Mapping of sugarcane crop area in the Paraná State using Landsat/TM/OLI and IRS/LISS-3 images

Mapeamento de área cultivada de cana-de-açúcar no estado do Paraná com imagens Landsat/TM/OLI e IRS/LISS-3 R E S U M O O conhecimento de estimativas confiáveis de áreas cultivadas de cana-de-açúcar é imprescindível para o agronegócio brasileiro por auxiliar no desenvolvimento de políticas públicas, na determinação dos preços aos produtores pelas usinas e permitir estabelecer a logística de escoamento da produção. O objetivo deste trabalho foi realizar o mapeamento de área da cultura de cana-de-açúcar para o estado do Paraná a partir de imagens dos satélites Landsat/TM/OLI e IRS/LISS-3, para as safras de 2010/2011 a 2013/2014. Os mapeamentos foram realizados por meio da classificação supervisionada de Máxima verossimilhança (Maxver) obtendo-se, em média, uma exatidão global de 94,13% e índice kappa de 0,82. As correlações com os dados oficiais do IBGE variaram de moderada a forte (0,64 ≤ rs ≤ 0,80) com concordância (dr) média de 0,81. Houve aumento de 2,73% (18.630 ha) de área com cana-de-açúcar no Paraná entre 2010/2011 e 2013/2014.


Introduction
Agricultural planning and the acquisition and provision of precise and fast information on the agricultural statistics can be made in a more objective way through orbital remote sensing (Aguiar et al., 2011;Mulianga et al., 2015;Schultz et al., 2015).
It is possible to map and estimate crop areas in large territorial extensions in short and medium term, as performed by Rudorff et al. (2010), who mapped and monitored sugarcane in the Center South of Brazil through the Canasat project using techniques of visual interpretation on satellite images.
Among the data of remote sensing, the MODIS orbital sensor has great potential in the mapping of agricultural areas (Adami et al., 2010;Souza et al., 2015).The temporal series of vegetation indices (VI) allow to monitor the dynamics of agricultural cultivation through the analysis of the VI spectrotemporal profile and, then, identify and map semi-perennial crops such as sugarcane (Xavier et al., 2006;Alves et al., 2014) or other annual crops such as soybean and maize (Arvor et al., 2011;Johann et al., 2016).
The sensors Thematic Mapper (TM) and Operational Land Imager (OLI) of the Landsat satellite and Linear Imaging Self Scanning (LISS-3) of the IRS satellite are widely used because of their better spatial resolution (Landsat with 30 m and IRS with 23.5 m) when compared to the MODIS, which allows a better distinction of the terrestrial targets, as demonstrated by studies on the mapping of areas of sugarcane and other agricultural crops (Silva et al., 2010;Vieira et al., 2012;Adami et al., 2012;Zhou et al., 2015).
According to the data of IBGE (2016), the state of Paraná produced, on average, in the period of five years comprehended between 2011 and 2015, 47.3 million tons of sugarcane per year, which represents 6.4% of the national production, being the 4 th largest producer.Thus, the state needs a rapid and efficient methodology to map these areas, especially because the Canasat project has been discontinued.
This study aimed to use images of the Landsat/TM/OLI and IRS/LISS-3 satellites associated with data of the TERRA and AQUA satellites of the MODIS sensor to map and estimate areas cultivated with sugarcane in the Paraná state.

