Testing automatic procedures to map rice area and detect

Testing automatic procedures to map rice area and detect phenological
crop information exploiting time series analysis of remote sensed
MODIS data
Giacinto Manfron a, Alberto Crema a, Mirco Boschetti a and Roberto Confalonieri b
a
CNR-IREA, Institute of Electromagnetic Sensing of Environment, Via Bassini 15, Milan, Italy
DI.PRO.VE., Department of Crop Science, University of Milano, Via Celoria 2, Milan, Italy
b
ABSTRACT
Rice farming, one of the most important agricultural activities in the world producing staple food for nearly one-fifth of
the global population, covers 153 MHa every year corresponding to a production of more than 670 Mton. Retrieve
updated information on actual rice cultivated areas and on key phenological stages occurrence is fundamental to support
policy makers, rice farmers and consumers providing the necessary information to increase food security and control
market prices. In particular, remote sensing is very important to retrieve spatial distributed information on large scale
fundamental to set up operational agro-ecosystem monitoring tool. The present work wants to assess the reliability of
automatic image processing algorithm for the identification of rice cultivated areas. A method, originally tested for Asian
tropical rice areas, was applied on temperate European Mediterranean environment. Modifications of the method have
been evaluated to adapt the original algorithm to the different experimental conditions. Finally, a novel approach based
on phenological detection analysis has been tested on Northern Italy rice district. Rice detection was conducted using
times series of Vegetation Indices derived by MODIS MOD09A1 products for the year 2006 and the accuracy of the
maps was assessed using available thematic cartography. Error matrix analysis shows that the new proposed method,
applied in a fully automatic way, is comparable to the results of the original approach when it is customized and adapted
for the specific study area. The new algorithm minimizes the use of external data and provides also spatial distributed
information on crop phenological stages.
Keywords: rice detection, phenology, MODIS, time series filtering, crop monitoring, Pareto boundary.
1. INTRODUCTION
Information on crop conditions throughout space and time, is a crucial issue to support the develop of more productive
and sustainable farming systems in order to answer to the continuous increment in food demand minimizing
environmental impacts such as greenhouses gas emission1 , water consumption, soil contamination and degradation2.
Developing countries are extremely vulnerable to the risk of food shortage and/or rice price increase; for the 2008’s
agronomical season shows a relative drop in the global rice production that determines a rice increase of 300%
corresponding to a value of 900 US$/t3. Changes in rice production and availability can produce food crises and grain
price variation that can be cause of social troubles due to economic and political changes. Developing countries countries
need scientific instruments able to provide them updated information about the status of crops, the crop quality condition
and the season characteristics to support early warning decision support systems. Moreover, in these countries up-to-date
crop status information during on-going season result very important to support food security initiative. On the other
hand in developed countries, such as the members of the European Union, rice is cultivated with more modern
techniques in intensive cropping system and it is not anymore a staple food. Rice production (3 Mton3) is lower than
consumption (- 1.07 Mton3) and it is cultivated in about 400.0003 ha mainly in few Mediterranean countries: Italy (49%),
Spain (29.7%), Greece (6.7%) and France (4.5%). Information on crop condition throughout space and time, is necessary
for the implementation of the Common Agricultural Policy. Regional/national scale crop monitoring, stress, early
warning and yield forecasting require the application of suitable techniques able to provide cost effective
spatial/temporal information on agro-ecosystems.
In this context, remotely sensing contribution can be very important in providing a continuous agro-ecosystem
monitoring tool for retrieving spatially distributed information on large scale. In particular multi-temporal satellite data
represent a suitable compromise between precision and feasibility. For these reasons especially in Asian countries,
several studies on mapping rice extent and estimate crop production using remote sensing data have been carried out
during the last years. Different sensors are used for rice monitoring and specific techniques exist for processing remotely
sensed data. Multi-temporal radar data are suitable for rice mapping due to the strong capacity in detecting water
presence typical of agronomic flooding, moreover in the tropics the all-weather day and night observation capabilities
represent a strong advantage for operational crop monitoring4,5,6. Optical remote sensing data represents another
possibility for rice monitoring at different scale. Some studies based on spectral classification procedures have explored
the potential of images from Landsat Thematic Mapper (TM) and National Oceanic and Atmospheric Administration
(NOAA) Advanced Very High Resolution Radiometer (AVHRR) in order to identify paddies respectively at local and
regional context7,8,9. In the last years Moderate Resolution Imaging Spectroradiometer (MODIS) sensor has been
strongly exploited to perform crop monitoring several examples exist for rice farming systems10,11,12. In particular Xiao10
developed a paddy rice mapping algorithm for the large-scale application in China using MODIS data. The method was
tested in South-West China and it focused on the detection of the critical phase of agronomic rice peculiarity such us
flooding, typical of this crop.
