Chlorophyll-a concentration is an important proxy for defining the tropic status of various bodies of water. Using remote sensing technology to retrieve chlorophyll-a concentration is an effective method for water eutrophication monitoring and a great number of algorithms for chlorophyll-a concentration retrieval are developed. These algorithms have different advantages and ranges of application. Because the optical characteristics vary in different bodies of water， it is hard to achieve desired results if blindly applying algorithms. In order to promote the further development of water quality remote sensing， the theory and data sources of remote sensing inversion are introduced.Then，domestic and foreign algorithms of retrieving chlorophyll-a concentration in water by remote sensing are summarized.The algorithms studied are categorized into six types by their architectural designs， namely： fluorescence peak and maximum peak algorithms， band algorithms， chlorophyll-a index algorithms，artificial intelligence algorithms，algorithm systems based on optical water types and analytical algori-thms.Each algorithm is presented systematically and its characteristics are analyzed.Then，all the aforementioned algorithms are compared regarding their applicable range of chlorophyll-a concentrations as well as water types.The applicability， merits and demerits of each category of algorithms are analyzed and concluded in order to provide reference for environmental and remote sensing researchers.The main conclusions are as follows：①the algorithm applicability for Case II waters is limited. More in-situ observations should be conducted to establish and supplement the database. Similarities and difference of various optical water types should be further studied to establish global algorithm systems based on optical water typologies； ②The combination of UAVs and hyperspectral sensors could provide new thoughts in monitoring inland water quality； ③Machine learning algorithms and mechanism models should be integrated to develop physical constrained models with high accuracy.
With the characteristics of high resolution， strong controllability and high cost performance， Unmanned Aerial Vehicle（UAV） remote sensing technology has been rapidly developed and applied in forest research. This study introduces the development of UAV remote sensing imaging platform， and the situation of forest investigation using UAV remote sensing technology at home and abroad. According to the investigation objects of individual tree and forest， the advanced methods of extracting forest parameters by UAV remote sensing technology are summarized. This study focuses on the analysis and discussion of various algorithms for obtaining forest parameters based on multispectral， hyperspectral and lidar sensors loaded in UAV platform， compares their advantages and limitations， and analyzes their best application scenarios. In addition， the application of UAV remote sensing in forest tree species classification and pest monitoring is also introduced. Finally， the application prospect of UAV in forest monitoring is discussed， which can provide theoretical basis and technical support for forest resources supervision in the future.
Monitoring and evaluating the state variables and fluxes of forest ecological processes on a long-term scale is one of the hot topics of current forest ecosystem observation research. In order to solve the problems of real-time transmission and storage of observation data of forest ecosystem observation stations， low data sharing， severe data fragmentation， weak construction of big data analysis platforms， and lack of real-time forest fire prediction and warning system， this paper used the remote sensing， eddy covariance technique， field investigation and Wireless Sensor Networks （WSN） technology to observe the spatial and temporal dynamics of carbon budget， water flux， and energy flux in Xiashu Forest Farm. We also developed a Data Visualization Platform Software （DVPS） to store and show the forest ecosystem observation data. Based on those data， DVPS can fulfill the monitoring of forest fires and the forecast and early warning of forest fire risks. DVPS also provides real-time display of carbon balance data， water flux data， energy flux data， and meteorological data between the forest ecosystem and the atmosphere， and the display of forest ecosystem service functions and its value. This experiment provides reference for the integrating of WSN， remote sensing and eddy correlation technology in long-term comprehensive observation experiments of forest ecosystems.
Rapid and accurate acquisition of forest structural parameters has been significant for forest resource investigation. In this study， photogrammetric and field-based tree height measurement of the Picea crassifolia were validated in the east and central of the Qilian Mountains. The individual segmentation algorithm using Canopy Height Model was applied to identify the position and height of the Picea crassifolia within each plot. The extraction accuracy of the average tree height was recognized the highest among the four indexes of maximum value， minimum value， mean value and standard deviation， with Root Mean Square Error （RMSE） values of 1.39 m and R2 values of 0.93（P<0.05）. Tree heights extracted from LiDAR data of Picea crassifolia were used to analyze the spatial distribution of tree height in the Qilian Mountains. There was a downward trend of the average forest canopy height from east to west in the Qilian Mountains. As the altitude rises， the forest canopy height showed a “unimodal” change， which peaked at change an altitude of 2 900 m. This study shown that UAV photogrammetric tree height measurements was a viable option for intensive forest monitoring plots. Additionally， it was shown that underestimated evident in field-based and UAV laser scanning tree height measurements could potentially lead to misinterpretation of results when field-based measurements are used as validation.
