Data Sources
The data contained in the STARS Geospatial Tool to Target Surface Water Irrigation for Crop Intensification is based on an analysis using Landsat 5, 7 and 8 imagery (Level 1T), free for download from http://earthexplorer.usgs.gov. We used the blue, green, red, near infrared and short wave infrared (SWIR) bands (30 m resolution). The South Western side of the study area is covered by Landsat path 138, row 44. The eastern side is covered by path 137, rows 44 and 45. The Landsat 5 and 7 scenes were calibrated to surface reflectance by the USGS. We calibrated the Landsat 8 images using the TOA-DOS approach. We also cross-calibrated Landsat 8 images using Landsat 7 imagery acquired within eight days before and after of the respective scene analyzed. Data on surface water salinity covering the period from 2002–2012 was obtained from the Bangladesh Water Development Board. Shape files of the most recent and reliable land elevation and soil salinity classes were collected from the Bangladesh Country Almanac (BCA) and Soil Resource Development Institute (SRDI 2000). The BCA land type shape file contains inundation classes including Highland, Medium-Highland 1, Medium-Highland 2, Medium-Lowland, Lowland, and Very Lowland. These reflect the depths at which flood water is encountered during the monsoon season (representing elevation class), i.e., no consistent floodwater, <90 cm, 90-180 cm, 180-275 cm, and more. Finally, road networks and other administrative features were obtained from Bangladesh’s Water Resources Planning Organization (WAPRO).
Cropland Identification through Remote Sensing
Cropland was identified and discriminated from other land uses using Landsat 5 scenes taken on either January 21 or 31, 2010. In late January, cropland we could easily separate crop land from forests, since vegetation cover on cropland is generally low at that time. Images were classified as cropland and "other" which included water, forest, urban areas and land used for fish production (ponds and ghers). To avoid misclassification due to calibration errors, raw images from the 2 Landsat paths were classified separately. We created segments with eCognition 9 (Trimble Navigation Ltd., Westminster, CO). Segments are image regions that are more homogeneous within themselves than with nearby regions and represent discrete objects or areas in the image. Each image region then becomes an analysis unit for which attributes can be measured and used during the classification.
The study area measured more than 3 million ha and systematic sampling of ground truth data for the cropland identification would have been a big endeavor. We therefore relied on high-resolution background satellite imagery available in ArcGIS 10.1 (ESRI, Redlands, CA) and visually classified > 250 segments for each class to create a training data set. High-resolution satellite imagery contains more detailed information than 30 m Landsat images or segments. We thus chose those segments for training for which the corresponding pixels in the high resolution images showed uniformity. This task was made easier by the fact that we had to identify just two classes, cropland and non-cropland. For each segment, we used these attributes: mean of the digital numbers of bands 1–5, as well as texture (all directions) (Haralick, 1973) and the normalized difference vegetation index (NDVI). Additionally, we calculated the ratio of the NIR band to the visible ones as follows:
Ratio NIR_Visible = ρband4 / (ρband1 + ρband2 + ρband3)
Subsequently, the Random Forest Classifier algorithm in WEKA (see Hall, et al., 2009) was used to generate the classification rules. Machine learning algorithms do not depend on normal data distribution assumptions and allow for lumping together distinct classes such as forest, water, urban etc. This reduces the effort needed to create distinct training classes. It also automatically chooses the relevant variables and discards the other ones. Once trained, the classifier was used to classify the remaining segments. Subsequently, a visual quality control of the automatically classified segments was conducted, again using high-resolution background imagery as a reference. Wrongly classified segments were manually reassigned.
Identification of Waterways and Surface Water Duration
We used Landsat 5 images to identify waterways (rivers, canals) acquired on October 26, 2009 and November 8, 2011, coinciding roughly with the end of the monsoon when waterways are at their maximum extent. The same approach as for the classification of cropland was used. Some waterways in the study area are ephemeral. We therefore checked for the presences of water in rivers, canals, and creeks using the Automated Water Extraction Index (AWEI; Feyisa, et al., 2014) with atmospherically corrected Landsat 8 images from March 21 and 30, 2014. AWEIsh was chosen because of its effectiveness in improving water extraction accuracy despite the presence of shadow resulting from trees lining rivers, canals, and water bodies. Using the same threshold as described in the Feyisa et al., (2014), AWEIsh values > 0 were assumed to be water pixels and values below 0 non-water pixels.
