Crop Mapping of Kent County, Maryland

Introduction

The first goal of the 1983 Chesapeake Bay Agreement was to reduce by 40% the amount of nutrients (Nitrogen, Phosphorus) entering the Bay by the year 2000. This would improve the oxygen levels in the Bay waters and allow aquatic life to flourish. Nutrients are contributed by point sources, such as sewage treatment plants, and non-point sources, like septic fields and agricultural land. Water treatment plants were the first to be regulated, but the passage of Maryland's Nutrient Management Law signaled that agriculture was next.

In order to better understand how nutrients, in the form of synthetic fertilizer and animal manure, travel from agricultural fields to the Bay, nutrient modelers use land use maps as a source of data for their models. Land use maps traditionally give broad definitions of agricultural land, with the only distinction being made between row crops and pasture. However, different species of row crops handle nutrients differently. For example grasses, such as corn, barley, wheat, and rye, receive supplemental nitrogen, and legumes, such as soybeans and alfalfa create their own. From a nutrient modeling perspective it would be useful to know where fields planted with crops that usually receive additional nutrients are located.

Corn field on left, soybean field on right.

Methods

The focus of this study was to determine the spatial distribution of corn, soybean, and wheat fields in Kent County, Maryland in 1998.

Image from July 1, 1998 Landsat of Kent County

Unlike other types of land use, the species present in agricultural fields may change dramatically from year to year, or even within the same year. This creates some unique challenges, and brings about the need to understand crop development schedules. Even anticipated trends can deviate as a result of weather, commodity prices, and management practices. One of the most common scenarios is the "double-cropping" of wheat and soybeans, meaning that two crops are grown on the same field during the same growing season, and the appearance of one or the other depends on what time of year the field is observed. Another common practice is crop rotation, in which corn would be planted in year one, and then wheat and soybeans would be double-cropped in year two. Although these practices are widespread, they are unpredictable because of the reasons mentioned above.

As a starting point, crop development calendars should be considered. The Maryland Agricultural Statistics Service provides weekly crop progress and weather reports from April through November every year. These reports contain information about the percentage of each crop that has been planted and harvested, as well as the general condition of a crop (poor, fair, excellent, etc.). A statewide crop calendar can be generated using the reports, but it is important to keep in mind that crop calendars will vary by region and by year.

Cumulative crop progress calendar for corn.

Crop calendars are useful for determining when greenness can be observed in different crops. A very simple measure of "greenness" is the normalized difference vegetation index (NDVI). A plot of the NDVI of the different crop types shows how their spectral appearance varies throughout the year. The spectral information comes from the Landsat family of satellites, which collect imagery at a spatial resolution of 30 meters every 16 days.

Graph comparing greenness for different crops

An NDVI uses two bands, but each date of Landsat imagery collected contains data from six bands. In order to utilize all available information, a more complicated classification method was used. The method used here is a "maximum likelihood classification" (MLC) which is a type of supervised classification. It is called "supervised" because it uses predefined training sites to generate numerical signatures for each class (corn, soybeans, and wheat). The signatures from the training sites are compared to the value of each pixel in the scene, to determine into which class the pixel belongs. The result is a classified image, identifying all pixels belonging to each class. There are other classification methods, but an MLC is widely used and well documented in image processing literature.

The training sites for this study were from three regions in Kent County. Actual field observations from 1998 were not available, but training sites were inferred by looking for fields exhibiting the spectral characteristics of each class. Landsat derived spectral characteristics of corn, soybeans, and wheat were identified during the summer of 1999 and the spring of 2000. Polygons of the fields were superimposed onto Landsat imagery, and the pixel values from within the polygons were extracted and used to generate the spectral signatures. Using this method, 33 corn fields, 22 full-season soybean fields, and 31 wheat fields (double cropped with soybeans) were identified.

Image of training site locations

After looking at the crop calendars, it was apparent that not all crops could be identified using a single observation date. So how many dates of imagery are needed to produce an accurate classification, and which dates used in combination are best? In 1998, there were three cloud-free Landsat image acquisitions made over Kent County: April 12th, July 1st, and August 2nd. A classification using four possible combinations of data produced some encouraging results.

Results

Since it was impossible to know exactly what was planted in every field in Kent County in 1998, one way to determine the accuracy of the classification was to compare the total acres classified in each class, to the Maryland Agricultural Statistics Service publication of total acres harvested in each class ("wheat" class reported here includes other small grains, such as barley and rye). It would be helpful to know the accuracy on a field by field basis, but that kind of data was not available.

MLC of April 12 + July 1 + August 2
Corn acreage underestimated, but accuracy of known features is good.

MLC of April 12 + July 1
Corn acreage underestimated, soybeans and wheat acreage overestimated, some known features are identified incorrectly.

MLC of July 1 + August 2
Acreage of all classes overestimated, many features identified incorrectly.

MLC of April 12 + August 2
Acreage estimates closest to NASS figures.

Conclusion

The classification results can be explained by considering the relationship between the periodic biological phenomenon and the climate conditions (phenology). In the April 12 observation, wheat acres are identified because they are the only class that is green. In the August 2 observation, wheat has been harvested, and corn and soybeans are spectrally unique due to differences in leaf and canopy geometry. In the July 1 observation, other vegetation (alfalfa, pasture, and forest) can be separated because they have a different development pattern than the classes of interest.

Graph of image acquisition dates and phenology highlights.

Discussion

This methodology showed promise for making an accurate classification map, although there are still several issues that should be addressed. It would be helpful to nutrient modelers to have other agricultural land, such as alfalfa and pasture, included in the classification. If the field polygons for the whole county could be identified before the classification took place, it would probably result in a more accurate acreage estimate. If a unique classification could not be produced every year, a probability matrix based on previous classifications would be useful to figure out what is there. This would also be useful when trying to do predictive nutrient runoff modeling.

There are also other methods that could be used to produce a classification. One of the most promising is the "decision tree classification" which allows for the inclusion of non -parametric data (variables without a distribution), and mixing between continuous and categorical data. The purpose for using the numerous techniques mentioned in the above section is to produce an accurate classification of agricultural land. Unlike other categories of land use, the species present in agricultural fields and nutrient management practices are constantly changing. Considering that agricultural land constitutes a majority of the land use in many counties in the Chesapeake Bay Watershed, it is important that an accurate method for monitoring this land is available.