Muli County and its surrounding areas in Sichuan Province are one of the most frequent and vulnerable forest fire areas in the China. In the past two years， there have been serious casualties of firefighters in forest fire fighting. In this paper， time-series and multi-source satellites images， UAV images and disaster site survey data are used to monitor the forest fire hot spot dynamically， and the process of fire spread in Muli County were analyzed. The results show that： The GF-4 Satellite images and 2 m/8 m optical satellite constellation images， which can be used for forest fire hotspots monitoring effetely. We believe that the grid can be determined as the forest fire hotspot， while the grid temperature value T≥360 K in the daytime or temperature value T≥330 K in the nighttime. 25 forest fire incidents in 6 fire sites were found and monitor from March 30 to April 6 in the area， and the progress of the fire in Muli and Xichang County were also investigated with the hot spot. Compared with the high-resolution unmanned aerial vehicle images and the collecting hot spot data of fire site， it shows that the accuracy of satellite forest fire hotspot monitoring can reach 89%. It is recommended to use time-series and multi-source satellite images to monitor the forest fires in this area continuously， and combine the fire site survey data of authoritative departments for timely fire warning. It is also suggesting that combine meteorological， topographic， vegetation and human factors to carry out fire cause analysis and risk assessment studies， to avoid causing loss of life and property again.
Fuzhou University was taken as an experimental area， this paper presented a fast extraction method of lodging landscape trees in typhoon disaster based on unmanned aerial vehicle remote sensing image， which can provide reference for the assessment of typhoon disaster losses and post-disaster reconstruction of the landscape department. Firstly， unmanned aerial vehicle remote sensing technology was used to obtain Pre and Post images during typhoon passing with a resolution higher than 10cm. After processing， Digital Orthophoto Map（DOM）and Digital Surface Model（DSM）were obtained. Then gaussian high pass filtering algorithm was used to highlight the edge information of tree trunk. And the best feature subset was selected by contrast filtering segmentation algorithm combined with maximum Relevance Minimum Redundancy（mRMR）feature selection algorithm. In addition， the tree trunk and non-tree trunk were detected according to the threshold value and Random Forest（RF）classification method respectively. At last， the tree trunk of lodging tree was simplified into skeleton line by using skeletonization algorithm， and the single tree trunk was extracted by using octo-neighborhood tracking method. The results show that a total of 71 lodging trees were detected using threshold classification based on single-phase UVA images， with an accuracy of 76.06% in the experimental area. However， the accuracy of lodging tree extraction improved by 12.73% based on RF classification，and the missed detection reached 25.39%. In order to compare the detection efficiency of lodging trees based on single-phase and two-phase images， combined the difference value of DSM in the two phases， threshold and RF classification were used respectively， with an accuracy of 89.66% and 87.30%， a commission of 17.46% and 12.70%．Research suggests that the single-phase image features can basically detect the lodging trees， and the multi-phase images analysis can effectively improve the detection accuracy of the lodging trees， providing an effective reference for the detection of the lodging trees under different data sources. According to the research， the UAV remote sensing technology can realize the rapid estimation of the number of lodging trees after typhoon.
Forests play an important role in ecosystem services， such as providing clean air， protecting biological habitats， and reducing global greenhouse gas emissions. Global Forest Change （GFC） maps the forest cover change with a high spatial resolution of 30 m every year， which has become an effective tool for monitoring the spatio-temporal changes of forest cover. Based on Google Earth Engine， using high-resolution global forest change data， we inspected the patterns and processes of deforestation combined with linear regression methods and spatial autocorrelation theory in Southwest China. The results show that in the past 19 years， the total forest loss in southwest China was 3.752 7 million ha. 2008 is a turning point， as the area of annual forest loss has a significant rising trend before（P<0.05）， and a fluctuant decreasing trend afterwards. The losses were mainly distributed in Guangxi province， southeast Guizhou Province and southern Yunnan Province. While forest loss are mainly located on the mountains（elevation below 2 000 m， slope below 19°）， the location of forest loss patches has moved to lower and flatter areas in recent years. From a spatial perspective， the Moran’s I index was positive， with an average value of 0.406 from 2001 to 2019.It also tells that most prefecture-level cities are neighboring by the cities with similar forest loss area. High-High clusters are mainly in Guangxi Province and southern Guizhou Province， and the aggregation of Low-Low are showed in Chongqing， Sichuan Province， and northern Yunnan Province. The potential factors that affect deforestation in Southwest China highlight the role of forestry activities and agricultural expansion， policy factors play an important role in the transformation of land use patterns. This study provides a basis for the future rational forest management and planning of forest resources， formulating more scientific and effective policies and programs.
