Corresponding author — B. D. Wardlow, tel 402 472-6729, email firstname.lastname@example.org Abstract
The global environmental change research community requires improved and up-to-date land use/land cover (LULC) datasets at regional to global scales to support a variety of science and policy applications. Considerable strides have been made to improve large-area LULC datasets, but little emphasis has been placed on thematically detailed crop map- ping, despite the considerable influence of management activities in the cropland sector on various environmental pro- cesses and the economy. Time-seriesMODIS250mVegetationIndex (VI) datasets hold considerable promise for large- area crop mapping in an agriculturally intensive region such as the U.S. CentralGreatPlains, given their global coverage, intermediate spatial resolution, high temporal resolution (16-day composite period), and cost-free status. However, the specific spectral–temporal information contained in these data has yet to be thoroughly explored and their applicability for large-area crop-related LULC classification is relatively unknown. The objective of this research was to investigate the general applicability of the time-seriesMODIS250m Enhanced VegetationIndex (EVI) and Normalized Difference Veg- etation Index (NDVI) datasets for crop-related LULC classification in this region. A combination of graphical and statisti- cal analyses were performed on a 12-month time-series of MODIS EVI and NDVI data from more than 2000 cropped field sites across the U.S. state of Kansas. Both MODIS VI datasets were found to have sufficient spatial, spectral, and temporal resolutions to detect unique multi-temporal signatures for each of the region’s major crop types (alfalfa, corn, sorghum, soybeans, and winter wheat) and management practices (double crop, fallow, and irrigation). Each crop’s multi-tempo- ral VI signature was consistent with its general phenological characteristics and most crop classes were spectrally sepa- rable at some point during the growing season. Regional intra-class VI signature variations were found for some crops across Kansas that reflected the state’s climate and planting time differences. The multi-temporal EVI and NDVI data tracked similar seasonal responses for all crops and were highly correlated across the growing season. However, differ- ences between EVI and NDVI responses were most pronounced during the senescence phase of the growing season. Keywords: MODIS, vegetationindexdata, crops, land used/land cover classification, U.S. CentralGreatPlains
their research states “a timeseries of the 16-day composite MODIS250m VI data had sufficient spectral, temporal, and radiometric resolutions to discriminate the region’s major crop types and crop-related land use practices (Wardlow et al., 2007, p. 307).” By using MODIS250m VI data, they were able to obtain meaningful results in detecting unique multi-temporal VI signatures for each crop class and for evaluating the crop classes’ average multi-temporal response patterns. In other words, the researchers illustrated the potential of MODIS250mvegetationindexdata for crop mapping, balanced against other factors such as cost, availability of resources, and time constraints. In their view, science-quality imagery from MODIS with global coverage, high temporal resolution (1-2 day repeat coverage), moderate to coarse spatial resolution (250m, 500 m, and 1 km), and distribution free of charge provides major advantages compared to other sources used for regional to global scale LULC mapping.
Dash, 2007; Goodin & Henebry, 1997; Tieszen, Reed, Bliss, Wylie, & DeJong, 1997 ).