Starting from the work of Xiao et al.10 we intend to test its adaptability to the Mediterranean tempered rice context with
the specific aims to: i) implement this methodology and evaluate it in relation to the different European environmental
and agricultural conditions; ii) to make it able to different context respect the ones it was developed; iii) test a different
mapping approach base on rice phenological interpretation.
2. MATERIALS
2.1 Study area
We focused our study on the temperate rice cultivations of Mediterranean region. Figure 1a represents the study area that
covers the main European rice crop districts where more than 84% of the total European rice is produced13. These
regions, from West to East, are: Ebro river’s mouth in Spain (A), Rhone delta in France (B), the North-West Italian rice
district of Piemonte and Lombardia (C), the North-East Italian rice cultivation in Verona province (D) and the
agricultural area of Po delta (E). A second study area was also choice on regional scale, consider only on Northern Italy
rice districts (Figure 1b) where more detailed ancillary data were available. The Xiao method and its adaptation on
Mediterranean condition was evaluated on the entire MODIS tile that cover the study area (Figure1a) while the new
method based on phenological detection was tested only on the second study area (Figure1b).
2.2 Satellite data
Following the satellite data process described by Xiao et al.10, we downloaded the 8-day composite MODIS Surface
Reflectance product MOD09A1 from the United State Geological Survey (USGS) data base server14. Every composite
provides reflectance data on 7 spectral bands at 500-m spatial resolution. This product type is derived from a multi-step
process that consider atmospheric, clouds and aerosol corrections and accounts for each pixel the best reflectance data
registered during the time composite window. MODIS products are organized in a tile system with the Sinusoidal
projection grid, each tile of the grid covers an area of 1200 × 1200 km, more or less 10° of latitude and longitude, data
are organized in the Hierarchical Data Format (HDF). The entire study area is included in the MODIS tile H18V04, the
extension of MODIS data is reported in figure 1a. Information on surfaces elevation and slope were derived from Global
30-Arc Second Elevation Dataset product15 (GTOPO30). This dataset was used in the original procedure to mask out
agricultural areas considered not suitable for rice cultivation. During the adaptation phase of the algorithm we decide to
take in account other two most precise satellite products as external information source. These two additional layer are
the MODIS global permanent water bodies product, MODW4416. at 250 m spatial resolution, and the Digital Elevation
Model (DEM) product from Shuttle Radar Topography Mission (SRTM)17.
Figure 1 – The study area covers the most important Mediterranean rice districts. On the left side, in red colour CLC2006 rice layer
(1:100.000) highlights different districts contained in the Sinusoidal MODIS tile H18V04: Ebro’s mouth (A), Rhone delta (B),
Piemonte-Lombardia area (C), South-Verona area (D) and Po delta (E). In green colour on the right side, the second study area centred
on Northern Italy and considered for further most precise analysis is presented. Rice distribution is reported from a more detailed
regional land cover map (1:10.000).
2.3 Thematic reference data
For the accuracy assessment of our moderate-resolution map products, two different reference data set layers were
collected and used as ground truth. The first one is the European Corine Land Cover product18 for the year 2006 (CLC2006), the nominal scale of this thematic map is 1:100.000 and the minimum mapping unit is 25 Ha. This land cover
product has, at the third thematic level, a total of 44 categories including the “paddy rice” class (code 213). A second
more precise thematic cartography (1:10.000 scale) was created by merging four different datasets derived from the
available regional land cover cartographic services: Piedmont, Lombardy, Veneto and Emilia Romagna. This North
Italian Regional Land Cover (RLC), was used to validate outputs on the second study area where more than 200.000 ha
of rice are cultivated and almost 50% of the total European paddy rice is produced.