Soil moisture is an important index in soil monitoring， which has an important impact on agricultural production， ecological environment and water resources management. With the remote sensing modeling and remote sensing inversion theory have gradually become important techniques and means to estimate soil indicators. Therefore， using the optical image data and radar image data， with Mohe City of Daxing'anling area as research area， to establish model of soil moisture inversion based on Landsat8 data and the model based on Landsat8 image data and high-resolution 3 remote sensing image data， the inversion results compared with the measured data analysis， and make evaluation on the model. The results showed that： （1） The surface temperature in the study area was inverted， and the TS-NDMI feature space was constructed by using surface temperature （Ts） and normalized difference humidity index NDMI. Combined with the measured data， it could be found that the inversion results of ts-NDMI feature space soil water inversion model were negatively correlated with the measured soil water content；（2） The soil moisture retrieval model based on GF-3 satellite data and Landsat 8 remote sensing data can get better retrieval results， and in areas with high vegetation coverage， the results obtained from this model are more accurate than those from a single optical data source， which provides a new way for the study of soil moisture in high vegetation coverage areas.
Precipitation dataset with high resolution are essential for accurate hydrology predictions and meteorology simulations over complex terrains. A regression model was built to downscale the Global Precipitation Measurement （GPM） IMERG precipitation data from 0.1° to 1 km on an annual scale， using vegetation， topography and geographical location features over the Tibetan Plateau. Then monthly precipitation data were obtained by disaggregating the annual downscaled estimates， which were calibrated with observations of local rain gauge stations. The major conclusions are summarized as follows： （1） Monthly GPM IMERG precipitation demonstrated good agreement with the rain gauge data during the period 2015 to 2017 （R2=0.79）， though GPM was slightly larger than ground observations； （2） Annual downscaled precipitation improved the spatial resolution of the GPM IMERG in the study area； （3） Monthly donscaled precipitation calibrated with rain gauge data reflected detailed characteristics with better predictive performance especially in summer or in wet regions.We concluded that the model can be used to obtain precipitation data with high spatial resolution from heavy rain to light one over the areas with complex tography， which is meaning for applications in hydrology and metorology studies.
The main function of radar altimeter is to measure the global average Sea Surface Height （SSH）. It can accurately measure the satellite ground distance by measuring the time interval between sending and receiving pulses. The accuracy of time measurement depends on the accuracy of altimeter clock. During the operation of satellite in-orbit， the frequency of radar altitude clock will drift slowly， which will affect the accuracy of satellite ground distance measurement and result in the measurement deviation of sea surface elevation. Based on the in-orbit test method of altimeter clock deviation of HY-2A satellite based on reconstruction transponder， this paper proposes an estimation method of altimeter clock deviation extracted from the time offset of satellite ground offset function curve， which is applied to the in-orbit calibration test of HY-2B satellite radar altimeter. The frequency drift of radar altimeter is measured， and the accuracy is better than 0.001 Hz. The results show that the performance of the atomic clock of HY-2B altimeter is stable， the range error caused by clock frequency deviation is millimeter magnitude， and the average range drift rate is 2.95×10-7 m/d.
The high-dimensional characteristics of the hyperspectral image and the high correlation between the bands have led to the problem of large data volume and information redundancy in the study of the feature recognition of hyperspectral images， which reduces the classification and recognition accuracy of hyperspectral images. Aiming at the above problems， a hyperspectral image classification method based on Local Fisher Discriminant Analysis （LFDA） combined with Genetic Algorithm （GA） to optimize Extreme Learning Machine （ELM） is proposed. First， the LFDA is used to reduce the dimensionality of the hyperspectral image data to eliminate information redundancy and retain the main features in the local neighborhood； then use GA to optimize the ELM， classify the feature samples after the dimensionality reduction， and improve the classification and recognition of the hyperspectral image Precision. The method proposed in this paper is applied to the research on the feature recognition of hyperspectral images in Salinas and Pavia University. The classification accuracy reaches 98.56% and 97.11% respectively， which verifies the effectiveness of the method in this paper.
