Background – Spatial Monitoring of Cropped Area

Famine has been a major focus for scientists studying semi-arid regions of Africa for many years. An information-based approach to food insecurity preparedness and planning successfully analyzes quantifies and evaluates agricultural and climate processes relating to food production (FEWS NET, 2004). Historically, remotely sensed precipitation and NDVI imagery have been used to identify areas with poor expected yields or pastoral conditions. This study describes some preliminary results that use MODIS NDVI imagery to estimate cropped area.

Production estimates can be derived by multiplying yield and area estimates. These production estimates may then be used to guide national food balance calculations and help determine food aid requirements. Numerically, we can write

(1)

where P is the food production over area A at average yield, Y. Crop yields are accurately estimated in semi-arid regions using grid-cell implementation of the Water Requirement Satisfaction Index (WRSI) model (Verdin and Klaver, 2002). The WRSI is driven by satellite rainfall estimates. Cropped area estimation, however, has yet to fully make use of such data and remains an arduous undertaking. This paper describes a technique to quantify cropped area using remotely sensed data. If successful this will allow more accurate predictions of food production, and consequently aid in the identification and mitigation of food security problems.

Study Area - Kenya

This study focuses on the country of Kenya, located in the Greater Horn region of Africa. Kenya's economy is heavily dependent on agriculture. “75% of Kenyans make their living from farming, producing both for local consumption and for export” (Kenyaweb, 2003). Kenyans employ both small-scale and large-scale farming, producing mostly maize and wheat. Because only 15% of Kenya’s land is suitable for farming – most farms are contained within Kenya’s Rift Valley – this study will focus on this area.

Goals - Variation about a baseline

The sustainability and ease of implementation of this effort are of principal importance, as the process should:

• Be repeatable on a yearly basis (applicable to specific dates within a growing season).
• Be easily disseminated to appropriate local representatives in developing countries.
• Be simplified and well documented for fast processing.
• Require inexpensive and readily available data to support repeatability and consistency.
• Be accurate and precise, effectively tracking average cropped area values, while capturing year-to-year variation in planting decisions and/or climatalogical effects on crop growth.

In order to achieve these goals the CHG/FEWS NET research group is exploring a combination of two distinct methodologies:

1. In the first step high-resolution satellite imagery (ETM+) will be used to create high quality maps of cropped area for a specific year. This step has high labor and data requirements, but will only need to be done once. The ETM+ interpretation phase will provide a baseline map, providing an accurate snapshot of cropped area for a given season in a particular country.
   
   
2. MODIS-based NDVI analysis will then be used to estimate year-to-year variations about this mean. This process is analogous to the combination of WRSI and historical yield data currently used in agro-meteorological modeling. This study focuses on the description of the MODIS-based crop estimation procedure.

Methods – Separation of crop signal using time series NDVI imagery

Imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS) is used (250m 16 Day NDVI composites), due to its strong adherence to all of the above requirements. In small cropping systems (abundant in semi-arid Africa), there are many farms and fields smaller than 250m2. Consequently, many of these pixels may lie on the edge of a small field, or contain parts of multiple fields, resulting in mixed pixels. Spectral Mixture Analysis (SMA) has been used extensively to unmix a pixel into its various components (Roberts, 1993) when target areas are significantly smaller than a single pixel.

SMA has typically utilized reflectance spectra, or the feature space, for 3-4 land cover types in a particular model. The Normalized Difference Vegetation Index represents greenness as a measure of the Near Infrared and Red band ratios (Sellers, 1985). I propose a new temporal unmixing approach, using the NDVI temporal signal to determine spectral contrast between cropland and other non-anthropogenic vegetation. An aggregate yearly cropped area value for a particular area can be attained using this technique.

Fractions of pure endmember pixels are determined for each pixel in an image, and fractional images are produced for each endmember. The shape of a temporal NDVI signature for various land cover types can be predicted based on its phenological characteristics (Figure 1). Temporal NDVI signatures are used in the SMA model to separate pixels into respective percentages of crop cover, non-anthropogenic vegetation, soil and water, ultimately estimating percentage of cropped area coverage.


Figure 1– Example of NDVI temporal behavior (phenology) for different land-cover types

Several independent data sources are simultaneously used to identify appropriate endmembers for the SMA model, including temporal NDVI profiles from the MODIS dataset, a maize density map and LANDSAT imagery over areas in the Sahel.