3. METHODS
3.1 Xiao approach for rice mapping
The Xiao et al.10 detection approach, is based on the identification of two diagnostic features of rice cultivation: the rice
transplanting/sowing moment that follows an intense agronomic flooding (well-known peculiarity of this farming
system) and the successive vegetative rapid growth event. According to these features, to perform rice detection it
becomes very important to be able to identify i) water surface condition (indicator of flood condition) and ii) vegetative
growth changes (indicator of crop grow) during the crop season. To perform this detection activity with remote sensed
data, several spectral indices were proposed. Time series of Land Water Surface Index19 (LWSI) were consider to
identify water presence and changes while Normalized Difference Vegetation Index20 (NDVI) and Enhanced Vegetation
Index21 (EVI) were used to describe seasonal vegetation behavior. The author10 performed a comparison of these indices
to identify the occurrence of flood/transplanting condition using the following criteria: LSWI +0.05 is greater than NDVI
or EVI values. In order to identify the rice rapid grow condition EVI must reach half of the maximum EVI value within a
temporal window of 5 8-date composite (about 40 days) subsequent to the previous flood event identification. When both
criteria are satisfy the pixel is considered a potential rice crop. Final rice map is obtained applying to the first detection
map several exclusion mask in order to reduce possible false positive detection.
Figure 2, shows five years of VI’s time series for a rice location in Italy. In general rice growing season is identified by a
Gaussian-like function of Vis. The graphs helps in understand the Xiao10 rice detection criteria. We can appreciate that
agronomic flooding conditions are always marked by a temporal exceeded of LSWI respect to EVI/NDVI indices and
how the rapid rice growing is described by EVI/NDVI increase.
Figure 2 - The seasonal dynamics of EVI, NDVI and LSWI+0.0.5 in a Italian rice site (Vercelli, lat. 45.2775N, lon. 8.4985E) for the
period 2008÷2012. According to the Xiao et al 10, rice culture is marked by a flooding, when LSWI+0.05 > EVI or NDVI (red
circles) and by a rapid growth indicated by VI increase (green rectangles). Black dot indicate presence of cloud contamination as
derived by MODIS quality flag.
Figure 3a reports a schematic flow-chart of Xiao et al.10 rice detection algorithm, in this paper it is mentioned as
“Original Method (OM)” in the following. The algorithm starts with a pre-process phase in which time series of spectral
indices (NDVI, EVI, LSWI and NDSI Normalized Difference Snow Index22) are calculated from the 46 8-day
MOD09A1 images. The successive analysis phase is aimed to detect potential flood areas and build the different
exclusion masks. Masks are made in order to exclude those features that could potentially negatively affect the seasonal
dynamics of vegetation indices (such us cloud or snow contamination) and those ones that could made commissions
error due to VIs temporal behaviour similar to the target (i.e. other vegetation classes). Cloud contamination is detected
in the time series for all pixels with blue band (B3) reflectance greater than 0.2 (20%) while snow cover is identified
when NDSI index is greater than 0.4 and SWIR band (B4) greater than 0.1. All data that present at least one of the
previous condition are removed from the time series. Several masks are designed to exclude land covers with temporal
series of VI similar to rice: “evergreen forest vegetation” is identified as those pixel having NDVI values greater or equal
to 0.7 in more than 20 (on 46) 8-days composites during the year; “evergreen shrublands, woodland and grassland
vegetation” areas are identified where LSWI value never grows up to 0.15 during the entire year; “permanent water”
mask excludes pixel in which NDVI were lower than 0.1 and lower than LSWI for at least 10 times (on 46) during the
year.
The finally detection phase of the algorithm allows to identified, as already mentioned, the two main agronomic features
of rice cultivation with the following steps. From potential flood time series, derived by comparing LSWI values to
NDVI or EVI, the data that match with the produced exclusion masks are eliminated obtaining the final flooding/rice
transplanting maps. Rice rapid growth time series are produced using the assumption that the EVI values reach half of
the maximum EVI value within the temporal window of 5 8-date composite (40 days). Finally, the pixels where both the
condition are satisfy are selected to made the final crop map. were also masked out the Areas with altitude greater than
2000 m and/or with slope greater than 2% are then masked out from the paddy rice map.