A spaceborne differential absorption barometric pressure radar operating in the 65~70 GHz strong oxygen absorption band is proposed to continuously acquire sea surface pressure data with global high temporal and spatial resolution. Through the analysis of the design requirements of the spaceborne differential absorption barometric radar system， the atmospheric profile data and the atmospheric absorption coefficient model are used to simulate and analyze the performance of the sea surface pressure differential absorption. The simulation results show that there is a linear relationship between the sea surface pressure in the strong oxygen absorption band and the differential absorption index. The RMSE of the sea surface pressure estimation obtained by the spaceborne differential absorption barometric pressure radar under clear sky at operating frequencies of 66 GHz and 69 GHz. The error is 2.6 mbar， and the rms error of the sea surface pressure estimation obtained under different cloud conditions is 3 to 4 mbar， which provides a reference and basis for the design and engineering implementation of the subsequent radar system.
Study on the spatiotemporal changes of global water storage is of great significance for understanding the global water cycle， arranging agricultural production， and preventing natural disasters. GRACE satellite provides direct observation for obtaining changes in water storage on a global scale. However， different solution models and methods lead to differences between the water storage products. To reduce the uncertainty in GRACE data， this paper combines the Global Hydrological Models （GHMs） and Land Surface Models （LSMs） data， and uses the Three-Cornered Hat （TCH） method to analyze the uncertainty of the existing products. On this basis， the idea of point surface fusion is introduced， using GRACE and model simulation data to select training points， and using machine learning methods to correct the accuracy of GRACE satellite data. This article takes California， USA as an example， and uses TCH， long-term quantitative analysis and groundwater well site data to verify the accuracy of the result. The results show that： （1） The correction results have lower uncertainty than the original GRACE data in the uncertainty analysis （GRACE CSR： 25.32 cm， PCR-GLOBWB： 33.10 cm， DBN： 13.85 cm）； （2） In long time series analysis， the correction results are smoother than the original data， reducing abnormal fluctuations； （3） In the verification of the well site， the results show that results have improved in correlation， root mean square error， and average absolute error.
Accurate identification and division of urban functional zones play an important role in rational planning of urban development and solving urban problems. Remote sensing images have rich spectral texture features， but it is difficult to characterize the social and economic attributes of buildings， while urban data such as social media data provide rich data resources for urban research and application， and supplement the internal characteristics of buildings missing from remote sensing images. In this study， multi-feature information of high-resolution remote sensing image and POI data is integrated and embedded topic model is used to mine its potential semantic information to identify urban functional areas. Three experiments were designed with two typical urban business districts in Ningbo as the study area to verify the effect and performance of the research method. The results show that this method can achieve 85.67% and 85.78% classification accuracy， and can accurately identify urban functional zones. At the same time， the multi-feature information of spectral， texture， geometry and POI feature combination can significantly improve the identification accuracy of urban functional zones， and the embedded topic model can mine the high-level potential semantic information of multi-features better than the three mainstream topic models of pLSA， LDA and STM.
Image fusion plays an important role in the remote sensing monitoring of coastal wetlands. However， the coastal wetlands are mainly composed of water， plants and so on， which leads to spectral distortion， and poor robustness of existing fusion methods. Therefore， in order to meet the demand of robust panchromatic / multispectral image fusion method for coastal wetland remote sensing monitoring， we propose a panchromatic / multispectral image fusion method （SSQI_PANSHARP） based on spatial-spectral quality evaluation， to integrate the advantage of different fusion methods. In this study， Hangzhou Bay coastal wetland was taken as the research area. Based on the 20 year image data of Landsat-7 and Landsat-8 from 1999 to 2018， the proposed method was fully verified from qualitative and quantitative aspects， as well as NDVI and NDWI. On this basis， the data of three periods in the same quarter are extracted evenly， while the land cover types of Hangzhou Bay Coastal Wetland in recent 20 years were monitored and analyzed. The results show that the SSQI_PANSHARP method has better spectral fidelity， and spatial structure information enhancement effect in the coastal wetland area， and strong robustness， which is suitable for coastal wetland monitoring needs. In addition， long-term dynamic monitoring showed that the proportion of construction land increased after 2010， the pond paddy field increased significantly， and the farmland green space decreased.
