The geomorphic characteristics of "thousands of gullies" in the Loess Plateau show significant self similarity in multi-scale space, and have obvious textural characteristics of local-irregular and macro-regular. Previous studies have shown that there have been specific research results on the selection of texture features, the uncertainty of scale effect, and the combination of texture features with other features in identification and classification of specific landforms. However, the current texture analysis methods are limited to the application of macro terrain classification. For the concept, classification, basic characteristics, and analysis methods of terrain texture, there is a lack of theoretical framework for application support. On the basis of the existing research results, this paper defines the Loess Plateau as the research scope, and puts forward the concept model of the Loess Plateau terrain texture, namely definition, characteristics, classification, and expression. In terms of the definition of terrain texture, this paper expands the scope of the definition. In addition to the existing macro morphological topographic texture, the terrain texture formed by the combination of the characteristics of typical loess geomorphic units (loess yuan, liang, mao, etc.) and the terrain texture formed by the slope characteristics of loess slope are proposed. This paper points out that the data expression based on Digital Elevation Model (DEM) will be more conducive to the quantification of terrain texture, especially the terrain factor derived from DEM can expand the feature space of terrain texture and enrich the data source of terrain texture analysis. In terms of the basic features of terrain texture, this paper puts forward three basic characteristics: regional difference, genetic complexity, and scale dependence. Among them, regional differences can be qualitatively distinguished by visualization or quantified by existing statistical methods, so as to effectively distinguish differences in texture between regions. In the classification system of terrain texture, this paper classifies the terrain texture based on its element saliency, texture origin, and visual form. Taking loess liang in the loess hilly and gully region as an example, a single loess liang can be regarded as a texture element. Through a certain arrangement and combination of several loess liang, the terrain textural characteristics of the loess liang hilly and gully region are formed. However, a single loess liang cannot express the texture features. This paper aims to build a conceptual model of terrain texture oriented to the Loess Plateau, and promotes the application and development of texture analysis method in Loess Plateau.
Spatial co-location mode analysis has been commonly used in data mining, which can be used to characterize the correlation between different urban service facilities and further quantify the distribution pattern of urban service industry. In this paper, a co-location pattern mining method with POI data is proposed to obtain the spatial correlation of urban service industry. Firstly, through the acquisition of neighboring instances, and selection and storage of homology candidate patterns, the second-order homology pattern of urban service industry can be obtained; Secondly, the industrial spatial correlation map is constructed to obtain the correlation structure between industries. Finally, the industrial spatial correlation graph density and spatial correlation significance are constructed to measure the tightness of urban service industry relationship. We select Chengdu, Lanzhou, Zhengzhou, Shenyang, Shanghai, and Shenzhen as experimental areas. The results show that there are both similarities and differences in the spatial correlation of urban service industry in different cities. Generally, service industries such as catering and shopping which are related to daily life have a strong spatial correlation with other service industries. These types of service industry are often spatially clustered. The administrative department has a weak spatial correlation with other service industries and often occupies a separate functional area. Based on the results of the co-location pattern mining for each city, we find that the co-location pattern between teahouses and residential areas is strong in Chengdu, which indicates a unique "tea culture". In Shanghai, foreign restaurants and leisure places show a co-location pattern, which indicates the internationalization characteristic of Shanghai. Both Chengdu and Shenyang show the strongest spatial correlation of service industry which is highly mixed. The spatial correlation of service industry in Lanzhou is moderate. While Shanghai and Shenzhen show the weakest spatial correlation of service industry. These two cities have a high-level economic development and show separated industrial zones. Zhengzhou also has a weak spatial correlation because of its "multi-center, group-like" structure. This paper uses the spatial co-location model to characterize the spatial correlation of the urban service industry, which can be used as references for future urban planning.
