Agricultural Research Management Information System

Home
Research Summary
Search
About Us
- ARMIS
- Brochure
Contact Us
Report
User Request
Data Input
Help
- Operation Manual
  - PDF
  - Video
- Program Area & Commodity
We have reached 37600 number of research entries at this moment.

- Logout

Research Detail

Home
Research
Detail

Mapping Seasonal Rice Cropland Extent and Area in the High Cropping Intensity Environment of Bangladesh Using Modis 500 M Data for the Year 2010

Murali Krishna Gumma
International Rice Research Institute, Los Baños, Philippines

Prasad S. Thenkabail
U.S. Geological Survey (USGS), Western Geographic Science Center, Flagstaff, AZ 86001, United States

Aileen Maunahan
International Rice Research Institute, Los Baños, Philippines

Saidul Islam
International Rice Research Institute, Los Baños, Philippines Bangladesh Rice Research Institute, Gazipur 1701, Bangladesh

Andrew Nelson
International Rice Research Institute, Los Baños, Philippine

Abstract/ Executive Summary:

Rice is the most consumed staple food in the world and a key crop for food security. Much of the world’s rice is produced and consumed in Asia where cropping intensity is often greater than 100% (more than one crop per year), yet this intensity is not sufficiently represented in many land use products. Agricultural practices and investments vary by season due to the different challenges faced, such as drought, salinity, or flooding, and the different requirements such as varietal choice, water source, inputs, and crop establishment methods. Thus, spatial and temporal information on the seasonal extent of rice is an important input to decision making related to increased agricultural productivity and the sustainable use of limited natural resources. The goal of this study was to demonstrate that hyper temporal moderate-resolution imaging spectroradiometer (MODIS) data can be used to map the spatial distribution of the seasonal rice crop extent and area. The study was conducted in Bangladesh where rice can be cropped once, twice, or three times a year. MODIS normalized difference vegetation index (NDVI) maximum value composite (MVC) data at 500 m resolution along with seasonal field-plot information from year 2010 were used to map rice crop extent and area for three seasons, boro (December/January–April), aus (April/May–June/July), and aman (July/ August–November/December), in Bangladesh. A subset of the field-plot information was used to assess the pixel-level accuracy of the MODIS-derived rice area. Seasonal district-level rice area statistics were used to assess the accuracy of the rice area estimates. When compared to field-plot data, the maps of rice versus non-rice exceeded 90% accuracy in all three seasons and the accuracy of the five rice classes varied from 78% to 90% across the three seasons. On average, the MODIS-derived rice area estimates were 6% higher than the sub-national statistics during boro, 7% higher during aus, and 3% higher during the aman season. The MODIS-derived sub-national areas explained (R2 values) 96%, 93%, and 96% of the variability at the district level for boro, aus, and aman seasons, respectively. The results demonstrated that the methods we applied for analysing and interpreting moderate spatial and high temporal resolution imagery can accurately capture the seasonal variability in rice crop extent and area. We discuss the robustness of the approach and highlight issues that must be addressed before similar methods are used across other areas of Asia where a mix of rainfed, irrigated, or supplemental irrigation permits single, double, and triple cropping in a single calendar year

Keywords: Seasonal rice mapping, MODIS, NDVI, Cropping intensity, Spectral matching techniques, Field-plot information, Bangladesh

Location:

Start Date:

End Date:

Program Area: Socio-economic and Policy

Non-commodity: Modeling

Objectives:

1. Produce seasonal rice extent maps of Bangladesh for 2010. 2. Determine seasonal rice areas at the district level for all 64 districts and for all three seasons in Bangladesh. 3. Establish the accuracies of (a) MODIS-derived rice crop extents by comparing them with independent field data and (b) MODISderived rice crop areas by comparing them with independent sub-national statistics.