3.2 Adapting of Xiao method for rice temperate areas
After the implementation of the Xiao approach, we conducted a set of three levels of additional improvement on it. The
purpose was to optimize the OM on the study area developing an “Adapted Method” (AM) for temperate areas
condition. With reference to the labeled boxes on figure 3a, the first level of improvements (L1), was done adopting
different thresholds values on cloud and snow exclusion masks. For what concerns the cloud mask, after several tests we
decided to decrease the original threshold on the blue band from value 0.20 to 0.18. For snow mask, according to the
research conducted by Cea et al23, we decreased the original NDSI threshold to 0.3 value. The second level (L2) was
done modifying the algorithm exclusion masks, to adapt them to the land covers of the study area. “Permanent water
bodies” mask was modified assuming that a temporal VIs series can be representative of permanent water bodies when
NDVI is contemporary lower than 0.15 and LSWI in more than 85% of the 8-day MODIS composites (> 39 cases). For
what concerns “Evergreen vegetation” mask threshold was changed assuming NDVI value greater than 0.7 on more than
42% of the cloud free pixel of the temporal series (> 19 cases). “Evergreen shrub land, woodland, and grassland” masks
were modified using this rule: for this vegetation covers LWSI index never grows over 0.05 value during the temporal
series. Moreover, another mask was added in order to eliminate “wetland areas” that generally show a VIs profile similar
to paddy rice fields. The criterion adopted is that flooding condition (par 3.1) occurs on more of 90% of the temporal
profiles data (> 41 cases). The last level of improvement (L3) involves the use of external datasets and local knowledge.
We used a high resolution (90 m) Digital Elevation Model (SRTM-DEM) to identify hill/mountainous area were rice
agriculture is not possible. For each MODIS pixel at 500m the standard deviation of the altitude (SD-DEM) was
calculated from about 25 DEM pixels. We considered that rice can be cultivated only in agricultural flat areas that have
an elevation standard deviation lower than the threshold value of 9%. The second external data consisted in the
MODW44 permanent water bodies product at 250 m of spatial resolution. Pixels belonging to this layer were mask out.
Finally, the third criteria was based on local knowledge of agronomic practices in Mediterranean temperate areas. We
applied a temporal mask that excluded all the flooding detection recognized by the algorithm that occurs before the 5th of
March and after the 25th of June. From regional agronomic calendar and literature, the period March-June is the one that
presents real flood events, sowing practices and crop emergences. These three changes of L3 were added to the previous
two of L2 and L1.
3.3 Testing a new signal dynamic based approach
A “New Approach” (NA) based on the idea to perform rice detection by analyzing the continuous temporal signal of
vegetation indices was also tested. The aim of this approach is to detect paddy rice in a most consistent and flexible way
minimizing the dependency by local threshold adaptation. Figure 3b, shows a schematic diagram illustrating the consept
of the new approach. Preliminary step of this work was the computation of EVI and LSWI spectral indices time series
and the extraction from the HDF dataset of the blue band (B3) and quality flags values related to cloud contamination
information (HDF layer 12: 500 m State Flag; HDF layer 8: 500 m Reflectance Band Quality).
A cloud contamination time series was build consider informations derived from both quality flags and blue reflectance
data. The cloud mask information were then used in the VI-smoothing step to eliminate the noises data that affect VI
time series. A temporal smoothing on original EVI data was applied using a local polynomial function that weight
observation in relation to cloud contamination on the based on Savitzky–Golay filter24. This filter is specifically
suggested for the interpolation of VIs time series11. The signal filtering is a very important step because makes the signal
interpretable to retrieve not only the rice detection parameters but also phenological crop information.
The smoothed signal was then analyzed calculating his first derivative, this analysis allowed to automatically identify all
local (relative) minima (were the first derivative changes from negative to positive sign ) and maxima (were the first
derivative changes from positive to negative sign) points. Minima points related to rice crop season (RICE CROP
FLOOD) were identify when a relative minima occurs in correspondence of a condition of LWSI+0.05 greater than EVI
index. Whereas CROP MAX conditions, maxima points corresponding to rice peak, were selected only when the
following criteria were verified: i) RICE CROP FLOOD must occur in a temporal window of 56 to 120 days (7÷15
composites) before maxima point; ii) maxima point occurs after a rapid growth period interpreted as a sequence of at
least three positive derivative points in a temporal window of five composites; iii) a maxima is followed by a sequence of
at least three negative derivative points in a temporal window of five; iv) the senescence period after the crop peak must
produce a decrease of 1/3 MAX-EVI value within a temporal window of 40 days (5 composites).