The spatial and temporal characteristics of residents' medical treatment reflect the service capacity and layout rationality of medical facilities. This study investigated the features and patterns of medical treatment using taxi trajectory data in Xiamen. We divided Xiamen Island into different research units based upon the central lines of roads. We presented a trajectory drift algorithm to extract the medical treatment OD data for tertiary hospitals. This algorithm deals with the positional error associated with trajectory data and can improve the extraction accuracy. Hospitalizing behavior was analyzed from the perspective of space and time. Finally, based on the residents' preference for hospitals, we discussed the spatial patterns of residents' medical treatment by K-means algorithm. The results show that: (1) Compared with traditional buffer analysis, the trajectory drift algorithm didn't require high positioning accuracy when extracting OD data for hospitals. OD data can be extracted more reasonably and completely only by shifting OD point's coordinates, with an accuracy increased by more than 30%. It was also applicable to all floating vehicle trajectory data; (2) The peak time of medical treatment occurred at 7 am and 2 pm, respectively. The number of medical visits was twice on weekends (including holidays) than working days. When the travel distance was greater than 1 km, the number of medical visits decreased with the increase of travel distance, following a Weibull function distribution; (3) Residents regarded the Zhongshan, the First Affiliated, and the Chinese Medicine Hospital as their first choice for medical treatment. There was a significant regional difference in choices of medical treatment, that is residents preferred nearby hospitals. The southwest of Xiamen Island had sufficient medical resources, and residents' average medical travel distance was less than 4 km. However, residents in northwest and southeast of Xiamen Island mostly had to travel about 10 km for medical treatment. The medical resources in these regions were relatively scarce and needed to be strengthened eagerly; (4) The service capacity of the nine tertiary hospitals in Xiamen Island was obviously different. The residents had a strong preference for the Zhongshan, the First Affiliated, and the Chinese Medicine hospitals, with evaluated preference values greater than 33%. The service scopes of these three hospitals basically covered the whole Xiamen Island, which indicated strong attraction and service capacity for the residents. The values of residents' preference for the other six hospitals ranged from 0 to 33%. These six hospitals mainly treated nearby residents, leading to weak attraction and service capacity. This study provides alternative methods to extract the spatiotemporal features of residents' medical treatment and supports the decision-making of optimizing the spatial configuration of medical facilities.
With the development of network technology, the analysis of internet public opinion plays an increasingly important role in dealing with the emergency. After the occurrence of natural disasters, it is helpful for the emergency management department to take effective emergency rescue measures in time to accurately grasp the characteristics of public opinion information and analyze its influencing factors. Based on the network public opinion data related to Typhoon Lekima, including micro-blog, WeChat, forums, websites, and other online public opinion data collected by the "Public opinion on Sina" system, this article analyzes the spatiotemporal characteristics of disaster public sentiment in the process of disaster. The influencing factors of the disaster public opinion information are also analyzed. The results show that the temporal distribution of public opinion information is consistent with the lifecycle of Typhoon Lekima. Compared with the grey EGM (1,1) model, ARIMA model has a higher applicability for short-term prediction of public opinion. The spatial distribution of public opinion is positively related to the severity of the disaster and also related to the economic condition and the network popularity in the affected area. The correlation between the severity of the disaster and the original public opinion information is stronger than that between the severity of the disaster and the transmitted public opinion information. The original public opinion information can better reflect the actual situation of affected areas. The study provides guidance for emergency departments to grasp the trend of public opinion and adjust emergency measures timely.
The natural environment in Northeast China has attracted a wide attention from the public. Land environmental changes such as land desertification, woodland grassland degradation, forest degradation, and land salinization have been the focus of scientific researches. Based on the propose of the "Beautiful China Mid-ridge Belt", researches on land environment near the northeast section of the line are continuing. In these researches, commonly used research methods include field survey and remote sensing images analysis. The knowledge graph technology is a time-saving and labor-saving method to explore the associations between data. So this paper aims to use the method of scientific knowledge map to excavate the current literature (4318 CNKI documents) on the topic of influencing factors of land environment change in Northeast China. The co-occurrence method is used to extract the target factors in literatures, in order to construct a more comprehensive scientific knowledge map. According to the knowledge map, we can summarize the main factors that affect land environmental changes in Northeast China and perform quantitative analysis. Finally we carry out a statistical analysis and select the top ten factors by word frequency that cause environmental changes, such as land desertification, salinization, grassland degradation, woodland degradation, forest degradation, and black land degradation. For example, for the factors that affect land desertification, the keywords with the highest frequency are ecological environment destruction, soil erosion, sandstorms, etc. For the factors that affect land desertification, soil erosion, soil desertification, and water shortages have the highest frequency. This method will provide new ideas for exploring the influencing factors of geography, climate, and other phenomena from the perspective of knowledge map and text mining in the future.