Remote sensing has been successfully applied to extract global land use and land cover patterns. The MODIS Land Cover Product (MOD12Q1) ( Friedl et al., 2002, 2009 ), for example, is currently one of the most commonly adopted land cover data sets for global biogeochemical models ( Jung et al., 2006; Potter et al., 2007 ). With monthly AVHRR Normalized Difference VegetationIndex (NDVI) products, the U.S. Geological Survey (USGS) developed the 1-km land cover database and identi ﬁed 159 seasonal land cover classes in the conterminous United States ( Loveland et al., 1995 ). Using satellite imagery at much ﬁner resolutions, various nation-wide efforts have also been made to extract more speci ﬁc land surface covers. For example, the U. S. Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) uses medium-resolution satellite images to produce annual Cropland Data Layer (CDL) maps for various states ( Boryan, Yang, Mueller, & Craig, 2011 ). Landsat imagery was also applied to develop the 2001 U.S. National Land Cover Database at 30-m resolution ( Homer, Huang, Yang, Wylie, & Coan, 2004 ). Howev- er, these land cover products were not designed to delineate C 3 and C 4
The goal of this study was to map rainfed and irrigated rice-fallow cropland areas across South Asia, using MODIS250mtime-seriesdata and identify where the farming system may be intensi ﬁ ed by the inclusion of a short- season crop during the fallow period. Rice-fallow cropland areas are those areas where rice is grown during the kharif growing season (June – October), followed by a fallow during the rabi season (November – February). These cropland areas are not suitable for growing rabi -season rice due to their high water needs, but are suitable for a short -season ( ≤ 3 months), low water-consuming grain legumes such as chickpea ( Cicer arietinum L.), black gram, green gram, and lentils. Intensi ﬁ cation (double-cropping) in this manner can improve smallholder farmer ’ s incomes and soil health via rich nitrogen- ﬁ xation legume crops as well as address food security challenges of ballooning populations without having to expand croplands. Several grain legumes, primarily chickpea, are increasingly grown across Asia as a source of income for smallholder farmers and at the same time providing rich and cheap source of protein that can improve the nutritional quality of diets in the region. The suitability of rainfed and irrigated rice-fallow croplands for grain legume cultivation across South Asia were de ﬁ ned by these identi ﬁ ers: (a) rice crop is grown during the primary ( kharif ) crop growing season or during the north-west monsoon season (June – October); (b) same croplands are left fallow during the second ( rabi ) season or during the south-east monsoon season (November – February); and (c) ability to support low water-consuming, short-growing season ( ≤ 3 months) grain legumes (chickpea, black gram, green gram, and lentils) during rabi season. Existing irrigated or rainfed crops such as rice or wheat that were grown during kharif were not considered suitable for growing during the rabi season, because the moisture/water demand of these crops is too high. The study established cropland classes based on the every 16-day 250m normalized difference vegetationindex (NDVI) timeseries for one year (June 2010 – May 2011) of Moderate Resolution Imaging Spectroradiometer (MODIS) data, using spectral matching techniques (SMTs), and extensive ﬁ eld knowledge. Map accuracy was evaluated based on independent ground survey data as well as compared with available sub-national level statistics. The producers ’ and users ’ accuracies of the cropland fallow classes were between 75% and 82%. The overall accuracy and the kappa coef ﬁ cient estimated for rice classes were 82% and 0.79, respectively. The analysis estimated approximately 22.3 Mha of suitable rice-fallow areas in South Asia, with 88.3% in India, 0.5% in Pakistan, 1.1% in Sri Lanka, 8.7% in Bangladesh, 1.4% in Nepal, and 0.02% in Bhutan. Decision-makers can target these areas for sustainable intensi ﬁ cation of short-duration grain legumes.
The GPR 250 REIT Index is a subset of the GPR 250Index and the selection criteria, calculations, periodic adjustments and all other fundamentals are exactly the same as that for GPR 250Index. Only companies that are organized as a REIT or similar structure are included in the GPR 250 REIT Index.
Four of the areas we selected for detailed sampling (areas “a”, “b”, “c” and “g”) are located in regions with high percentages of good quality pixels. Area “e” is situated in a transitional region, characterized by good quality occurring for approximately half of the months considered. Further interpretations for areas “d” and “f” should take into consideration the limited number of monthly composites labeled as having good quality. Reduced image quality can affect the analysis of timeseries and multiple methodologies have been proposed to deal with time-series reconstruction through gap filling and spike removal. However, no definitive method or procedural consensus exists involving those topics. We agree with  that the application of smoothing and filtering should be considered in a case-by-case basis. If judged necessary, these methodologies can be easily incorporated into our workflow as a pre- processing step. In particular, a spike removal approach could be used when identifying low- frequency/low-amplitude variations in phenology, often characterized by gradual transitions in vegetation response. In other situations, including the current work, the nature of the analyses requires the ability to identify abrupt changes in dataseries.
Accurate land cover mapping on a regional scale in the XUAR is useful for regional climate and environmental modeling. In this study, we evaluated the accuracy of seven global land cover products over the XUAR and found that significant discrepancies exist. Furthermore, the novel XUAR Landcover 2010 product was derived based on an automatic decision tree classification procedure employing the TWOPAC classification software. An extensive MODIS-derived EVI timeseries was utilized as the input data, covering six MODIS tiles with 46 dates each, which were first preprocessed and then used to extract phenological metrics. After post-processing, including the SRTM digital elevation model and parameters derived thereof, good accuracies of 79.78% for the overall produc ts and accuracies ranging from 22.52% to 99.2% for the individual classes could be attained. For selected areas within the XUAR, we also compared the results with higher resolution Landsat data and found that small-scale types, such as salty lands and the differentiation between deciduous forest and grasslands, can be captured. We consider that the XUAR Landcover 2010 product is a solid input for the modeling community or for future studies on regional land cover change. The novel product in this study can be shared with interested researchers active in the XUAR area.
