ESA UNCLASSIFIED - For Official Use
Spacial Explicit Rice crop calendar database from 14 Years of MODIS Imagery: patterns and trend in rice
seasonality preliminary results
Mishra B., Nelson A.
Framework • Why Rice
• Importance of information on crop seasonal dynamics Background
• Time series analysis
• PhenoRice approach
Use of MODIS data for rice monitoring • Mapping results
• Multi-year analysis : first test in Senegal
An Asia wide Rice Calendar for Asia (RICA) database • Maps
• Preliminary results of trend analysis
Rice is the world’s most important staple crop. It is
the second largest crop in terms of harvested area after wheat, but is by far the most important in terms of human consumption (FAOSTAT 2012).
Why Rice ?
Framework
Price control Sustainable development Food security Changes in rice production and availability can cause food crises and grain price variation
Meeting future increases in demand for rice is challenged by increasing climatic variability which limits
yields especially in developing countries.
Sustainable production should meet GHG emission and water use taking into account local cultural
importance
20 countries in Asia produce 90% (654Mt) of global production 87% (142Mha) of the area
Why do we need rice crop calendars?
Sustainable and climate smart production requires the adoption ofpractices that are tuned to local conditions. Appropriate planting (or
transplanting) and harvesting dates are part of these practices.
Knowledge of current dates is required to establish a
baseline of current practices and to propose improved practices
such as earlier planting or the adoption short duration or stress tolerant varieties to reduce the risk of yield reduction or losses during the
season.
Similarly, knowledge of these dates is important for crop
monitoring and early warning systems that provide policy related
information for national and regional food security.
Date information is fundamental for modelling the exposure of the rice crop to biotic and abiotic stresses and crop growth simulation modelling.
Why do we need RS
based calendars?
46 46 46 182 182 199 Laborte et al. 2017 RiceAtlas Dry season Wet season Busetto et al. 2019 MODIS based SoS Dry Dry Busetto et al. 2018 S1 based 2016 SoS Dry season Wet season SenegalRICA - RIce Calendar for Asia
Exploiting available rice phenology algorithm (PhenoRice) and multi-year MODIS data to generate Planting (SoS) and Harvesting (EoS) date estimates for major rice growing countries of Asia
Methodological solution to automatically handle pixel based and
continuous season detection to generate a spatial explicit and “consistent” database (RICA) at regular spatial aggregation scale (e.g. hexagon grid) Validate the approach by analyzing mapping spatial patterns and
comparing values with available reference data
Investigate multi-year dynamics (preliminary results)
Rice identification and dynamics: Phenorice
Starting from the work of Xiao et al. (2005,2006) , CNR-IREA and IRRI developed a tool for monitoring rice crop seasonal dynamics (Boschetti et al 2017)
Phenorice 1) detects rice cultivated areas and 2) analyses their phenology and agro-practices.
Methods
The method exploits a rule
based method which identifies a pixel as a rice crop when
1) a clear and unambiguous flood condition is detected
2) a clear crop growth signal is
Temperate
Tropical
Single season Double season Triple season
0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + + + ++ ++ ++ + + + Vietnam 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + ++ + Bangladesh 10 15 20 25 30 35 40 45 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + + + ++ ++ + + Philippines 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + ++ + Benin 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + ++ + Turkey 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + ++ + Spain Lat [° ] 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + + ++ + 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + + ++ + China 0 10 20 30 40 50 60 70 -2 0 0 0 2000 6000 year EVI + + ++ + India Egypt Winter crop and summer Rice Legend EVI NDWIb2.b6 NDFIb1.b5 NDWIb1.b6 NDFIb1.b7 Cloud contamination (B3 >0.18) MOD35 Cloud (State flags MOD09A1)
! ( !( ! ( ! ( ! ( !( ! ( ! ( ! ( ! ( ! ( ! ( MAD BRA USA VNMPHL BGD CHN BEN EGY TUR ESP 170°0'0"E 170°0'0"E 130°0'0"E 130°0'0"E 90°0'0"E 90°0'0"E 50°0'0"E 50°0'0"E 10°0'0"E 10°0'0"E 30°0'0"W 30°0'0"W 70°0'0"W 70°0'0"W 110°0'0"W 110°0'0"W 160°0'0"W 160°0'0"W 9 0 °0 '0 " 7 0 °0 '0 "N 5 0 °0 '0 "N 3 0 °0 '0 "N 1 0 °0 '0 "N 0 °0 '0 " 3 0 °0 '0 "S 3 0 °0 '0 "S 7 0 °0 '0 "S 7 0 °0 '0 "S + + + + Flooding Crop establishment Emergence Heading Maturity
Rice SoS Sites
Phenorice results and
performance
Overall: Y = 0.91 X + 8.8; r^2 = 0.97; ME = -2.6; MAE = 9.2 Dry: Y = 0.67 X + 24.1; r^2 = 0.63; ME = 2.4; MAE = 5.6 Wet: Y = 0.5 X + 105; r^2 = 0.24; ME = -10.2; MAE = 14.9 IND PHL PHL SN SNInput DATA
MODIS products (LP DAAC)
• 250 m 16 day composites VI products (MOD13Q1 MYD13Q1)
• 1km 8 days LST composite (MOD11A2 v.005) 42 Tiles x 14 years = 27,048 image
Reference data Rice Atlas
(Laborte, 2017)
Method
Download MODIS timeseries [42 tiles × 14 years]
Pixel level identification of rice SoS and EoS dates with
PhenoRice
Spatial aggregation of SoS
and EoS information Season identification per spatial unit
Frequency histogram of SoS and EoS per
hexagon* PhenoRice algorithm and cloud screening Gaussian fitting of number and timing of seasons MODIS tiles (42) EVI smooth NDFI SoS Emergence Cloud 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 -200 -100 0 100 200 300 400 V I [ -] DOY EVI raw PoS EoS
From pixel detection to RICA database
From pixel detection to RICA database
Problem:
- How to aggregate single pixel detection?
