Crop Classification: Evaluation of Automating Crop Classification Methods (NIH/NCI)

ehasl research project Knowledge of the spatial distribution of specific crop types is important for many environmental and health studies. In many instances such studies need crop types maps over large geographical regions (e.g., multi-county, entire state) for multiple years in order to determine statistically significant relationships between environment and disease occurrence. For example, in a study of agricultural chemical use and occurrence of cancer, once location of crops can be determined, important parameters such as pesticide use can be estimated and incorporated into an environmental model for exposure assessment. Such maps covering extensive geographical regions can only be derived from satellite imagery such as Landsat TM.

Landsat satellite imagery has been collected since the early 1970's and has been successfully used to classify many different crop types. However, the classification process can be very time consuming using traditional methods (i.e., supervised, unsupervised). In general, traditional methods require a remote sensing analyst to interact extensively with the computer system during the image classification process. Supervised methods require ground truth data to be collected and manually entered and the resulting training statistics to be carefully validated and refined prior to classification. Unsupervised methods require the analyst to evaluate and label each of the resulting clusters using ground reference data. In addition, an accuracy assessment process is necessary to provide a confidence level. A more comprehensive description of common classification and accuracy assessment methods can be found in Lillesand and Keifer (1987). Classification of a single Landsat scene (approx. 170x185 km) can take several days to months depending on the complexity of the land cover types and imagery. New methods that automate the interpretation process are essential if we are to meet the needs of environmental research applications in a timely and cost effective manner.

Our long-term objective is to develop a crop type map for 66 counties in eastern Nebraska, the site of several epidemiologic studies of the association of pesticide use and cancer. Cropping patterns are to be mapped for several years, beginning in the early 1970's. This paper describes the method used to identify a single crop. We selected corn to demonstrate our method because it is the predominant crop in Nebraska and many other Midwest states. Corn has the largest area of all of the US crops (USDA National Agricultural Statistics Service, web site http://www.usda.gov/nass) and has the greatest use (pounds applied) of pesticides and fertilizers. Since the 1980's, more than 90% of corn acreage received nitrogen fertilizer and herbicide treatments. In 1992, herbicides accounted for 91 percent of all pesticides applied to corn. The digital map resulting from the classification method discussed in this paper represents the distribution of corn in three categories (highly likely to be corn, likely to be corn, and unlikely to be corn). Further work is planned in the near future to classify other major crop types in the region (e.g., sorghum and soybeans).

We have developed a method that significantly reduces the time required to identify major crop types within a Landsat image. Our method also provides a level of confidence for each individual pixel classified. Specially designed software was developed which uses agriculture statistics to automatically extract spectral training data from selected target areas on the satellite image. The entire satellite image is then classified based on a combination of training data and agricultural statistics. Current and historical agricultural statistics (e.g., total hectares of crop harvested) are readily available at county level using state or national agricultural statistical providers. These data are collected for major crops by regular site visits to a selected sample of fields throughout the growing season. For our purpose, agriculture data are used in two ways to automate the classification process. First, the data are used to identify an area within the image to collect training statistics for a particular crop type. And secondly, the agriculture data are used to automatically process the remainder of the satellite image on a county by county basis. The agriculture data are used in lieu of collecting ground reference data per traditional methods (e.g., field measurements). The Mahalanobis distance measurement is used in the final map product to provide a measure of confidence which is important for further modeling efforts.

To demonstrate the feasibility of this approach, we produced a map for a single crop type (corn), using a Landsat Multispectral Scanner image in eastern Nebraska. Thirteen counties (3.35 million hectares) were classified in less than 15 minutes. The resulting map classifies the land area as either 'highly likely to be corn', 'likely to be corn', or 'unlikely to be corn'. Ground reference data from three counties were used to assess the accuracy of our method. The resulting average classification accuracy of 89 percent is comparable to traditional methods.

Lillesand, T.M. and Kiefer, R.W., 1997. Remote sensing and image interpretation, 2nd edn (New York: John Wiley & Sons).

View the project in PDF Format ( 1.0 MB download )