To evaluate the vegetation restoration effect, anthro- pogenic and climatic impacts should be considered. Vege- tation cover change represents the most direct response of vegetation to climate changes and human activities (Zhao et al., 2012). Akiyama and Kawamura (2003) analyzed the land cover change over 1979–1997 and indicated that the areas with productive grasslands decreased while low-productivity grasslands increased. The seasonal change in vegetation (phenology) is a key parameter for studying and analyzing climate change and vegetation responses (the feedback be- tween the land surface and the atmosphere), which can im- prove the simulation quality of carbon, water, and energy ex- changes between the atmosphere and the land surface (Ma et al., 2013). The vegetation productivity and phenology in the temperate region of China has already changed in response to the dramatic climatic changes (Jeong et al., 2011; Piao et al., 2006, 2010; Peng et al., 2011). Earlier studies indicated that the recovery of vegetation from long-term degradation is related to the increase in precipitation (Eklundh and Ols- son, 2003; Sop and Oldeland, 2013). The growing body of evidence suggests that climate warming has advanced the bi- ological spring in temperate China (Chen et al., 2005; Piao et al., 2006; Zheng et al., 2002). Additionally, longer growing seasons, particularly earlier spring vegetation green-up, may significantly enhance the vegetation productivity in temper- ate and boreal regions (Cong et al., 2013; Hu et al., 2010; Kimball et al., 2004).
Thank you to Paul White (USDA soil scientist, former manager of Kansas State University soil microbiology laboratory) for instruction in analysis of microbial biomass and allowing use of laboratory facilities. Efficient Microbes (EM)™ Original solution was donated for this study by Sustainable Community Development, L.L.C. (Kansas City, Mo.). Trees for Life (Wichita, Kans.), a non-profit agriculture and educational organization, aerated the cow manure slurry with fresh bakers yeast donated by the American Institute of Baking, Manhattan, Kans. The J.C. Pair Horticultural Center of Kansas State University staff (Tami Roesch, Richard Ryer and Mike Shelton) is greatly appreciated for assisting with the planting and maintenance of the collard crops. We are grateful to the Trees for Life staff volunteers who harvested collards.
present research data set show that the assumption that dryland and irrigated corn and sorghum yields are approximately normally distributed seem to hold.
The variability in yield that is explained by environment is much higher than the variability that is explained by genetics within each cropping system. This result signifies how environment plays a significant role in determining crop yield. This further emphasizes the notion that crop recommendation should be environmental specific. Corn yield variation explained by environment is much higher than sorghum. A relative stability of sorghum yields across environmental variations compared to corn yields in this analysis support previous findings (Boyer 1970; Beadle, 1973; Stone et al., 1996; Fischer et al., 1982). Obviously, dryland yields are relatively more environmental dependant than irrigated yields just because one environmental factor, water, is less of a limitation in irrigated systems.
Desta forma, para estudos em escala regional - especialmente os realizados por meio de composições multitemporais -, é possível realizar as análises das mudanças no uso e cobertura da terra por meio de índices de vegetação por diferença normalizada (NDVI) ou índice de vegetação realçado (EVI) do sensor MODIS, os quais são capazes de evidenciar variações sazonais, interanuais e de longo termo de parâmetros estruturais, fenológicos e biofísicos da vegetação (Huete et al., 2002; Correia et al., 2006). As séries temporais do NDVI e EVI poderão evidenciar se as mudanças ocorridas são devidas à sazonalidade climática ocasionada pela alteração no regime de chuvas ou se ocorrem por mudanças no uso da terra (p. ex. desmatamentos) ou, ainda, se são causadas por queimadas. Para tanto, torna-se importante avaliar a capacidade de as séries temporais de NDVI e EVI propiciarem informação suficiente para possibilitar a diferenciação na mudança de comportamento espectral da vegetação, de modo a predizer se houve ou não desmatamento ou alteração nas classes de vegetação nativa e de uso da terra. Nesse contexto, o objetivo geral deste estudo é avaliar o uso de séries temporais MODIS NDVI e EVI para detectar desmatamentos no bioma Cerrado.