Finally, following Boschetti et al.25 previous study, the smoothed VI signal was been analyzed to interpret two further
rice phenological features: Start of season period (SoS), end of season period (EoS). These two, were respectively
identified when occurs a 10% increment of VI index after the CROP FLOOD moment and when occurs a decreasing of
50% of the EVI index after the CROP MAX point.
DATA
DATA
MODIS 8-day composites of surface reflectance product (MOD09A1)
NDSI
NDVI
EVI
Wetland
Cloud
Permanent
water
L1
L2
MASK
EVI
LSWI
Cloud
Evergreen
vegetation
Snow
INDICES ANALYSIS
LSWI
QF
INDICES
B3
TEMPORAL SIGNAL ANALYSIS
INDICES
B3
MODIS 8-day composites of surface reflectance product (MOD09A1)
VI Smoothing
Derivative analysis
Relative
MIN
POTENTIAL FLOOD
DETECTION
Relative
MAX
POTENTIAL FLOOD DETECTION
RICE CROP FLOOD
RICE & PHENO DETECTION
RICE DETECTION
MAPS OF FLOODING AND RICE TRANSPLANTING
Expert knowledge
SD-DEM
RAPID GROWTH IDENTIFICATION
MODW44
L3
FINAL MAP OF PADDY RICE FIELD
A
CROP MAX
RICE DETECTION
SD-DEM
RICE PHENOLOGICAL STAGES
B
Figure 3 - Schematic diagrams illustrating different algorithms adopted for mapping paddy rice fields starting from 8-day MODIS
MOD09A1 satellite data. A) The Original Method (OM) from Xiao et al. (2005) approach and its adaptation (AM) to temperate area.
Changes (L1,L2 and L3) are highlighted by grey boxes of the figure. B) A new signal dynamic based approach (NA) based on the
detection of key phenological stages.
3.4 Evaluation of classification performances
The thematic maps produced in this work were evaluated deriving accuracy measures from error matrix calculation . The
use of the error matrix (or confusion matrix), is a standard methodology used to evaluate remote sensing derived maps.
Information about errors is computed by a pixel to pixel comparison between thematic map and reference map. Xiao’s
method (OM) result and its adaptation to Mediterranean condition (AM) were evaluated on the entire MODIS tile using
CLC06. The preliminary results on the new approach (NA), based on temporal signal analysis, were evaluated at
regional scale over the Italian rice district exploiting the spatially detailed information provided by the RLC thematic
map. On this second study area the performance of the three mapping methods were compared using the Pareto boundary
analysis approach27. Thematic cartography produced by automatic analysis of low/medium resolution satellite data can
be affected by the so called “low resolution bias” when compared to high resolution reference data. Low resolution bias
is the inaccuracy introduced by the difference in spatial resolution between high and low resolution data and not related
to the performances of classification algorithm. The bias is linked to the characteristic of the features on the ground and
it is a function of shape, size and fragmentation of the target under analysis. The Pareto Boundary proposed by Boschetti
et al.27 can be a method to assess the potential maxima accuracy that a classification method can achieve in a specific
experimental condition as a function of the actual fragmentation of the environment. The comparison of different
classification results respect to the Pareto boundary in the Commission Error (CE) – Omission Error (OE) space allows
to rank the algorithms performance. It is also possible to compare classification methods results in relation to the types of
error committed defining appropriate User’s cost function. A general formulation to compute different isoline cost
functions is provided in Boschetti et al27. In our case we select the function parameters in order to define the more
conservative method able to minimize CE.