At present， the quantification of debris flow material sources is mainly depended on field survey， which is time-consuming， with limited spatial coverage and strong subjectivity. Comparatively， remote sensing-based detection method provides a more reliable way for extracting areas of debris flow material sources because of its characteristics of frequent observation， large scale coverage and high precision. In this study， we developed an object-oriented classification method to extract the source area based on Sentinel 2 image and ALOS digital elevation model data， according to the spectral and topographic characteristics of the source area. Compared with visual interpretation method， this method was automatically conducted and can identify the type difference of the material sources. Take the Shuzheng village basin as a case study， the method precisely extracted the three key sources for debris flow （slump-mass sources， gully sediments sources and slope wash sources） before and after the earthquake. The results show that： （1） Based on the validation sample points collected from high-resolution images of UAV and Google Earth， the material sources extraction accuracy of the proposed method is 85.71% before the earthquake and 88.34% after the earthquake， and the corresponding Kappa coefficients are 0.77 and 0.76 respectively. （2） Compared with the pixel-based remote sensing classification method， the accuracy of the proposed method before and after the earthquake is 14.28% and 22.70% higher， and it has a better performance， especially for the recognition of small areas of slump-mass. （3） Before and after the earthquake， due to disasters such as collapses and landslides， the total source reserves increased from 1.85 million m3 to 3.99 million m3. The main source type is the slump-mass source， accounting for 70.80%. In general， this study provides a semi-automatic extraction method based on high-resolution remote sensing image for the extraction of debris flow sources， which will provide important scientific support for debris flow disaster prevention and risk assessment.
By extracting and preprocessing ASTER remote sensing data of Salamat Basin and Doseo Basin， analyzing anomaly minerals information of clay， carbonate and ferrum， it carried out the remote sensing interpretation of structure and surface geological features， acquired hydrocarbon remote sensing anomaly area and conducted the petroleum geological evaluation. The results of surface structures and geological interpretation indicate that tectonic characteristic of the Central African Rift System is a shear zone while Salamat Basin and Doseo Basin are pull-apart basins that formed when the shear faulted. Salamat Basin and southern Doseo Basin are distributed with fractures and both sides of the fault have development of echelon fault or fold obliquely crossing with the main fracture. The main shear zone in the macro-view image appears converse "S" to the north east. Comparing with the known oil well， it shows that clay mineral alteration anomaly at the well is obvious， followed by carbonate alteration anomaly， and the iron ion alteration anomaly is weakest. Through analyzing on remote sensing area of Salamat and Doseo Basin， the Level 1 anomaly areas are mostly distributed in the fracture development zone， showing that seepage of oil and gas is controlled by fracture， and the fracture zone provides favorable channel for the seepage， which makes the ground oil and gas anomaly information obvious. The research has extracted alteration anomaly area which provides reference for further exploration.
Meeting the expectation of ship targets detection on sea surface， considering the need of wide swath and high resolution of azimuth， a method for detecting ship targets on sea surface based on wide swath spaceborne scatterometer is proposed. In this method， the scatterometer uses a sector beam rotating scanning antenna， transmits wideband LFM signal， carries out azimuth high-resolution processing and two-dimensional ship target detection on sea surface， and finally the ship target observation data in multi azimuth are obtained. Using observation data， such as azimuth， Doppler frequency and one dimensional high resolution range image， ship length， ship speed and the included angle with satellite movement direction are calculated， ship type is preliminarily and recognition， the function of auxiliary detection is realized. The simulation results show that this method can detect the ships on the sea， get the effective ship information， and expand the ability of auxiliary detection.Furthermore，we present suggestions for the future research to improve the ability of Ship Target Detection and Recognition.
With the rapid growth of urbanization， the heat island effect has become a hot issue in today’s society. Most of the existing researches focus on the analysis of heat island intensity change and landscape pattern influence. In fact， the extraction results of remote sensing heat island are affected by many factors. Image selection is particularly critical. Taking Shijiazhuang as an example， Landsat （different seasons and different vegetation status in the same season） and ASTER night images were used in this paper. These data were used to analyze the impacts of season， farmland growth state， daytime and night and other factors on the extraction results of remote sensing heat island. Research show that， in the seasons with flourish crops growth， high average temperature， remote sensing heat island extraction effect is better and the intensity of heat island is larger. After the harvest of farmland crops， the land is bare. To ensure the quality of remote sensing heat island extraction， night data should be selected. The experimental results can provide reference for the selection and analysis of remote sensing data in the process of heat island research.