To increase the sensitivity of detecting discrete changes in montane forest ecosystems, it is, therefore, advantageous to employ more robust approaches of dense timeseriesanalysis that can also consider seasonality [ 37 , 38 , 41 – 43 ]. The recent development of a range of algorithms including LandTrendR [ 44 ] for the Landsat archive and Breaks For Additive Season and Trend (BFAST) [ 21 ] for Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) datasets, has enabled near real time monitoring in the health and dynamics of ecosystems worldwide [ 38 ]. BFAST has previously been applied on a Landsat time-series using the Normalized Difference VegetationIndex (NDVI) in a montane forest in Ethiopia. It successfully identified small-scale forest disturbances, events characterised by a small number of pixels with varying magnitudes of change, from discrete clearings of the forest to progressive reductions in the canopy cover [ 22 ]. It has been proven to be robust in determining both discrete and abrupt changes in forest cover in a variety of habitats [ 42 , 45 , 46 ], even in regions with high cloud contamination, such as the Kenyan montane forests. Landsat data with a relatively fine spatial resolution (30 m-pixels), spanning more than four decades, could be ideal for studying forest disturbances using LandTrendR or BFAST. However, significant gaps in the archive exist over East Africa for the mid-1990s and early 2000s. The number of available Landsat scenes for the study sites at the time of analysis, and the significant gaps are shown in Appendix A . A trade-off in the spatial dimension, as well as the depth of the historical records, is to use imagery from MODIS at a very fine temporal (16-day) but coarser spatial (250m pixel) resolution, to detect forest disturbances in this region [ 46 – 48 ].
In the absence of sufﬁciently long-term studies that could adequately identify synergistic effects of cropping practices to improve WUE, cropping systems simulation models may be used. Kirkegaard and Hunt (2010) used the Agricultural Production Sys- tems Simulator (APSIM, Keating et al., 2003) to simulate multiple management effects (minimum tillage, weed control, crop rotation, planting date, and genotype selection) on wheat yield and WUE in southeastern Australia. They found combinations of management practices simulated over a 48-year period increased yields more than implementing any single practice, and that WUE increased from 6.0 kg ha −1 mm −1 for a baseline conventional till W–F system to 15.2 kg ha −1 mm −1 for a system in which all ﬁve of the suggested management practices were employed. Saseendran et al. (2010) simulated several centralGreatPlains dryland cropping systems and reported 16-year average wheat yields that were the same for the W–F and W–C–F no-till systems. The simulated average wheat yield for the W–M–F rotation was numerically greater than for W–F or W–C–F, but not signiﬁcantly so. The simulated WUE values were 3.13, 4.51, and 3.89 kg ha − 1 mm − 1 for W–F, W–C–F, and W–M–F,
City Safety is an auto-braking mechanism that intends to detect a possible road collision and, as a consequence, to prevent it. Activation of this system occurs if an object appears in front of a vehicle and the driver does not respond in order to avoid it. On rare occasions, it may happen that according to the driver’s perception, the City Safety activation is not needed. Thus, this is a false activation of City Safety. The purpose of this thesis was to develop a machine learning algorithm that will classify activations and determine what influences this classification. Data used for the analysis was represented as a multivariate timeseries. A length of a timeseries was 40, constituting to 8 seconds and the activation of City Safety was at the time equalled to 4 seconds. The variables represented the sensor fusion signals, such as a speed of the host vehicle or a distance to the target object. There were 223 events with a ground truth label. 71 events were used for testing, as they were annotated after the start of the research. The rest (152 events), where therefore used for training and validating. Due to the small sample size, 100 random train-val splits were applied, creating 100 different data setups which were used for modelling. Two machine learning models were proposed: (1) Baseline, and (2) Pseudo-labelling model. Both models reached quite high performance (82% accuracy on tests sets) for the optimal set of parameters.
where b (i) is around 0, and a (i) , d (i) around 1.
Fig. 1. Clustering the EVI temporal pattern in multidimensional feature space.
