Digital Soil Classification of Proximal and Remotely Sensed Data at the Field Level.

A fundamental requirement for effective soil management at the field level is an understanding of the spatial distribution of soil classes. Owing to the cost of soil data acquisition, proximal sensors (e.g. Electromagnetic (EM) induction instruments) and remotely sensed data (e.g. digitized aerial photos) are increasingly being used as surrogates. However, the use of ancillary data to identify soil management classes requires objective methods. In this paper we describe a methodology to first identify classes, using fuzzy k-means (FKM) analysis of two proximal sensors (i.e. EM38-v and EM31-v) and three spectral brightness bands (i.e. Red (R), Green (G) and Blue (B)) extracted from a digitized aerial colour photograph of an irrigated cotton growing field located in the Namoi River valley of Australia. The iteration of fuzzy exponents (?) and various indices, including the fuzziness performance index (FPI) and normalized classification entropy (NCE), enabled determination of ? = 1.4 and k = 4 soil management classes to be selected for further investigation. The classes formed contiguous units, or soil management zones, with Class A representing red sily clay soil profiles (Dermosols) along the eastern margin of the field, whilst Classes C and D demarcate the location of grey heavy clay soil profiles (Vertosols) in the southeast and northeast corners of the field, respectively. Class B is located spatially between these classes. Using fuzzy canonical analysis we find that the EM38-v and EM31-v signal data contribute most to the discrimination of Class D, whilst R contributes to the discrimination of Class A. Class C is distinguished from Class D based on having the lowest R and G. Validation of the FKM approach was confirmed by calculating average soil property (i.e. clay and sand content, cation exchange capacity (CEC), ECe, pH1:5 and exchangeable sodium percentage (ESP)) variance (i.e. S2Z) and the total within class variance (i.e. S2T) of a particular map (i.e. k = 2-8 classes). We conclude that the use of the FKM algorithm to classify proximal and remotely data, provides an understanding of the spatial distribution of soil management zones which differ owing to variable gypsum requirement (GR). Although a small parcel of the field was not well accounted for, given the high confusion index (CI) associated with it, we suggest that two options are available to better discern this area and include a) the expansion of the ancillary data set through the collection of additional proximal and remotely sensed data at the farm level or alternatively the inclusion of additional sensor data at the field level.