4. RESULTS
4.1 Application and adaptation of Xiao Method to rice temperate areas
Figure 4a reports the rice detection map obtained implementing the OM algorithm; the results are able to highlight the
major European rice district of Italy (figure 4a), France (figure 4c) and Spain (figure 4d). Despite this good performance
in detection is evident a strong presence of positive errors (high CE) especially in the North part of the MODIS tile
where rice is not cultivated. Error matrix analysis for OM map reveals a very high overall accuracy (OA = 94%)
however due to the small presence of rice in the study area the User Accuracy (UA) is of only 3% as a consequence of
the high Commission Error (CE = 97%). The producer accuracy (PA) shows that only the 56% of the rice was correctly
identified corresponding to 44% of Omission Error (OE). Further analysis reveal that the CE is primarily distributed in
the following CLC classes: (CLC code 211) “not irrigated arable land” class, (36%), (CLC code 312) “coniferous forest”
(12%), (CLC code 311) “broad leaved forest” (8%), (CLC code 313) “mixed forest (4%)” and (CLC code 512 ) “water
bodies” (2%). Errors in these land use classes should have been minimized by the OM masking procedure; the results
demonstrated that the adopted criteria are not sufficient reliable in our case study. Figure 4b shows the map obtained
with the AM algorithm. The results clearly highlights a sensible reduction of positive errors. L1 adaptation, on snow and
cloud masking, did not show a significant CE decreasing while L2 stage of adaptation was able to provide a 5 points CE
reduction. This result is quantified in a decreasing of 3.821.000 Ha of falsely detected rice. Despite this improvement the
map still shows an high level of CE (92%). The last algorithm adaptation (L3) based on external information and on
local expert knowledge, reduced the CE of a further 21% increasing the UA to a 27%. We finally observe that it is
possible to make a local adaptation of the Xiao method to decrease the CE of the algorithm, however the masks
adaptation do not provide any additional improvement in the rice detection. This finding underlines the need to
investigate different detection approach.
a
c
b
d
e
f
Figure 4 – Results of rice detection in Mediterranean temperate area after the analysis of MODIS 8-day reflectance data (at 500m
spatial resolution) with the OM algorithm (a, c, d) and after the implementation of the AM algorithm (b, e, f). On figures c, d, e, f,
black lines are the rice extention derived from CLC 2006 and light-grey areas are the algorithm’s rice detection.
4.2 First results of new signal dynamic based approach
After the analysis conducted on OM and AM, we explore another approach based on vegetation indices temporal signal
analysis. The method is based on automatic identification of phenological stages from the continuous VI signal25 this
analysis allows to separate a crop signal, characterized by a strong seasonality, from the natural vegetated surfaces.
Moreover, the peculiarity of agronomic flooding in rice farming can be used, as already proposed by Xiao10, as a
diagnostic detection criteria to isolate this crop from the other cultivation. Preliminary results of this approach are
reported in figure 5, that shows the automatic rice detection for the Italian rice district without using external information
such as exclusion mask and temporal mask derived by crop calendar. Accuracy metrics shows an OA of 98% with a very
satisfying UA of about 87% as a consequence of very few commission errors (13%). The rice detection (green color) in
fact occurs mainly within the boundaries provided by the RLC thematic map (Figure5b). However the method,
completely automatic, resulted highly conservative showing a strong OE (68%). The Pareto boundary method (Figure
5b), allowed us to compare the different classification approaches respect the maximum potential accuracy that can be
achieved in the specific environment of the case study. Figure 5e shows that the target in our case study, the rice fields, is
characterized by small patches respect to the MODIS pixel size. This condition can produce a strong omission error due
to the low resolution bias. The Pareto analysis, according to Boschetti et al.27, reveals how the different methods fall in
region of reciprocal indifference in the OE-CE space because no method dominates the others. For instance, the NA has
an higher OE and a lower CE respect the others two and vice versa. In this condition it is not possible to identify which
method is the best without choosing an a-priori criteria in relation to the user attitude. In our case we prefer a
conservative method primarily able to reduce CE. Figure 5b reports in the OE-CE space the isolines of cost function
designed to prefer conservative classification results. In this way is possible to better solve the classification rank
problem according to user needs. Analysis of figure 5b shows that the AM and NA classifications belong to the same
isocost function and so provides comparable performances. However, it is important to underline that the NA method
was applied in an un-supervised way while the other requires local expert knowledge.
b
c
d
e
Figure 5 – Rice detection result after the implementation of the NA method on over the regional study area (a); a zoom on the
Piedmont-Lombardy rice district as result on OA method (c), AM method (d), and NA method (e) and the comparison between the
classification results of the considered algorithms (OM, AM and NA methods), using the Pareto boundary analysis method (b).