The construction of new urbanization puts forward practical requirements for a suitable human settlement environment in the new period. By using GIS technology, based on multi-source data such as Nighttime Light (NTL) data, traffic network vectors data, Points of Interest (POI) data, and statistical yearbooks, we selected the economic level, transportation accessibility, historical culture, and public service as impact factors of the suitability of human settlement environment (weights were 0.36, 0.27, 0.17, 0.20, respectively) on the basis of 500 m × 500 m grid unit. We quantitatively evaluated the human suitability of human settlement environment in Anhui Province in 2017 using a synthetical index method to construct the human suitability evaluation model. The results show that: (1) the human suitability index of human settlements in Anhui Province ranged from 0.83 to 87.10. The human settlement environment could be divided into five types: highly suitable areas, relatively suitable areas, moderately suitable areas, general suitable areas, and critical suitable areas. The area of moderately suitable areas was the largest, accounting for 68.72% of the whole province, while the area of highly suitable areas was the smallest, accounting for only 1.24% of the whole province. The spatial heterogeneity of human settlement suitability was characterized by "Multi-core" and "striped" patterns; (2) the transport accessibility and public service were the main factors that led to difference in human suitability of human settlements in the province, with an average index of 94.18 in the moderate suitable areas, and an average contribution rate of 34% in all types of regions. Besides, the historical culture had a significant impact on higher and critically suitable areas, with an average contribution rates of 10.51% and 10.53%, respectively, while the economic level had the most significant contribution to the highly suitable areas, with an average contribution rate of 22.02%; and (3) nearly 90.86% of the population in the province was concentrated in the regions with human suitability index of 43.00~66.00 (i.e., moderate to the high suitable areas), which implied that the human quality of human settlements matches the population distribution. In conclusion, our evaluation results objectively reflect the baseline of the human settlements in Anhui Province.
The exposure of spring is usually difficult to be monitored over mountainous terrain. In this study we investigated the performance of statistical models (Weight of Evidence) and two machine learning models (Random Forest and Classification and Regression Tree) in predicting the potential exposure positions of spring water in Beijing. A total of 1378 springs from field survey were used for model training and validation. The environmental factors included elevation, slope, aspect, topographic wetness index, stream power index, distance to rivers, distance to faults, lithology, normalized difference vegetation index, and land use. The predicted results from the three models are validated using the receiver operating characteristics curve. The area under the curve for the Weight of Evidence model was 0.80, while that for Classification and Regression Tree and Random Forest the AUC was 0.81 and 0.86, respectively. Therefore, the Random Forest model has the best prediction performance. Moreover, the Random Forest model revealed that lithology, distance to faults, and distance to rivers had the greatest impact on the spring exposure. This study shows that the machine learning method has good prediction ability and is expected to be applied in future spring protection and restoration researches.
Drought is one of the most frequent and widespread climate extremes, causing devasting social, economic and ecological damages. It is of key importance to evaluate drought reliably and effectively. In this study, in order to assess global drought grade, the Random Forest (RF) algorithm was used to establish the drought grade assessment models for the 11 climate zones in the world. We chose monthly mean precipitation, mean temperature, maximum temperature, minimum temperature, soil moisture, evapotranspiration (ET), Normalized Difference Vegetation Index (NDVI), and Sun/Solar-induced Chlorophyll Fluorescence (SIF) as explanatory variables and drought grades based on Standardized Precipitation Index (SPI) as target variable. The SPI on different timescales of 1 month, 3 months, 6 months and 12 months were labeled as SPI1, SPI3, SPI6 and SPI12, respectively. The data from 2007 to 2012 were used as training data of the assessment models while those from 2013 to 2014 were used as prediction data. The results showed that: (1) The temporal scale of SPI influenced the model accuracy. Among the models with drought grade based on SPI1、SPI3、SPI6 and SPI12, the one with drought grade based on SPI1 had the highest accuracy (60%~75%) and prediction performance. The model with drought grade based on SPI1 was able to capture 90.91% of the drought records in the global emergency events database (EM-DAT). It could capture 78.47% of the drought duration month in the EM-DAT. The agreements with records and drought duration month in the EM-DAT indicated the good performance of the drought grade assessment model based on 1-month SPI and RF algorithm. (2) The drought grading criterion had little impact on the model performance. Users could select criterion I (drought/not drought) or criterion II (severe/not severe) depending on the real needs. (3) The relative importance of each explanatory variable depended on both the temporal scale of SPI and climatic differences. Precipitation was the most important factor for the drought grade based on SPI1. The importance of precipitation decreased and the ones of other explanatory variables such as temperature, soil moisture, NDVI, and ET increased as the timescale of SPI increased. The importance of variables except precipitation showed differences in different climate zones. Among the tropical, subfrigid, and tundra climate zones, temperature or ET is relatively important for drought. Soil moisture is relatively important in dry climate zone and precipitation is the most important in mild temperate climate zone, while vegetation is relatively important in the humid continental climate zone.
