Feature selection and similarity analysis of crop fields for mapping organic carbon concentration in soil

High-resolution, remotely sensed imagery of bare soil has been used to quantitatively map the spatial variation of soil organic-C concentration (SOC) with a greatly reduced cost compared with grid sampling. However, the procedure requires that each crop field must be sampled and mapped separately. Examination of a remotely sensed image revealed that some fields had similar image properties. If similar image properties also indicate the similarity of field properties such as soil organic carbon distribution and crop residue amount, fields could be placed in groups with similar image properties. Therefore, maps could be developed for a group of fields instead of a single field. The cost of mapping could be further reduced in proportion to the number of fields in the group. In this study, 50 rectangular areas were selected from a 1999 digital orthophotograph to examine different features for use in measuring field similarity. Three features, including color histograms, color slopes, and wavelets, were extracted from the 50 rectangular field areas. Two similarity measurement methods, statistical clustering with Euclidean distance and the Ward neural network system, were tested for grouping fields with similar image characteristics. The Ward neural network system, using the color histogram feature, gave the best result for grouping similar fields. Based on this result, two bare surface fields were selected from a NASA ATLAS image. Soil samples were collected from the two fields. The similarity of the two fields was computed with the Ward neural network system using the color histogram features. Because of the high similarity between the two fields, maps of soil organic carbon concentration for the two fields were developed with a single mapping procedure. The resulting maps were checked based on a second set of soil samples that were different from that used in model development. There was a good agreement with an r2 value of 0.81.