Methodology:

Methods The seasonal rice mapping methodology involved the following ten steps: (3.1) Temporal series of reflectance data. (3.2) Generating class spectra by performing an unsupervised classification on NDVI MVC. (3.3) Composing an ideal spectral data bank. (3.4) Grouping classes with a decision tree algorithm. (3.5) Grouping classes with spectral similarity values. (3.6) Spectral matching techniques. (3.7) Identifying and labeling classes. (3.8) Resolving mixed classes. (3.9) Sub pixel area estimation. (3.10) Accuracy assessment. 3.1. Temporal reflectance data and NDVI data A 315 layer stack of reflectance data for all seven bands for across 46 weeks was generated using MODIS imagery (Thenkabail et al., 2009). The 8-day NDVI images were prepared by using Eq. (1) and monthly MVCs of NDVI for January through December (12 layer stack) was prepared using Eq. (2). 3.2. Generating class spectra by performing an unsupervised classification on NDVI MVC Class spectra were generated through an unsupervised ISODATA cluster algorithm on the 12-band monthly MVC NDVI. Unsupervised classification was used instead of supervised classification in order to capture the range of variability in phenology over the study area, particularly in large study areas where the NDVI signatures of most of the potential classes are unknown. The unsupervised classification was set at a maximum of 100 iterations with a convergence threshold of 0.99 (Leica, 2010). ISODATA classification using progressive generalization led to an initial 100 classes (Cihlar et al., 1998). The MODIS NDVI time-series spectra were then plotted for each of the 100 classes for labeling. 3.3. Composing an ideal spectral data bank Ideal spectral signatures were generated using time-series data that were extracted from 118 observation points. Each of the points chosen to generate the ideal spectral signatures represents a definitive crop type and/ or cropping system such as ‘‘irrigated-groundwater-rice-rice-rice’’ (meaning the rice field is irrigated by groundwater and is rice during all three seasons), ‘‘irrigated-groundwater-rice-fallow-rice’’, or ‘‘deepwater-rice-fallow-water’’. Multiple points with the same crop type/system, even though distributed spatially in discrete patches were combined to create a single ideal spectral signature, for that cropping system (between 5 and 17 points per spectra) resulting in 15 ideal rice signatures and a 9 ideal signatures for other classes. 3.4. Grouping classes with a decision tree algorithm A decision tree was applied to the 100 NDVI signatures obtained from 100 classes that resulted from the unsupervised classification to obtain twelve distinct groups. The decision tree is based on monthly NDVI thresholds at different crop growth stages in the season. The months and threshold values were chosen based on knowledge of the crop calendar from local experts, field observations as well as published rice crop development stages. 3.5. Subsequent grouping of classes using spectral similarity values There are several spectral matching techniques (SMTs) (Thenkabail et al., 2007) to reduce the grouping into similar land use classes, and in this case we selected spectral similarity value (SSV) (Homayouni and Roux, 2003) that has previously performed well in analyzing spectral signatures for agricultural crops such as rice (Thenkabail et al., 2007). SSV was calculated for each class combination, 3.6. Spectral matching technique Classes with similar SSVs were grouped and then matched against ideal spectra. The 100 classes obtained from the unsupervised classification include crop and non-crop lands. Each of those classes was investigated and grouped into similar or near-similar broad classes. We use an example to illustrate the process base on nine similar cropland classes obtained by the DT algorithm in Fig. 4b1, Section 3.4. The nine classes (class numbers 24, 40, 41, 45, 47, 50, 52, 54, and 59) have similar or near-similar signatures as determined by their SSVs. The nine classes are then matched with the nearest ideal signature. This resulted in three class spectra (classes 41, 45, 47) matching perfectly with ideal spectra ‘‘4’’, which is labeled as ‘‘irrigated-surface water – double crop – rice in boro season – fallow in aus season – rice in aman season – large scale’’. The same process is followed for all cropland classes until all class spectra are matched to ideal spectra. 3.7. Identifying and labeling classes The combination of decision trees and spectral matching allows for rapid and accurate identification and labeling of classes as illustrated. However, further affirmation of the class labeling requires steps B through D especially in cases where we did not have a sufficiently rich ideal spectral data bank. Whenever there was ambiguity in the class matching we used various sources of information to increase our confidence in the matching decision. We performed visual interpretation of the phenology from the 8-day NDVI and LSWI time-series to distinguish between irrigated and rainfed systems or to confirm deepwater systems for example. We also relied on visual interpretation of high resolution imagery from Google Earth (where available) to confirm the presence of any rice bunds or irrigation structures. Finally we referred back to relevant information from our field plot data to correctly class match the class. 3.8. Resolving mixed classes When a study area contains many distinct land cover classes over a large spatial extent, there is a risk that some of the classes from the unsupervised classification may contain several sub-classes or mixed classes. These mixed classes were resolved by extracting them from the stack, reclassifying them, and applying the methodology above on these new classes in order to separate them. 3.9. Sub-pixel area estimation With the use of moderate spatial resolution imagery in areas where land use patterns change over sub-pixel distances, it is inevitable that many MODIS pixels will contain more than one land cover class. The labeling of classified land cover maps at this resolution suggests that each pixel in that class is 100% pure, when this is certainly not always the case. One approach is to use higher resolution imagery with spectral and spatial resolutions capable of accurate rice area estimation. This requires a sample of imagery across the study site that captures representative rice classes in each season. Since this was beyond the scope of this study we estimated the sub-pixel rice area for each rice class from the 191 detailed ground data observations following previous methods (Thenkabail et al., 2007). The ground data observations include a visual estimate of the proportion of the 500 m 500 m area that surrounds the observation point under different land use (water, built-up area, cropland, etc.). If our ground data observations are representative of the rice systems and we have sufficient observation points per class, then we can estimate a reliable rice area fraction (sub pixel area, or SPA) for each class based on the average rice area across all observation points in that class. The SPA information is applied to each class to estimate the actual rice area for that class. This SPA rice area estimate is compared with the published sub-national rice area rather than the MODIS pixel rice area. As seen in Section 3.2, the number of field points per rice season is not proportional to the published rice area statistics. If we assume that the statistics are an accurate assessment at district level – an important assumption for our area accuracy assessment – then care must be taken when interpreting the SPA for those under represented classes.