Given that the automatic extraction of features from high resolution remote sensing images (HR remote sensing images) cannot be fully realized, this paper proposes an object oriented semi-automatic method to extract typical natural objects from HR remote sensing images based on spectral and textural features. We introduce the concept of "seed points" that provides limited seed points from the ground objects to be extracted to realize the auto-extraction of large-scale typical natural ground objects under the premise of ensuring the extraction accuracy. We firstly segment the initial images based on the Minimum Spanning Tree algorithm, and then the spectral and textural features of each object are calculated according to the average normalized value and standard deviation of the image gray. Secondly, users provide foreground samples through interactive selection of "seed points", and then find the regional expanded foreground samples with the least merging cost based on Region Adjacency Graph (RAG). Thirdly, the Gaussian Mixture Models (GMMs) can better adapt to the color of the image, effectively capture the subtle differences between the foreground and the background, and approximate any probability distribution with arbitrary precision. They have been widely used in image processing and model building. This paper uses GMMs to estimate whether the current object belongs to the foreground or background models. We then update the mean field in the fully connected conditional random field through the feature space Gaussian filter. The fully connected condition random field can describe the relationship between each node and all other nodes. Finally, the contour extraction results are obtained according to the extraction criteria of different natural objects and the global information description with the fully connected conditional random fields. In this study, we utilized aerial images and GF-2 images to verify our method, and extracted five typical natural objects including woodland, grassland, cropland, bare land, and waters. The results show that the total accuracy and Kappa value of typical natural features extracted from aerial images were 0.959 and 0.948, respectively. They increased by 20.757% and 0.268, respectively, compared with the SVM method. The total accuracy and Kappa value using Gaofen-2 (GF-2) remote sensing images were 0.959 and 0.941, respectively. They increased by 1.698% and 0.133, respectively, compared with SVM. Our results indicate that the proposed method can effectively extract the natural objects in HR remote sensing images with less interaction and time.
In flood disaster monitoring, fast and accurate detection of inundated area and flooded disaster region is of great value for flood control and post-disaster reconstruction work. This paper takes the 2017 Saint Louis flood in the United States as an example. Based on Sentinel-1 SAR data, the method of combining change detection and threshold was used to achieve large-scale flood inundation extraction. Firstly, the SAR data were pre-processed with sigma radiation calibration and Refined Lee filtering, which were effective in improving the contrast of land and water bodies, as well as attenuating the coherent speckle noise. Secondly, the difference image between the reference image and flooded image was defined by change detection methodology and the image histogram was divided by the quantile threshold method to extract the submerge area. Finally, image post-processing was performed on the thresholded results to reduce misclassification. Areas not close to the water surface and whose slope was higher than 3 degrees were defined as non-flood region for exclusion using the digital elevation model. Then, the small particle noise and holes were removed by morphological filtering to achieve large-scale flood inundation extraction. The boundary information was retained while keeping the original size of the flood category unchanged. Heavy rainfall was the main cause of the 2017 extensive flooding in Saint Louis. The low-lying northern river bend area was the most severely affected, inundated for up to two months while the main city suffered less damage due to its high terrain and timely flood protection. Until now, there have been few studies on the effectiveness of different synthetic aperture radar data polarization modes in relation to flood detection. The Sentinel-1 VV/VH polarization data were compared with the flood inundation extraction range obtained from the Sentinel-2 optical image during the same period. Then, the flood detection applicability of the polarization mode was evaluated based on the comparison results. The scattering response characteristics in the multi-polarization patterns were analyzed by plotting the back-scattering cross-sectional lines for different periods of each polarization. The results show that both Sentinel-1 VV and VH polarization data can identify flood with a high accuracy of over 82%. Compared with VH polarization mode, VV polarization mode has fewer false positives. In the same region, the scattering degree of Sentinel-1 VV polarization signal was 28% smaller than that of VH, showing more sensitive information from the flood. Therefore, Sentinel-1 VV polarization mode is more suitable for monitoring the inundation range of flood disaster.