Assessment of Land-use Intensity and Fallow Land Identification
Land-use intensity was determined on the base of a total of 44 Landsat 7 and 8 images acquired between 31 December and April 10 of 2011-12, 2012-13, and 2013-14. Crop productivity mainly depends on the amount of light that is intercepted by a crop during its life cycle (Monteith and Moss, 1977). The enhanced vegetation index (EVI; Huete et al. 2002), is a measure of the quantity of light intercepted for photosynthesis. It was successfully used by Schulthess, et al. (2012) to predict maize yield at the field level in Bangladesh.
We therefore measured the intensity of crop productivity by quantifying the maximum EVI reached by the most widely grown field crops in the study area, including lathyrus, fallows, wheat, mustard, mung bean, Boro rice, and maize. We extracted EVI trends from at least 10 known fields for each of the above crops in each of the 3 years. Following extraction, EVI values for each of the main crop types were plotted as a function of the number of days before or after January 1, until the 100th day of the year upon which the observation in question was made, corresponding roughly to the first two thirds of the Rabi dry season. We grouped each of the cropland types into three intensity classes, including (1) fallow land (2) low-intensity cropland, comprised of lathyrus, lentil, and mungbean, neither of which are typically fertilized, weeded, or irrigated, and which are broadcasted resulting in sub-optimal crop stands, and (3) high-intensity cropland, including wheat, Boro rice, maize, and mustard, all of which are more intensively grown with higher crop management intensities.
After checking for normality and homoscedasticity, we subjected data from the date upon which the maximum EVI value (corresponding to maximum LAI as a measure of peak productivity) was observed in each class to a one-way ANOVA using JMP 8.0.2 (SAS Institute Inc., Cary, NC) for the 2011-12, 2012-13, and 2013-14 dry seasons. F-tests indicated significance (P<0.001) between classes in each of the three seasons analyzed. Separation of means with the Tukey-Kramer’s range test at α = 0.05 showed that the fallow, low-intensity, and high-intensity classes were consistently different and independent in each season.
Because of the significant differences between cropland use intensity classes, we then set thresholds to separate classes to be used for all subsequent EVI analyses. Thresholds were set as the mid-distance point between the lower boundary for the standard deviation of the lowest maximum EVI observation for the high-intensity cropland types, and the uppermost boundary of the standard error for the highest EVI observation for the low-intensity crop types. This conservative process was used to distinguish the low-intensity and fallow crop classes for each season studied. The last step consisted of the extraction of the maximum EVI value for each pixel of the calibrated Landsat scenes for the entire study area in order to broadly map the three land-use intensity classes for cropland.
Creation of a buffer area around rivers, canals, and creeks
Since the efficiency of axial flow pumps decreases with lift height, and because they can only push water horizontally without gravity feed within a limited distance (Santos Valle et al., 2014), we created a 400 m buffer around those waterways in which water was present in late March. The 400 m width of the buffer was chosen as an empirical value, assumed reachable under most circumstances given feedback from irrigation service providers using the pumps. Intensive agricultural practices can result in sedimentation and nutrient loading of watercourses. Riparian buffers planted with species capable of ameliorating these problems could aid in mitigating the negative effects of crop intensification. We consequently reduced the 400 m buffer further excluding a 15 m strip adjacent to rivers and canals from cropping. This resulted in a 385 m wide buffer, potentially suitable for surface water irrigation.
Interpolation and Temporal Evolution of Surface Water Salinity Dynamics
Salinity concentrations in the Bangladesh tidal estuary vary in time, with salinity typically increasing as the dry season progresses. This results from the gradual reduction of southward river, canal, and creek water flow following the monsoon season, with important ramifications for irrigation water quality. To account for temporal changes in water salinity, we created four data sets based on the median of the observed data from the second halves of the months January to April over the 11-year period (2002–2012). Each data set was interpolated using Indicator Kriging to create a surface map of salinity. Salinity of river water is being measured at stations on the main rivers only. No data exist for the other water bodies. We therefore used kriging to give a good approximation of the salinity levels of smaller rivers, canals and creeks. Those maps were then classified into three water salinity classes, including 0–2 dS m–1 (high–quality), 2–4 dS m–1 (medium–quality), and >4 dS m–1 (low–quality).