Material and Methods
The Paraná state is located in the Southern region of Brazil, situated between the parallels 22º 29' S and 26º 43' S and meridians 48º 2' W and 54º 38' W (Figure 1).
The images of the Landsat/TM/OLI and IRS/LISS-3 satellites were made available by the United States Geological Survey (USGS, 2012) and by the National Institute for Space Research (INPE, 2016), and have, respectively, spatial resolutions of 30 and 23.5 m.
The supervised classification method of Maximum Likelihood (Maxver) was used to map sugarcane areas, following the example of other studies (Phongaksorn et al., 2012;Silva Junior et al., 2013;Mulianga et al., 2015).This algorithm makes the 'pixel-to-pixel' classification of the images, using the multispectral information of each pixel obtained in previously acquired training samples to find homogeneous regions (classes).It is assumed that these classes have normal distribution, estimate the mean values of each class and the covariance matrix of the training samples to then compute the statistical probability of an unknown pixel of the image to be classified (Nery et al., 2013).
Here, the acquisition and selection of the set of training samples used in the classifications were performed in images from off-season periods of summer crops (maize and soybean) and winter crop (wheat), in false-color compositions RGB-564 for Landsat/OLI and RGB-453 for IRS/LISS-3 and Landsat/ TM.This step also used the spectro-temporal profile of the vegetation index EVI (Enhanced Vegetation Index) (Huete et al., 1997) of the TERRA and AQUA satellites of the MODIS sensor, so that the training samples were homogeneous and exclusively from sugarcane plantations.To obtain a maximum classification error of 5%, Eq.1 (Niel et al., 2005) was used to define the minimum size of training samples (540 pixels) for each orbit/point of the studied satellites and years.This procedure facilitated the learning of the Maxver classification in the mapping of sugarcane area, performed in the program ENVI 4.5 (EXELIS, 2016).The spatial accuracy of the generated mappings was evaluated using the overall accuracy (OA), Kappa agreement index (KI), Omission Errors (OE) and Inclusion Errors (IE) (Congalton, 1991), taking as reference the maps of the Canasat project in the crop years from 2010/2011 to 2013/2014, made (1) by the INPE (Rudorff et al., 2010).To obtain a sampling error of approximately 1.58%, 4000 samples are necessary, according to Barbetta (2007).Thus, a stratified sampling was applied, randomly selecting 2000 sampling pixels in each Sugarcane class and 2000 pixels in the Not-sugarcane class.
The last step consisted in the extraction of the mapped sugarcane area using a routine developed in the IDL programming language (Esquerdo et al., 2011), which counts the number of mapped pixels and multiplies it by the area of each pixel for each municipality.These data of municipal area of each crop year were compared with the municipal data of harvested area provided by the IBGE (2016) based on the statistical indicators: Spearman's correlation coefficient (r s ), because the data do not exhibit normality by the Shapiro-Wilk test at 0.05 significance level; mean error (ME) and the refined version of the Willmott's index of agreement (dr) (Willmott et al., 2012), which measures the accuracy of the values estimated by the mapping in relation to the data from the IBGE.

Results and Discussion
The results of the spatial distributions of the areas cultivated with sugarcane in the Paraná state evidenced a higher concentration in the mesoregions Northwest, Central North and Pioneer North (Figure 2), which corroborates official data of harvested area (in hectare) of the crop between the years 2008 and 2012 (IBGE, 2016).
The accuracy analysis for the mappings with the Kappa index varied from 0.8 (2010/11) to 0.84 (2013/14), indicating excellent thematic quality (KI ≥ 0,81) (Table 1) according to Landis & Koch (1977).Phongaksorn et al. (2012) Despite the size of the areas and different studied regions, these results were similar to those found in the present study, indicating that the mappings of sugarcane areas in Paraná with the Maxver classifier showed excellent results, allowing reliable estimates of areas, in a fast and objective way.
One of the reasons that led the Maxver classifier to be excellent was the method of acquisition of training samples, which was not exclusively based on the RGB false-color compositions of the Landsat and IRS, but also on the analysis of the spectro-temporal profile of the EVI vegetation index of the MODIS sensor, considerably decreasing the spectral confusion in the mapping process.In addition, the state does not have extensive areas of pasture, which normally generate a lot of spectral confusion in the mapping of sugarcane area.
The mean overall accuracy (OA) obtained through the Maxver classification in Paraná was 94.31% (Table 1), above the  The high accuracy obtained in this study indicates that the methodology is efficient in the mapping of sugarcane crop area; however, the accuracy must also be analyzed individually by the errors IE and OE.The samples randomly selected on the mappings of sugarcane crop had a variation from 6.24% (crop year 2013/14) to 10.31% (crop year 2010/11) of IE (Table 1).
In the samples randomly selected on the not-sugarcane areas, there was a variation between 14.43% (crop year 2013/14) and 16.42% (crop year 2010/11) of IE, which represents the values of the sugarcane crop not selected by the classification.The OE varied between 15.53% (crop year 2013/14) and 19.11% (crop year 2010/11), and that is the proportion of reference samples that were excluded from the class to which they belonged.Vieira et al. (2012) obtained IE of 5.1% for the Sugarcane class and 6.6% for the Not-sugarcane class, and the OE were 9.7% for the Sugarcane class and 3.4% for the Not-sugarcane class.Mulianga et al. (2015), mapping sugarcane areas, found IE of 4.2% and OE of 20%.Adami et al. (2012) obtained a mean overall error for sugarcane of 2% with an inclusion of mean error of 2%, compensated by the omission of the mean error of also 2%, and concluded that the Canasat maps have excellent thematic quality.Although the Canasat maps show higher accuracy, the procedure for their generation consists in a more costly and time-consuming method, compared with the Maxver, which allows estimates of area prior to harvest.
Although the OE and IE obtained by Adami et al. (2012) are lower than those obtained in the present study in Paraná, the use of Maxver supervised classification in this study has the great advantage of being automatic and agile in the estimate of sugarcane areas in the state, which consequently allows anticipation of agricultural statistics by the official organs or companies interested in the information.
In addition, it should be highlighted that the evaluation of spatial accuracy in Paraná was made with large number of samples (4000) generated from the reference of the Canasat maps, which gives statistical reliability in the analyses presented here.
The area obtained by the mapping in the Paraná state had an increase of 18.63 thousand ha (2.73%) in the period between 2010/11 and 2013/14 (Table 2).Rudorff et al. (2010) reported an area increment of 1.88 million ha (73.2%) between 2003/04 and 2008/09 in São Paulo, concluding that the large growth in this period occurred due to the demand for ethanol caused by the advent of vehicles equipped with the Flex-fuel technology from 2003 on in Brazil.
In the comparison between the official data of IBGE and those of the mappings (Table 2), there was a mean overestimation of 6.03% in the four evaluated crop years.The greatest difference occurred in the crop year 2012/13 (8.74%), while the lowest difference occurred in the crop year 2013/14 (3.03%), showing higher KI, which can be partially explained by the greater availability of images for acquisition, being free from clouds on the dates of off-season periods of the summer crops and quality of the OLI sensor images.Silva et al. (2010), based on the analysis of information on sugarcane cultivation area of the Canasat, observed a growth of 3.21 million ha (76.2%) between 2005/06 and 2009/10, which occurred mostly because of the construction of new processing facilities.In Paraná, the growth was 77.5%, but in the comparison of the last crop years (2008/09 com 2009/10) there was a decrease of area expansion and also in other states.
According to Aguiar et al. (2011), the decrease in harvested areas in the São Paulo state from the crop years 2008/09 and 2009/010 was influenced by the economic crisis, which also corroborates with the stabilization of growth in the estimates of areas observed in the Paraná state by the mapping generated through Maxver in the period between 2010/11 and 2013/14 (Table 2).
The official data of area from IBGE and those estimated in the mappings of sugarcane areas showed correlations (r s ) (Figure 3) that varied from 0.64 (crop year 2013/14) to 0.80 (crop year 2012/13), although they overestimated the area cultivated with sugarcane.
The mean error (ME) indicated that, on the state scale, the area obtained by the mapping was overestimated only between 51.67 ha (2013/14) and 141.28 ha (2012/13), compared with the official data from IBGE, showing a consistency between what was mapped and the official data.
The refined Willmott's index of agreement (dr) evidenced a high accuracy of the estimated values (sugarcane mapping) in comparison to the official data (IBGE) in all crop years evaluated, since the trend line of these data (Figure 3) is close  In general, these statistical indicators show that, although there are differences between the estimates of sugarcane area obtained through the Maxver method and the official data, it is possible to know, besides estimating the sugarcane area earlier, the spatial distribution (Figure 2) with which the crop is cultivated in the state.