Fine-scale crop classification has always been a hot topic in the field of agricultural remote sensing, which is of great significance for crop yield estimation and planting structure supervision. The emergence of deep learning provides a new way to improve the accuracy of crop classification. Recently, the Convolutional Neural Network (CNN), a representative algorithm of deep learning, shows obvious advantages in processing high-dimensional remote sensing data. However, the application of CNN in crop classification based on multispectral data is still rare, and the classification accuracy dependent on the different feature information of crops is hard to evaluate. In this paper, a crop classification method based on feature selection and CNN for multispectral remote sensing data is proposed to improve fine crop classification. This study used Sentinel-2 remote sensing images as data source. Based on the reflectance of 13 multispectral bands and 10 vegetation indices including normalized difference vegetation index, ratio vegetation index, enhanced vegetation index, etc., the Relief F algorithm was used to rank the contribution of multidimensional features. According to the rank of feature contribution, the features with high contribution were selected and optimized by group training to obtain the best feature collection. Therefore, a CNN-based classification method based on feature selection was designed. Based on this, we classified and mapped the main crops including rice, corn, and peanut in Yuanyang County, Henan Province, with an overall classification accuracy of 96.39%. Meanwhile, the support vector machine and simple CNN were also used to classify the main crops in the research area for comparison. We found that the CNN-based classification method based on the optimal feature collection had the highest classification accuracy, followed by simple CNN, and the support vector machine had the worst performance. The main conclusions of this research are as follows: (1) The Relief F algorithm was effective to sort the contribution of different features. In total, we obtained 24 optimal feature subsets, with a training accuracy of 99.89%; (2) The CNN-based classification method using the optimal feature collection can extract the high-precision difference in features to the greatest extent and realize the fine-scale classification of crops. Compared with simple CNN and support vector machine, the CNN method based on the optimal feature collection has obvious advantages.
Various optical vegetation indices have been widely used in vegetation monitoring and crop yield estimation. However, the temporal availability of optical vegetation indices derived from visible and infrared remote sensing bands is usually a problem in many studies. Currently, the Vegetation Optical Depth (VOD) derived from space-borne microwave radiometers that is unaffected by cloud cover has been found to be proportional to the vegetation density and water content, which shows a great potential in crop monitoring using remote sensing. The vegetation information captured by the two remote sensing approaches is different and complementary since they come from satellite sensors with different spectrum ranges. In this study, we focused on the synergistic use of optical remote sensing data and microwave remote sensing data to estimate wheat yield more accurately. We selected the VOD estimated by the L-band microwave radiometer on board of the SMAP mission, and the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), the Leaf Area Index (LAI), and the Fraction of Photosynthetically Active Radiation (FPAR) product retrieved from the MODIS satellite sensor as the input variables for winter wheat yield estimation using neural network regression models. We compared the performances of back propagation neural network, genetic algorithm back propagation neural network, and particle swarm optimization algorithm back propagation neural network regression models for estimating wheat yields. The results show that the significance values (P) of the three neural network regression models were all less than 0.001, which indicated that all models have passed the significance test. The genetic algorithm back propagation neural network regression model was the best compared to the other two neural networks regression models, with the highest correlation (R=0.755) and the lowest root mean square error (RMSE=529.145 kg/hm2), mean absolute error (MAE=425.168 kg/hm2), and mean relative error (MRE=6.530%). Moreover, in order to analyze the advantages of different optical vegetation indices in crop yield estimation, we also established another two different genetic algorithm back propagation neural network models that used NDVI and LAI, and that used NDVI, EVI, LAI, and FPAR optical data for winter wheat yield estimation as a comparison. By comparison, the correlation (R) of the model established by the microwave and optical remote sensing data increased by 0.163, 0.229, and 0.056, respectively; while its root mean square error (RMSE) decreased by 122.334, 158.462, and 46.923 kg/hm2, respectively. The combination of multi-source remote sensing data can improve the accuracy of model results to a large extent.