Source of Information: ISPRS Journal of Photogrammetry and Remote Sensing Volume 91, May 2014, Pages 98-11

Article Link (Online Journal):

Funding Source:

Budget:

Total Budget (Tk.) :

Achievement/Findings:

Decision trees and supervised class labeling to map seasonal rice areas using hyper-temporal 500 m MODIS NDVI time-series data and intensive field-plot information. The accuracies of the rice area for each crop season [(boro (December/January–April), aus (April/May–June/July), and aman (July/August–November/ December)] were determined by correlating the MODIS-derived sub-national (district-level) seasonal rice area statistics with the Bangladesh Bureau of Statistics sub-national statistics. The R2 values were 0.96 for boro rice, 0.93 for aus rice, and 0.96 for aman rice. These statistical results also showed that the MODIS data overestimated rice area by 6% for boro, by 7% for aus, and by 3% for aman relative to the sub-national statistics. The overall accuracies of the five rice classes, during the three seasons, varied from 78% to 90%. However, rice versus non-rice accuracies exceeded 90%. Almost all intermixing was only between rice classes. The remote sensing based cropping intensity of rice determined in this study for Bangladesh was 149% across the country and was found to be 26% lower than previous non-remote sensing estimates. Mapping seasonal rice areas is the first step in characterizing important rice-growing environments for sustainable development and livelihoods. Precise up-to-date seasonal rice maps and statistics such as these are important inputs for assessing the impact of abiotic stresses such as droughts and floods, which regularly affect the region and are predicted to increase in frequency and intensity in a changing climate. This approach was appropriate for accurate identification rice systems, rice cropping intensity and rice area estimates in most rice growing environments and seasons in Bangladesh. We documented the problem areas and discussed the possible shortcomings of the method as well as suggesting adjustments to the methodology in light of the findings of this study and forthcoming sensors. We suggest that this methodology can be improved and adapted for mapping rice in other countries where cropping intensity is high and where rice cultivation is extensive, including much of South, South East and East Asia where much of the world’s rice is grown. The research makes a broad contribution to the methods and products of the Group on Earth Observations (GEO) for monitoring agriculture areas, Agriculture and Water Societal Beneficial Areas (GEO Agriculture and Water SBAs), the GEO Global Agricultural Monitoring Initiative (GEO GLAM), the global cropland area database using Earth observation data, and studies pertaining to global croplands, their water use, and food security in the 21st century.

Knowledge Management: Journal