Accurate radiometric calibration is the fundamental of quantitative remote sensing. In this work， the Microwave Radiation Imager （MWRI） on the Chinese meteorological satellite Fengyun 3C （FY-3C） is intercalibrated against the Microwave Imager on the Global Precipitation Measurement （GMI） using the double difference method. First， the FY-3C MWRI data， GMI data and the fifth edition of the European Centre for Medium-Range Weather Forecast Re-Analysis （ERA5） data are resampled into a 1°×1° regular grid space. Then， matching observations are collected according to matching criteria， and simulations in both FY-3C MWRI and GMI channels at top-of-atmosphere are calculated using the ocean microwave radiative transfer model. Next， the double differences and theoretic observations in FY-3C MWRI channels are computed. Finally， the intercalibration coefficients are determined， and the FY-3C MWRI data are re-calibrated. The results show that， against GMI， the observations in FY-3C MWRI channels are underestimated， especially for the low frequency channels， and the calibration bias decreases with the frequency increment. The calibration biases of FY-3C MWRI ascending （MWRIA） data are 1.0 K~2.0 K lower than that of FY-3C MWRI descending （MWRID） data. At the standard scene brightness temperatures defined by the Global Space-based Inter-Calibration System （GSICS）， in 10V/H， 18V/H， 23V， 36V/H and 89V/H channels， the calibration errors of MWRIA are -6.7±0.3 K，-8.7±0.7 K，-2.9±0.7 K，-2.0±0.8 K，-2.4±0.7 K，-4.0±0.8 K，-2.4±1.4 K，-1.3±1.0 K and -0.4±1.8 K， respectively； the calibration errors of MWRIA are 7.9±0.7 K，-9.7±0.9 K，-4.3±0.9 K， -3.0±0.8 K，-3.5±0.9 K，-5.1±0.8 K，-3.0±1.1 K，-2.4±0.6 K and -1.0±2.1 K， respectively.
The surface roughness reflects the small fluctuations of the surface and is an important surface parameter in the microwave remote sensing inversion of soil moisture. In this paper， the Lightning River Basin is used as the research area. First， the surface roughness of different ground objects is measured using the pin-plate method， and then a series of processing such as perspective transformation， digitization， tilt correction and period correction are performed on the measured data. Surface roughness results. Studies have shown that when measuring the surface roughness using the pin plate method， in order to correct the influence of the measurement attitude of different sections， tilt correction is required to reduce the calculation bias， and the ground surface with periodic ridge and ridge structure such as carrot field and cauliflower field needs further Perform period correction. By analyzing the surface roughness of typical objects in the Lightning River Basin， it is found that the surface roughness of the grassland in this area is small， the surface roughness of the crop area is generally large， and the surface roughness of various types of objects is from small to large They are grassland， cauliflower field， corn field， carrot field， potato （harvest） field， and potato （unreceived） field. Finally， the correlation analysis between the surface roughness and the airborne microwave radiation and scattering observations was conducted， and it was found that there was no obvious relationship between the surface roughness measured at the ground single point and the airborne scale microwave radiation scattering characteristics.
High spatial resolution satellite remote sensing provides redundant data for large-scale agricultural land management. Due to the variety and complexity of land parcels， it is still a great challenge to accurately and timely map land use types. Introducing parcel boundary has proven an effective strategy to integrate spatial and spectral features， to improve the classification accuracy. However， current feature extraction methods always treat each parcel as a whole and use only mean feature values of all pixels inside each parcel. This approach cannot well adapt to the scenarios that parcels include more than one types of spectral similar target. To this end， this paper proposes a spectral clustering based feature extraction method to better model the complexity of parcels. BJ-2 images and ground surveying data from 2 typical areas in the Chongming County in Shanghai were selected to experimentally evaluate the proposed method. The results show that： （1） Compared with the direct spectral averaging method， the proposed method can effectively improve the accuracy of land parcel classification； （2） Introducing the clustering features into the typical feature combination can further improve the accuracy of land parcel classification. And the improvement mainly lies on categories with unstable mixing ratio of internal pixel spectrum， such as vegetable field and corridor. The proposed method provides an effective alternative to classify parcel types especially for parcels including more than one spectral similar target.
