A decision tree approach to predict humification of organic soils

Peat and other organic-rich soils are an important global carbon store with more carbon being held in northern peatlands than in the atmosphere at any one time. Disturbance such as erosion or peat mining creates aerobic conditions which lead to the loss of carbon to the atmosphere as the now exposed organic matter decomposes. This decomposition can be expressed as the degree of humification. Earlier work with HyMap and the ASD spectroradiometer has shown that the degree of humification of peat can be estimated spectrally, but that water masks the humification signal in the shortwave infrared (SWIR), related to relative amounts of residual lignin, cellulose and other biochemical components (McMorrow. et al., 2004). This paper identifies the threshold at which humification and moisture signals are separable in simulated HyMap reflectance spectra. It then proposes a two-stage decision tree approach; first stratifying samples into moisture classes based on moisture-specific spectral indices, then applying a second set of indices to predict humification separately for each moisture class. The indices were developed using a controlled laboratory drying experiment on 34 blanket peat samples from the Pennines, UK. Samples of exposed blanket peat representing the range of humification were saturated with water and gradually oven-dried in 20 stages until they were totally dry. Gravimetric water content and ASD spectra were recorded at each drying stage on each sample. HyMap spectra were simulated from the ASD spectra. Humification was measured using the colorimetric method in which the transmission of light at 624 nm through a solution of peat in sodium hydroxide was recorded. Transmission is inversely proportional to humic acid content, a commonly used measure of the degree of humification. Eight spectral indices were tested to predict moisture content, namely single band reflectance, normalised difference reflectance, band slope, band ratio, first derivative of reflectance, second derivative of reflectance, depth below the continuum and factor analysis. Inverse forward stepwise regression was run to predict the moisture content for 60% of the samples. The models were tested with the remaining 40% of samples. The second derivative at band 86 (1710 nm) predicted moisture with an accuracy of 92% and avoids the A SWIR atmospheric water absorption bands. Samples then were divided into six predicted moisture levels (>50,