In this study, the complexity and enormous amount of time-series EVI datasets may lead to the difficulty of obtaining the actual number of clusters. Therefore, to provide maximum effectiveness of the clustering algorithm, we first consider the number of clusters of 25 which was then evaluated based on a statistical measurement of how separate that pattern is to patterns in its own cluster compared to patterns in other clusters. The separability analysis was applied to discriminate among high detailed significant patterns that were theoretically defined to portray the specific characteristics of each peat swamp area in the study site.
Since its inception in the 1950s (, ), k -nearest neighbor ( -NN) still receives regular interest among researchers; both in the theoretical aspect and the practical aspect. Its discrimination procedure is simple but powerful and needs virtually no modification to handle multi-class problems, i.e. it just obeys the majority vote for the classes among the k nearest neighbors of the sample being considered. k -NN decision rules gained theoretical acceptance since its early age of development;  developed their notions of consistencies between sequences of decision functions and showed that a formulation of k -NN is consistent with a reference decision rule. Many notable points are worth mentioning in their work. They initiated the field of nonparametric classification, the distribution generating the examples need not be assumed to be Gaussian or any other parametric distributions. The reference decision rule mentioned in their work as the ''likelihood ratio procedure''  is closely related to what is known today as the Bayes classifier. The Bayes classifier is the best classifier that will yield the lowest possible expected misclassification given that we know the distribution of the data; it will be discussed in detail later. They established that whenever the number of available examples approaches infinity and are dependent of such that and , the decision of k n NN will get arbitrarily closer to that of the likelihood ratio procedure with high probability. For example, one may choose k n t be
Since the integrated periodogram can be seen as a function, we shall use specific techniques for functional data. Nowadays functional data are present in many areas, sometimes because they are the output of measurement processes, other times for theoretical or practical reasons. There are several works on the statistical analysis of functional data and, particularly, on their classification. For example, a penalized discriminant analysis is proposed in Hastie et al. (1995); it is adequate for situations with many highly correlated predictors, as those obtained by discretizing a function. Nonparametric tools to classify a set of curves have been introduced in Ferraty and Vieu (2003), where authors calculate the posterior probability of belonging to a given class of functions by using a consistent kernel estimator. A new method for extending classical linear discriminant analysis to functional data has been analysed in James and Hastie (2001); this technique is particularly useful when only fragments of the curves are observed. The problem of unsupervised classification or clustering of curves is addressed in James and Sugar (2003), who elaborate a flexible model- -based approach for clustering functional data; it is effective when the observations are sparse, irregularly spaced or occur at different time points for each subject. In Abraham et al. (2003) unsupervised clustering of functions is considered; they fit the data by B-splines and partition is done over the estimated model coefficients using a k-means algorithm. In a related problem, Hall et al. (2001) explore a functional data-analytic approach to perform signal discrimination. Nevertheless, many of these procedures are highly sensitive to outliers. A simple idea to classify functions is to minimize the distance between the new curve and a reference one of the group. The approach presented in this paper follows this idea. As a reference function of each group we shall take the mean of the integrated periodograms of its elements. Later this curve will be substituted for a more robust representative.
4.3.2 Single Classes
Grassland vs. Forest
The corresponding single-date classification of the NDVI images confirms the seasonal FI patterns, with their peak in accuracy around acquisition period 18 in early October. The minor local FI maximum in spring was not evident in the average classification accuracies, since only single years exhibit this trend. 2010 stands out with unusually high classification accuracies, in excess of 91 % from a single composite image, and a long phase of very good separability. During early 2010 the NDVI values of Grassland were unusually low compared to other years, while Forest did not suffer from such extremely depressed NDVI values, and as a result classification at this time period outperformed the best result of each other year by more than 5%. The successive classification process shows very similar results to the classification of all classes together. Starting from the image with highest separability, the classification accuracy converges to its natural classification limit of 90 % at around 10 images.