Conclusions
1.The Maxver classification method showed excellent results in the mapping of sugarcane in Paraná, allowing reliable estimates of area, in a fast and objective way.
2. The obtained mappings showed optimal performance in the analysis of spatial accuracy using, as reference, the Canasat maps.The high accuracy obtained in the evaluation indicates that the methodology is efficient in the mapping of areas cultivated with sugarcane and can also be analyzed individually, from the analyses of the errors IE and OE.
3. The use of MODIS sensor images for the acquisition of sampling sets through the analysis of the spectro-temporal profile of EVI was essential to decrease the spectral confusion in the process of selection of homogeneous training samples, which caused the performance of the Maxver classifier to be more efficient.

Figure 1 .
Figure 1.Paraná state with location and distribution of the mills and distilleries with the Landsat/TM/OLI and IRS/ LISS-3 scenes , using the Landsat-/TM satellite in a sugarcane area in Thailand, obtained a KI of 0.92.Silva Junior et al. (2013) made an automatic detection of fires in sugarcane plantations and obtained the best result (KI = 0.90) with the Maxver classification with the treatment of radiometric correction.Vieira et al. (2012), using segmentation and Object-Based Image Analysis (OBIA) in Landsat/TM/ETM images, mapped sugarcane in the Northern São Paulo state and obtained a KI of 0.87.With the same method, Schultz et al. (2015) found maximum KI values of 0.89, in an area to the South of the same state using various scenes of the Landsat-8 satellite.

Figure 2 .
Figure 2. Mappings of the sugarcane crop areas in the Paraná state for the crop years (A) 2010/2011; (B) 2011/2012; (C) 2012/2013; (D) 2013/2014 4. The use of the Maxver supervised classification has the great advantage of being automatic and agile in the estimate of sugarcane areas in the state, which consequently allows to assist in the development of public policies and an earlier provision of agricultural statistics.

Table 1 .
Accuracy indices obtained for the mappings of sugarcane area through Maxver