Description
Woodchip samples with varying ferrihydrite surface coatings were incubated under different bulk redox conditions to investigate how iron availability and redox fluctuations affect localized redox gradients and hotspots. The spatial distribution of iron oxidation states in these samples can help identify reducing zones in the wood, with potential implications for understanding the effect of microscale redox heterogeneity on other redox-driven biogeochemical processes such as denitrification. However, resolving the spatial distribution of iron oxidation states is analytically challenging. Although XANES spectra can be used to infer iron oxidation states at individual points, they are not practical to acquire at sufficient granularity to develop two-dimensional maps of iron oxidation state and reveal spatial patterns of redox heterogeneity. Instead, we used multi-energy μXRF maps collected at five energies to register the equivalent of a five-point sparsely sampled XANES spectrum at each pixel. Using a limited number of XANES measurements from each sample, we trained a machine learning model to predict a proxy for iron oxidation state derived from the XANES spectra. The model’s features were constructed from the five-energy μXRF intensities at the pixel corresponding to each XANES measurement. The trained model was subsequently applied to the five-energy μXRF maps for all pixels to predict two-dimensional maps of iron redox-sensitive spectral variation in each sample. This approach can help extend sparse measurements into high-resolution spatial predictions, which could enable future investigation of microscale iron redox heterogeneity in complex environmental samples.