Satellite remote sensing is well proven as an effective means of discriminating different crop types (Chen et al., 2012; Fang, 1998; Forkuor et al., 2015; Karkee et al., 2009; Van Niel and McVicar, 2004; Zhang et al., 2015) and for monitoring crop phenology (Fangping et al., 2007; Ganguly et al., 2010; Pan et al., 2015; Potapov et al., 2008; Qiu et al., 2015; Sha et al., 2016; Wu et al., 2010). In particular, several studies have engaged in discriminating rice cropping systems using remotely sensed imagery over parts of China (Sakamoto et al., 2006; Xiao et al., 2006; Xiao et al., 2005). The use of vegetation indices to capture the spectral information contained in multi-temporal remotely sensed data has been shown to provide valuable information on the seasonal development of crops (Ganguly et al., 2010; Liang et al., 2011; Wu et al., 2010). The most common of these vegetation indices used in crop monitoring is the Normalised Difference VegetationIndex (NDVI) (Holben et al., 1980; Huete et al., 2002; Rouse et al., 1974; Tucker, 1979). Several studies have explored the use of satellite derived NDVI to measure vegetation health (Elmore et al., 2000; Lenney et al., 1996; Rouse Jr et al., 1974) and to estimate and predict crop yield (Kastens et al., 2005; Mkhabela et al., 2011; Ren et al., 2008). The availability of frequent satellite data as provided by the Advanced Very High- Resolution Radiometer (AVHRR) (Kastens et al., 2005) and the Moderate Resolution Imaging Spectroradiometer (MODIS) (Huete et al., 2002; Mkhabela et al., 2011) has allowed for the development of frequent NDVI products over large areas. In particular, the daily revisit capability of MODIS (resulting in 8 – 16 day composite products), has made its data a common choice for crop phenology studies and monitoring systems on regional or country-wide scales (Begue et al., 2014; Bolton and Friedl, 2013; Chen et al., 2012; Fangping et al., 2007; Ganguly et al., 2010; Liang et al., 2004; Mkhabela et al., 2011; Ren et al., 2008; Son et al., 2014; Zhang et al., 2015; Zhang et al., 2003). However, the limitation of these products is their coarse spatial resolution (MODIS – 250 to 500 m and AVHRR – 1.09 km at nadir), making them incapable of detecting variability at fine scales because of mixed pixel effects whereby a single pixel in the image may cover two or more fields or management units (Peng et al., 2011; Xiao et al., 2005). Hence, there is limited capacity to use this data to provide crop growth information at the localised scale that is appropriate to farmers.
Rainfall, as a key factor, controls active variations in vegetation status especially in arid and semi-arid regions as vegetation health in those regions is vulnerable to the quantity, duration, and frequency of rainfall (Dutta et al. 2015 ; Kundu et al. 2015 ). Spatio-temporal dynamics of growing season mean 1-month SPI for the period of thirty years from 1989 to 2019 shows that year 2003 was the driest and year 2007 was the wettest in the study area. The dry (2003) and wet (2007) years were consid- ered to establish the relationship among growing season mean 1-month SPI and growing season mean NDVI and VCI. Analysis of spatio-temporal dynamics of growing season mean 1-month SPI for the dry year (2003) showed that the SPI ranges from extremely dryness ( ≤− 2.00) to severely dryness (− 1.50 to − 1.99) conditions prevailed in western, central, southwestern, and northwestern parts of the district. The moderately dry (− 1.00 to − 1.49) and near normal (− 0.09 to 0.09) conditions were observed in northern, northeastern, and southeastern parts of the district (Fig. 2a ). However, during the wet year (2007), the SPI ranges from near normal (− 0.99 to 0.99) to extremely wet (> 2.0) (Fig. 2b ), and no region was found under moder- ately dry (− 1.00 to − 1.49), severely dry (− 1.50 to − 1.99), and extremely dry (≤− 2.00) conditions. The near normal (− 0.99 to 0.99) conditions were noticed in the western parts of the district. The western and northwestern parts of the district were charac- terized by moderately wet (1.00 to 1.49) conditions. However, very wet (1.50 to 1.99) conditions were observed in the south- western parts of the district. Interestingly, during the wet (2007) year, majority of the study area was under extremely wet (> 2.0) condition. Analysis revealed that SPI could be used to identify the dryness and wetness, their intensities, and spatio-temporal extent (Dutta et al. 2015 ; Kundu et al. 2020 ). Furthermore, growing season mean 1-month SPI was compared with the corresponding vegetation indices, such as growing season mean NDVI and VCI, derived from temporal MODIS250mdata to assess the impact of rainfall on status of vegetation.