- How to define «seasons»? - How many?
- When start end?
Mixtools analysis
Automated statistical technique to identify “main modes” of the distribution of SoS
and EoS Rice Sesons
Rule 1: agronomic criteria minimum
time between two consecutive rice crops (> 80 days)
Rule 2: > 2% of the total number of
rice detections within the spatial unit
Rule 3: number of models for SoS
must equal the number for EoS
NO OK OK
Season identification and validation
Mix tool model resu lts CHN_Heilongjiang CHN_Guangdong IDN_Nusa_Tenggara_Timur IND_Gujarat_Anand VNM_South_East_Binh_Thuan IDN_Irian_Jaya_BaratResults
Planting dates
Detection between: 61 to 150 doy
Detection between: 151 to 240 doy
Detection between: 241 to 330 doy
Season 4 Season 2 Season 1
Season 3
Product assessment
Comparison of the planting and harvesting dates between RICA and RiceAtlas.
Each circle indicates a single (sub) region; the area of the circle is the log10 value of the rice area in that (sub) region corresponding to RiceAtlas.
SoS EoS
MAE = 28.8 R^2 = 0.80
MAE = 35.2 R^2 = 0.81
Work in progress
Assign each detection to RICA season using Bayes Decision
Boundaries method Timeseries MODIS data
PhenoRice Processing Timeseries SoS, EoS
Spatial Extension Hexagonal tessellation
Average seasons for each unit defined by RICA
Compute annual statistics of SoS and EoS for each season using a threshold
in all hexagons
Assumption:
• Number of rice pixels in 14 years should be GT 200 in a hexagon to compute SoS and EoS.
• Difference between the SoS of two seasons should be more than 80 days.
• At least seven estimates in the timeseries of SoS (EoS) is necessary to compute the trend for each season in a hexagon.
Data processing work flow
Trend and p-value computation using linear
Preliminary results
Orissa (ongoing changes due to monsoon rainfall anomalies ?)
Preliminary results
Trend Sign. RICA RiceAtlas < -2 > 3 Gumma et al 2015Asia wide rice crop calendar
We process 2003 – 2016 MODIS timeseries with PhenoRice
algorithm to generate pixel level SoS and EoS maps
Appropriate data handling was developed to generate regular
hexagon crop calendar by spatially and temporally averaging multi-year pixel level planting (SoS) and harvesting (EoS) date estimates
RICA V0
First Asia wide HR rice crop calendar reflects local spatial variations
Provides multi season information for each unit
@administrative level is in agreement with Rice Atlas
The dataset is under analysis in collaboration with IRRI to perform
expert base assessment to
validate it at local scale
identify potential artefacts
Perform further refinement (mask and specific region rules)
Conclusions and
Future Work
Multi-year analysis
2003 -2016 hexagon level estimates can be exploit for multi-year
analysis
A preliminary methodology was implemented to attribute each
detection to the more plausible crop season (need to trend analysis)
The dataset generated can be of help to identify rice dynamics
in relation to land use changes (crop intensification, expansion…)
consequence of agro-practices (e.g. varietal changes)
effect of climate/enviromental process (e.g. rainfeed monsoon)
Hot spot of anomalous behaviour
Assessment/validation undergoing
Exploitation of expert knowledge (collaboration with IRRI)
Investigation of publication and grey litterature