Reclassification and Application of Soil Salinity and Inundation Land Types Shape Files
The publically available soil salinity map provided by SRDI (2000) comes with various classes, some of them being "mixed", i.e., a polygon may belong predominantly to one class, but may also contain data from another class. To simplify the analysis, we reclassified all data into three classes including <2, 2–4, and >4 dS m–1 by assigning the highest reported value in each class as the identifier for the new class.
Matrix of Land Suitability Based on Soil Surface Water Salinity
Since either high soil and/or surface water salinity are severe constraints for crop production, we created a matrix as shown in the Table below as a heuristic tool to simplify the analysis. These thresholds take into account crops that are rather salt intolerant, such as maize. Crop species and even cultivars within a crop species can vary greatly in their ability to withstand soil salinity (Ayers and Westcot, 1989).
|
Table 1. Analytical matrix to land potential for cropping surface water.
|
|
Salinity (dS m–1)
|
|
Soil
|
Water
|
|
|
0 – 2
|
2.1 – 4
|
> 4
|
|
0 – 2
|
High Potential
|
Medium Potential
|
Marginal Potential
|
|
2.1 – 4
|
Medium Potential
|
Low Potential
|
Marginal Potential
|
|
>4
|
Marginal Potential
|
Marginal Potential
|
Marginal Potential
|
|
|
Intersection of the Layers and Suitability Analysis
Cropland, EVI, surface water salinity, soil salinity, hydrozone, and land type layers were intersected to assess the suitability of cropland for sustainable intensification. Lastly, a subset of the land within the 385 m buffer was created. This resulted in a geospatial database that can be queried for extraction of descriptive statistics, which is the product of this analysis and the subject of this website.
Users can now query this database for appropriate irrigation technology targeting to plan irrigation instillations that will raise crop productivity and move farmers from rice-fallow or rice-low intensity systems into rice-high intensity systems.
Note: This methodology has been distilled from the following publication: Schulthess, U., Krupnik, T.J., Ahmed, Z.U., and A.J. McDonald. Technology targeting for sustainable intensification of crop production in Southern Bangladesh. The 36th International Symposium on Remote Sensing of Environment. Berlin, Germany. 15 May 2015. Available at: (Click here).
REFERENCES
Ayers, R.S. and D.W. Westcot. 1989. Salinity problems. In: FAO, editor Water quality for agriculture. FAO, Rome, Italy. p. 1-32.
Feyisa, G.L., H. Meilby, R. Fensholt, R. Simon and S.R. Proud. 2014. Automated Water Extraction Index: A new technique for surface water mapping using Landsat imagery. Remote Sensing of Environment 140: 23-35.
Huete, A., K. Didan, T. Miura, E.P. Rodriguez, X. Gao and L.G. Ferreira. 2002. Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment 83: 195–213.
Monteith, J.L., Moss, C.J., 1977. Climate and the efficiency of crop production in Britain [and Discussion]. Philosophical Transactions of the Royal Society B: Biological Sciences 281, 277-294.
Hall, M., E. Frank, G. Holmes, B. Pfahringer, P. Reutemann and I.H. Witten. 2009. The WEKA data mining software: An Update. SIGKDD Explorations 11: 1.
Haralick, R.M., K. Shanmugam and I. Dinstein. 1973. Textural features for image classification. IEE Transactions on Systems, Man, and Cybernetics SMC-3: 610–621.
Santos Valle, S., Qureshi, A.S., Islam, M.S., Hossain, M.A., Gathala, M.K., Krupnik, T.J., 2014. Axial flow pumps can reduce energy use and costs for low-lift surface water irrigation in Bangladesh. Cereal Systems Initiative for South Asia Mechanization and Irrigation (CSISA-MI) Project, Research Report No. 1. CIMMYT, Dhaka, Bangladesh
Schulthess, U., J. Timsina, J.M. Herrera and A. McDonald. 2012. Mapping field-scale yield gaps for maize: An example from Bangladesh. Field Crop. Res. 143: 151-156.
SRDI. 2000. Soil Salinity Bangladesh. Soil Resources Development Institute, Dhaka, Bangladesh.