Based on the complete data sequence and abundant spectral information of the moderate resolution imaging spectrometer, remote sensing feature indices play an irreplaceable role in studying the stage, trend, and law of wetland ecosystem development. There are some limitations in selection of remote sensing feature indices for the traditional inter-class distance discrimination, such as over-dependence on statistical characteristics of data, indefiniteness of the inclusion index and the ecological significance of the target area, and poor applicability of the classification model. Based on this, Baiyangdian Wetland Nature Reserve in Hebei Province is selected as the research area. The method of Class Separation Discrimination (CSD) and Class Distance Discrimination (CDD) is proposed to construct the optimal remote sensing feature index. In addition, QUEST algorithm and Markov distance discrimination method are used to construct the classification decision tree model for the Baiyangdian Wetland extraction, which overcomes the shortcomings in the selection of traditional inter-class distance index. The results show that, firstly, the overall classification accuracy of the remote sensing feature index selected by the method combining CSD and CDD reach 91.32% and the kappa coefficient is 0.88 for the wetland information extraction in the study area. Compared with the traditional Classification and Regression Tree (CART) method, the classification accuracy is improved by 1.67%. Secondly, the selected optimal remote sensing feature index has a clear ecological meaning for the type of wetland extracted. For example, the redoxing process of emergent water plants under alternate dry and wet conditions determines that iron oxide index (IO) can be successfully selected to further separate the mixed cultivated land and emergent water plants. Furthermore, according to the OLI image of the study area in 2017, the decision tree model based on the combination of CSD and CDD and the decision tree model based on CART algorithm are applied to the classification of the OLI image in the study area in 2019. The overall classification accuracy and kappa coefficient of the model based on the combination of CSD and CDD are 86.97% and 0.83, respectively. The model based on cart method cannot meet the classification requirements. The research results show that the model based on the combination of CSD and CDD has good applicability and stability over years. In a word, the combination of CSD and CDD can effectively avoid the limitations of traditional remote sensing feature indices, and improve the applicability of classification model. It is a beneficial attempt to combine remote sensing feature index selection algorithm with decision tree in wetland information extraction.
The remote sensing technology has been widely used in water body extraction and monitoring. Nowadays, many researchers are committed to improving the accuracy of water body extraction. Discrete particle swarm optimization algorithms can obtain high accuracy and robust classification results in remote sensing image classification, which has been widely used to extract the water bodies. However, the applicability and accuracy in water body extraction using discrete particle swarm still need to be assessed and verified. Here we proposed two new methods for water body extraction based on the discrete particle swarm optimization algorithm, namely, Spectral Matching Coupled Discrete Particle Swarm Optimization method (SMDPSO) and Maximum Entropy coupled Discrete Particle Swarm Optimization method (MEDPSO). Based on Landsat8_OLI remote sensing data, we selected the imageries with four environmental elements, i.e., ice and snow, clouds, mountain shadows, and buildings. The conventional method was used to extract the water body, and the results were compared and verified with two commonly used water index methods (NDWI, MNDWI). The results show that: (1) SMDPSO and MEDPSO methods can quickly find the best water bodies in the four experimental areas, and the two methods were all applicable for the study areas. Using the NDWI and MNDWI methods, water bodies can be misclassified with ice, snow, clouds, shadows, and buildings, and the extraction accuracy was low; (2) The SMDPSO method can identify small rivers and discrete water bodies. The overall water body extraction accuracy was high, but the extraction accuracy was low in complex environment. The MEDPSO method can not only identify small water bodies, but also suppress background information interference in the extraction process which cannot be realized by the other three methods. The overall accuracy of the four experimental areas was above 97.8%, which was higher than the other three methods; (3) By introducing the discrete particle swarm optimization algorithm into the water body extraction methods, the regional integrity of each method can be enhanced, and the accuracy and automation of the water body extraction can also be improved; (4) The machine learning methods such as the maximum entropy model, and image information such as spectrum, shape, and texture, as well as terrain information can be used to identify water bodies. These methods can achieve higher accuracy in the water body extraction. These results provide scientific references for the application of discrete particle swarm optimization algorithms, as well as the selection of water body extraction methods using remote sensing data.
