the IP data set, the average accuracy reaches 95.0%, the overall accuracy 97.4%, and Kappa 0.97 by training with 5% data sets. In the UP data set, OA, AA, and Kappa reach 97.6%, 97.1%, and 0.97, respectively, by training with a 0.5% data set. The overall accuracy in the KSC data set can reach 99.2%. The network has been proven to strong feature extraction and classification ability.This method effectively improves the classification accuracy of hyperspectral images in the case of small sample sets and studies the effect of training and input data sizes on the classification accuracy. The classification accuracy of the network is improved with the increase in the training or input data. However, redundant information generated by a large amount of training data and excessive input data does not help improve the classification performance.
Abstract:Swidden agriculture is a widespread but controversial traditional land-use type in the tropics, especially in mountainous Laos with high percentage of forest cover. Driven by population growth, forestry policies, and climate change, swidden agriculture has been experiencing rapid evolution itself and drastic transformations into commercial plantations, such as rubber plantation. However, the remote sensing monitoring of tropical swidden agriculture has always been challenged, primarily because of the spatial and temporal dynamics in agricultural and forest cover, marginal feature compared with modern agriculture, and fragmentation and random distribution of swidden patches, hence with many unsettled issues and very limited information on its involved population, exact distribution and spatio-temporal dynamics. To explore the application potentials of machine learning algorithms in monitoring swidden agriculture, with two Landsat Operational Land Imager (OLI) images acquired in April, or the peak of the 2016 dry season, a support machine algorithm (SVM) was modified by masking out the information of construction land to improve the classification accuracy, or an overall accuracy of 95% and a Kappa coefficient of 0.81, followed by the examination of spatial (e.g., district-level) differences of freshly-opened swidden in Phongsaly Province, Laos, and their characteristics to local settlements and varied-level roads as well as topographical features. The results showed that: (1) Swidden agriculture remains an important land use type in Phongsaly because the newly-opened swidden was about 987.93 km2 (6.10% of the province) in 2016. More swidden patches were detected in the south and west parts of the province, with a fragmented distribution. (2) The area of newly-opened swiddens at district level ranged between 100—210 km2, with Samphanh District ranking the first (1/5) and Boonneua District the last (1/10). (3) Approximately 90% of newly-opened swiddens were concentrated within five km to residential points, particularly within four km. Similarly, these swiddens exhibited a distance decay pattern along the minor roads, tracks and major roads, in particular within a distance of five km of minor roads. (4) The newly-opened swiddens were mainly distributed in low mountain area (500—1000 m) with slope gradients of 15°—25° and aspects of southeast, showing slight variations among the districts of Phongsaly Province. This study provides reference for exploring machine learning algorithms in remote sensing monitoring of swidden agriculture in transition in the tropics.
Abstract:Glaciers are precious solid freshwater resource for humans. Since the 1990s, the glaciers in China have been in accelerating trend in melting due to global warming, so as to increase the sizes of surrounding ice lakes and form new ice lakes. This may result in geological disasters because glacial lake outburst have the characteristics of suddenness, great destructiveness, short duration, and wide distribution. Hence many research work has focused on the monitoring glacial lake changes.This study proposed an improved method to extract ice lakes based on the NDWI-NDSI combination. In this study, the NDSI (Normalized Difference Snow Index) was used to generate the land masks to separate the lands (foreground) and the mixed regions of land and ice (background). Then the threshold segmentation of masked NDWI was conducted in order to precisely extract the areas of glaciers. In the evaluation experiment, the ice lakes of the Bujiagangri glaciers located in the eastern section of Tanggula in southeastern Tibet were used as study region and 16 Landsat images covering the study area from 1988 to 2018 were used as test data. The experiment results indicated that the proposed method in this study can effectively extract the glaciers and reduce the misclassification compared with the methods using NDWI. By the proposed method, it can be found: (1) the area of the ice lakes has increased nearly 2 times (from about 2.7666 km² to about 5.2308 km²) due to the glacier evolution; (2) the annual increase rates of the areas of glacial lakes in this region is about 0.1230 km²/yr; (3) among the 12 glacial lakes, L-04, L-05 and L-10 glacial lakes have the largest areas (0.5—1.0 km²) and enlargement areas (0.5—1.0 km²). This indicated that the glaciers have greatly rapidly melted and severely retreated in the past 30 years, which may result in potential threats to the personal and property safety of downstream residents and Sichuan Tibet Highway (G317). Further investigations will be conducted to verify the usability of the proposed method in this study in other regions containing the glaciers with different geographic conditions.