Moreover, the application of the signal dynamic based approach of NA allows to retrieve not only a rice map, but also
phenological stages maps. Figure 6, reports the four phenological rice detection maps, MIN-map (i.e. RICE CROP
FLOOD), SoS-map, MAX-map and EoS-map, obtained by applying this new algorithm on North Italy for the year 2006.
The rice phenological stages estimation fully matches with the rice crop cultivation calendar in Italy. In particular, we
observed that MIN points, interpreted as a flood condition with possible plant emergence, occurs in late March – April.
SoS points, interpreted as the start of plants rapid grow, occurs always on April – May. MAX points, in correspondence
to rice peak/flowering, occurs in July and finally EoS, proximal to the physiological maturity, occurs always on
September.
MIN
SoS
MAX
EoS
Figure 6 – Results on temporal distribution of the key phenological stage detected after the implementation of the NA method. MIN:
temporal stage close to rice emergence; SoS: temporal distribution close to the rapid growth of the culture; MAX: stage interpreted as
close to the rice flowering period; EoS: temporal distribution of the crop maturity.
5. CONCLUSIONS
This study wanted to evaluate reliability of automatic image processing methods for the identification of rice cultivated
areas and for monitoring of crop status and development. This study was addressed to evaluate the exportability of a well
know automatic rice detection method, developed in the tropics, to the different condition of temperate Mediterranean
environment. The method proposed by Xiao et al10 for Asian tropical areas have been studied and implemented to test its
performance on Mediterranean temperate area. This method was able to identify all the European rice district however
the map produced show very high commission errors. The application of this methodology is constrained by the choices
of specific thresholds for flood detection and exclusion masks used to reduce commission with other targets. Results
shows that a local tuning is required when applied to a different context respect the ones where it was developed. A
multilevel adaptation approach was tested to modify the original algorithm to the experimental conditions; results
demonstrated that it is possible to reduce false positive detection of more than 25 % in particular exploiting external data
and local expert knowledge. Finally, a new approach based on temporal series analysis has been tested. This method, is
based on the identification of peculiar moment of a crop through the analysis of temporal VI-signal, that can be
univocally related to a rice crop. With this approach is also possible to interpret seasonal rice crop features such as
flooding, start of the fast growing period, flowering and maturity. This method can detect rice crop area in a conservative
way with a limited use of fixed criteria and external exclusion masks. The final rice map produced has a UA of 87% and
provides very few commission error (< 13%). However the method shows a strong omission. Low-resolution thematic
maps, created at regional/continental scale, cannot achieve 100% accuracy due to the low-resolution bias. In particular,
European rice farming is characterized by small and discontinuous fields that potentially increase the low resolution bias.
The Pareto boundaries derived for the Italian rice district case study, shows maximum accuracy combination with a very
high omission error up to 90% in case of 0% CE. Pareto boundary analysis highlighted that the new method, applied in a
fully automatic way is comparable to the results of the Xiao approach when it is customized and adapted for the study
area. These preliminary results encourage to further improve the temporal signal analysis approach to perform rice
detection and further provide information on rice crop in order to move from a simple detection activity to a monitoring
system for rice farming. In conclusion, this work highlights that a proper processing of MODIS satellite data offers a
reliable solution to study and monitoring on large-scale rice cultivations and it is therefore a cost-effective tool for the
retrieval of spatially and temporally distributed information on the cropping system.
REFERENCES
[1] Verge, X.P.C., Kimpe, D., Desjardins, R.L., “Agricultural production, greenhouse gas emissions and mitigation
potential”, Agricultural and Forest Meteorology, 142, 255–269 (2007).
[2] Bouman, B.A.M., Humphreys, E., Tuong, T.P. Barker, R., “Rice and water”, Advances in Agronomy, 92, 187–237
(2007).
[3] http://ricestat.irri.org:8080/households/
[4] Shao, Y., Fan, X., Liu, H., Xiao, J., Ross, S., Brisco, B., Brown, R., Staples, G., “Rice monitoring and production
estimation using multi-temporal Radarsat ”, Remote Sensing of Environment, 76, 310–325 (2001).
[5] Chen, J., Lin, H., Liu, A., Shao, Y., Yang, L., “A semi-empirical backscattering model for estimation of leaf area
index (LAI) of rice in southern China”, International Journal of Remote Sensing, 27, 5417–5425 (2006).