The spatial-temporal fusion technology is an effective tool to blend observations from sensors with different spatial and temporal characteristics. The ESTARFM algorithm has good applicability to areas with fragmented land, and is susceptible to meteorological conditions, and has important practical significance for resource and environmental monitoring in southern China. However, how to select the fusion scheme and set the parameters of the fusion model to achieve the best fused time series vegetation index data is still unclear. This paper takes Nanchang County, a typical county in the middle and lower reaches of the Yangtze River, as an example, to analyze the parameters sensitivity of the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) for fusing multi-temporal Landsat and MODIS data. Based on Normalized Differential Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI), we systematically compared the performance of vegetation index fusion using the RI scheme (band reflectance was fused firstly and then the vegetation index was calculated) and IR scheme (vegetation index was calculated firstly and then directly fused). The results show that: (1) The sensitivity of parameters (sliding window size and number of similar pixels) in ESTARFM displayed similar characteristics in both the fusion of band reflectance and vegetation index. With the increase of sliding window size and number of similar pixels, the R 2 and SSIM values of band reflectance and vegetation index increased firstly and then remained steady or decreased, while the overall fusion error (MAE and RMSE values) decreased firstly and then remained steady or increased. There was an optimal parameter setting range in the application of ESTARFM model. The parameter sensitivity analysis is required to determine the optimal parameter range when adopting ESTARFM algorithm in different regions; (2) Compared with the RI scheme, the IR scheme had a higher fusion accuracy (R2RI-NDVI=0.866, R2IR-NDVI=0.953, R2RI-EVI =0.814, R2IR-EVI =0.930). It produced less outliers and noise during the fusion process and can effectively weaken the "pattern spot" and preserve spatial details and texture features, resulting in a high similarity with the real image. In addition, based on Landsat and MODIS multi-temporal images, the ESTARFM algorithm can also be used to generate high-temporal-resolution images to approximately replace the cloud and cloud shadow areas in Landsat images, which can effectively overcome the "cloud pollution" phenomenon in cloudy and rainy areas and improve remote sensing data quality. Our results provide a reference for the application of spatial-temporal fusion model in the complex environment with fragmented land and changeable planting systems.
Bosten Lake is a typical inland lake in the arid zone. The change in the lake area is strongly related to local natural and cultural environmental changes. Based on the GIS and RS technologies, this paper combines Landsat imagery and MODIS data, including a total of 2289 scenes, with JRC GSW water mask products to characterize the interannual and intraannual changes of the area of Bosten Lake from 2000 to 2019 through the Google Earth Engine (GEE) platform using index methods. We use the 2019 Sentinel-2 images to compare and analyze the results. To quantify the the causes of the changes, we analyzed the human activities and daily meteorological data of Yanqi, Korla and Bayanbuluk meteorological stations during 2000-2018. Results show that: (1) the GEE is efficient for integrating multi-temporal high-resolution remote sensing data to analyze the temporal change of lake area, especially the intraannual change. Compared with Landsat-5/7/8 and MOD09GQ data, the lake shoreline extracted based on Sentinel-2 images shows more details due to their high temporal and spatial resolution; (2) during 2000-2013, the total lake area decreases by 181.66 km2 with a decreasing rate of 13.98km2/a; while during 2013-2019, the lake area increases by 133.13 km2 with a increasing rate of 22.19 km2/a; (3) Intraannually, the lake area shows an upward trend from Mar. to Jun., keeps peak until September, and decreases from Oct. to Dec. and (4) the interannual change of Bosten Lake area has no significant correlations with the changes of evaporation, precipitation, and accumulated temperature within the watershed. While the intraannual change of Bosten Lake area shows strong correlations with those meteorological varabiles.