Keywords:Glacier;global warming;glacial lake changes;Bujiagangri;glacial lake disaster;NDWI-NDSI
Abstract:Recently, Low-Rank Representation (LRR) has been widely used in hyperspectral remote sensing imagery classification. How to accurately classify ground objects by LRR has become a challenge in hyperspectral remote sensing research. The LRR based on elastic net (ENLRR) and the extended kernel version of ENLRR (KENLRR) are proposed to solve the above-mentioned problem. LRR classification method can make full use of the global information of the image. Its basic idea is to represent the whole test image by using the linear combination of as few training samples as possible, reconstructing the target image according to the representation coefficient matrix and training samples, and calculating the class of each pixel by the minimum reconstruction error criterion. The main idea of ENLRR is to introduce an elastic net into the LRR model, which replaces the rank function with the combination of nuclear and Frobenius norms of the coefficient matrix. To better classify nonlinear data, a modified KENLRR method is proposed by introducing kernel tricks in the ENLRR algorithm, and the neighborhood filter kernel function is adopted to map the original data into a high-dimensional feature space, which can obtain spatial-spectral joint information for better classification. In the experiments, three popular hyperspectral datasets are adopted, the proposed methods and the SVM, KNN, ELM, LRR, MFLRR, LSLRR, and KLRR comparison methods are used to carry out classification. Based on the experimental results, the proposed methods are effective in accurately distinguishing ground objects and have good stability and adaptability. In comparison with LRR method, the overall classification accuracies of ENLRR and KENLRR are improved by 4.55% and 6.74% in the Washington DC dataset, 14.22% and 23.30% in the Purdue Campus dataset, and 8.45% and 15.40% in the Gaofen-5 (GF-5) Yellow River Delta dataset. Therefore, the KENLRR method can provide the best performance for hyperspectral remote sensing imagery classification. The high-quality classification results provide technical support for analyzing the distribution pattern of ground objects, and prove the superiority of the proposed methods in hyperspectral remote sensing imagery classification。
Abstract:Rotation Forest (RoF), a powerful ensemble classifier, has obtained many successful applications in hyperspectral image classification. However, the data often has the problem of class imbalance. Consequently, the traditional RoF algorithm focuses on identifying the classes with majority samples, ignoring the accuracy of minority samples. The SMOTE (Synthetic Minority Oversampling Technique) algorithm increases the number of minority samples by simulating the way of generating new samples, thereby achieving the effect of balancing the categories of the data set. However, the SMOTE algorithm is mainly used in the data preprocessing stage and has the risk of increasing artificial noise when dealing with multi-class problems. Therefore, a novel dynamic ensemble algorithm based on SMOTE and RoF is proposed in this work to increase the classification accuracy of the multi-class imbalanced hyperspectral data. The proposed algorithm uses a dynamic sampling factor technology to merge the class distribution optimization with the base classifier. This algorithm not only realizes the adaptive generation of class balance data set but also reduces the influence of noise on the base classifier. In this experiment, three public hyperspectral images are used to test the performance of the algorithm, They are Indian Pines, Salinas and Pavia University. Four comparison algorithms are also selected, including random forest, traditional RoF, RoF algorithm with random oversampling, and SMOTE data preprocessing. The overall accuracy, average accuracy, F-measure, Gmean, minimum recall rate, ensemble classifier diversity, model training time, and McNemar test are the algorithm evaluation criteria. The experimental results demonstrate the effectiveness of the proposed method. The novel method not only obtains obvious classification advantages but also increases the recognition accuracy of minority samples while maintaining the overall classification accuracy of the data.
Abstract:Most proposed hyperspectral image band selection methods only consider the problem of band information redundancy and ignore the noise level of the selected bands. Accordingly, the representative band subset may contain high-noise bands, which is not conducive to subsequent semantic segmentation, image classification, and other applications. In response to this problem, this work proposes a noise-robust band selection method based on Pearson correlation coefficient, Information Entropy and Noise Level, referred to as PIENL.In the proposed PIENL method, the Pearson correlation coefficient is first used to calculate the correlation between the bands, and the band correlation matrix is constructed. Then, the spectral bands of the hyperspectral image are divided into several subspaces of the same size, and an optimal subspace division objective function adapted to the Pearson correlation coefficient is constructed to adjust the division points of the subspace. Finally, a new band information measurement criterion is proposed, which observes the band information entropy and noise level at the same time and uses the noise level as a penalty item in the objective function of the optimization problem. According to this criterion, the spectral band with high information entropy and low noise level in each subspace is selected as the representative band.Experiments were conducted on three public hyperspectral datasets of Indian Pines, Salinas, and Washington DC. Different band selection methods are evaluated using the average correlation degree of bands, classification accuracy, and the noise robustness. The experimental results show that this proposed PIENL method demonstrated outstanding band selection performance in terms of class separability, average correlation of representative bands, and noise robustness compared with the other advanced band selection methods.The PIENL method has strong robustness to noise and has achieved significant results on hyperspectral datasets containing noise bands. We can conclude that: (1) The similarity measurement method based on the Pearson correlation coefficient is more suitable for measuring the spectral difference between the noisy hyperspectral image bands compared with Euclidean distance; (2) Considering both information entropy and noise level to measure band information is helpful to select representative bands of hyperspectral image; (3) The representative bands selected by PIENL have better class separability. Compared with other advanced band selection methods, the overall accuracy of PIENL method is improved by 3%—13%, 1.5%—6.0% and 1%—6% respectively on the three datasets with high-noise bands removed. The overall accuracy is improved by 6%—11%, 2%—8% and 3%—7% respectively on the three datasets containing high-noise bands. This also shows that PIENL has better performance on hyperspectral images that contain high-noise bands.