Data processing

Each MODIS composite scene was stacked temporally in a manner optimized to examine and separate 'temporal profiles', depicted in Figure 1. Figures 2, 3 and 4 illustrate the data smoothing algorithm that was performed. Each pixel’s temporal profile was examined and compared to 2 separate quality images. The profiles were systematically 'corrected' (linear interpolation) only if the NDVI quality imagery indicated a sufficiently poor value for the date in question. By incorporating NDVI quality information, the smoothing process successfully eliminated known “bad” data points, while preserving any authentic, albeit extreme, temporal NDVI trends within the data. Figure 4 illustrates the improvement in image quality. While the non-smoothed image on the right contains multiple data 'craters' (the red marks dotting the image), the smoothed image on the left has successfully eliminated these poor quality data points. These techniques may also provide improved NDVI monitoring tools.

Figure 2– Smoothing using NDVI 'quality' and 'usefulness' meta-data

Figure 3 – Spectral snapshot, smoothing using 'quality' and 'usefulness' meta-data

Figure 4 – NDVI Image before smoothing (left), after smoothing (right) 

Techniques – Spectral Mixture Analysis

Endmember selection

Several independent data sources were simultaneously used to identify appropriate endmembers for the SMA model, including temporal NDVI profiles from the MODIS dataset, a maize density map (high density areas shown in brown shading) and LANDSAT imagery over Kenya between the years of 2002 and 2004 (Figure 5).

Figure 5 – MODIS (upper left), Landsat (upper right), temporal NDVI spectra (bottom left) and sample selected endmembers (bottom right) in Kenya’s Rift Valley area. Maize high-density overlay is shown in brown hatch over the MODIS image.

Pure Maize endmembers show expected phenological properties of a summer-winter dual crop cycle, and are clearly separable from forest and other non-anthropogenic land cover. Further analysis is needed to separate single season and dual season cropped areas, as well as areas sown and harvested at irregular times.

Preliminary Results

The results shown below are of a preliminary nature, and are being expanded and validated further. They are provided here simply as a qualitative assessment of the outlined methods.

Spectral Mixture Analysis was performed using 3 endmembers: maize, forest and shade.

Figure 6 SMA Result Image

Figure 6 represents the maize fractional result image in the Northwest Rift Valley area. The original maize density map is overlayed (cyan) for validation purposes. The model clearly identifies areas of highest density maize, and shows much promise for future quantitative analysis.

Future Work - Validation

The estimation techniques in this report will be further refined, using known characteristics of local cropping patterns to incorporate such factors as latitudinal variation and multiple growing seasons. Also, more exhaustive and quantitative validation will be performed using cropped area estimations derived from higher resolution data. Landsat images will be analyzed and used to compare with the estimates from MODIS analysis for multiple growing seasons and study areas. This will allow for validation of not only the spatial precision of the proposed technique, but its accuracy through time as well.

This study represents exploratory work. Further investigation is required – particularly in the areas of endmember selection and validation. As more is learned about the agricultural practices in the Rift Valley area, endmember selection will be modified to accurately depict the phenological properties of maize. Wheat, an increasingly important crop, will also be mapped. Large scale SMA has shown deficiencies when phenological characteristics vary greatly across large areas of study. The models presented in this study will be subset to account for localized effects.

If this study is validated, it will provide a valuable aid in estimating cropped area over large spatial areas that contain highly mixed areas of cropland and other vegetation types. These estimates may be used to develop more accurate and timely estimates of annual food production. Ultimately, this data may be used to assess food security stability. Countries at risk for food insecurity are in great need of timely and accurate crop production assessments. It is the aim of this study to provide decision makers with a reliable, low cost and simple method for determining cropped area, and, ultimately aid in the nutritional stability in countries throughout in the developing world.

References

FEWS NET Web Site. 12 Dec. 2004. Famine Early Warning Systems Network. <http://www.ed.gov/index.html>.

KenyaWeb Web Site. 2003. Kenyaweb.com, Inc. <http://www.kenyaweb.com>

Roberts, D.A., Smith, M.O., Adams, J.B. (1993), "Green Vegetation, Nonphotosynthetic Vegetation, and Soils in AVIRIS Data," Remote Sensing of Environment, 44: 255-269.

Sellers, P.C. (1985), “Canopy reflectance, photosynthesis and transpiration”, Inernational Journal of Remote Sensing, 6:1335-1372.

Verdin, J. and Klaver, R. (2002). “Evaluating The Performance Of a Crop Water Balance Model in Estimating Regional Crop Production.” Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings. 2002.