[6] Zhang, Y., Wang, C., Wu, J., Qi, J., Salas, W.A., “Mapping paddy rice with multitemporal ALOS/PALSAR
imagery in southeast China”, International Journal of Remote Sensing, 30:23, 6301-6315 (2009).
[7] Fang, H., Wu, B., Liu, H., Huang, H., “Using NOAA AVHRR and Landsat TM to estimate rice area year-by-year”,
International Journal of Remote Sensing, 19:3, 521-525 (1998).
[8] Okamoto, K., “Estimation of rice-planted area in the tropical zone using a combination of optical and microwave
satellite sensor data”, International Journal of Remote Sensing, 20:5, 1045-1048 (1999).
[9] Van Niel, T. G., McVicar, T. R., Fang, H., Liang, S., “Calculating environmental moisture for per-field
discrimination of rice crops”, International Journal of Remote Sensing, 24, 885–890 (2003).
[10] Xiao, X.M., Boles, S., Liu, J.Y., Zhuang, D.F., Frolking, S., Li, C.S., William, S., Berrien, M., “Mapping paddy
rice agriculture in southern China using multi-temporal MODIS images”, Remote Sensing of Environment, 95 (4),
480–492, (2005).
[11] Peng, D., Huete, A.R., Huang, J., Wang, F., Sun, H., “Detection and estimation of mixed paddy rice cropping
patterns with MODIS data”, International Journal of Applied Earth Observation and Geoindormation, 13, 13-23,
(2011).
[12] Biradar, C,M., Xiao, X.M., “Quantifying the area and spatial distribution of double- and triple-cropping croplands
in India with multi-temporal MODIS imagery in 2005”, International Journal of Remote Sensing, 32:2, 367-386,
(2011).
[13] http://ricestat.irri.org:8080/wrs/
[14] http://glovis.usgs.gov/.
[15] http://www.edc.usgs.gov/products/elevation/gtopo30.html.
[16] http://www.glcf.umd.edu/data/watermask/
[17] http://www2.jpl.nasa.gov/srtm/
[18] http://www.eea.europa.eu/data-and-maps/data/corine-land-cover-2006-clc2006-100-m-version-12-2009.
[19] Xiao, X., He, L., Salas, W., Li, C., Moore, B., Zhao, R., “Quantitative relationships between field-measured leaf
area index and vegetation index derived from VEGETATION images for paddy rice fields”, International Journal
of Remote Sensing, 23, 3595– 3604 (2002).
[20] Tucker, C.J., "Red and photographic infrared linear combinations for monitoring vegetation", Remote Sensing of
Environment, 8, 127-150 (1979).
[21] Huete, A. R., Liu, H. Q., Batchily, K., vanLeeuwen, W., “Comparison of vegetation indices global set of TM
images for EOSMODIS Remote Sensing of Environment”, 59, 440–451, (1997).
[22] Hall, D. K.., Riggs, G. A., Salomonson, V, V., “Development of methods for mapping global snow cover using
Moderate Resolution Imaging Spectroradiometer (MODIS) data”, Remote Sensing of Environment, 54, 127-140
(1995).
[23] Cea, C., Cristobal, J, Pons, X,. "An improved methodology to map Snow Cover by means of Landsat and MODIS
imagery", Geoscience and Remote Sensing Symposium, 4217 - 4220 (2007).
[24] Chen, J., Johnsson, P., Tamura, M., Gu, Z., Matsusshita, B. and Eklundh.”A simple method for reconstructing a
high-quality NDVI time-series data set based on the Savitzky–Golay filter”, Remote Sensing of Environment, 91,
332–344 (2004).
[25] Boschetti, M., Stroppiana, D., Brivio, P. A., Bocchi, S., “Multi-year monitoring of rice crop phenology through
time series analysis of MODIS images”, International Journal of Remote Sensing, 30:18, 4643 – 4662 (2009).
[26] Jönsson, P., Eklundh L., “Seasonality Extraction by Function Fitting to Time-Series of Satellite Sensor Data”,
Geoscience and remote sensing, 40(8), 1824-1832 (2004).
[27] Boschetti, L., Flasse, S.P., Brivio P.A. (2004), “Analysis of the conflict between omission and commission in low
spatial resolution dichotomic thematic products: The Pareto Boundary”, Remote Sensing of Environment, 91, 280292 (2004).