Multi-scale digital soil mapping with deep learning

Abstract

We compared different methods of multi-scale terrain feature construction and their relative effectiveness for digital soil mapping with a Deep Learning algorithm. The most common approach for multi-scale feature construction in DSM is to filter terrain attributes based on different neighborhood sizes, however results can be difficult to interpret because the approach is affected by outliers. Alternatively, one can derive the terrain attributes on decomposed elevation data, but the resulting maps can have artefacts rendering the approach undesirable. Here, we introduce ‘mixed scaling’ a new method that overcomes these issues and preserves the landscape features that are identifiable at different scales. The new method also extends the Gaussian pyramid by introducing additional intermediate scales. This minimizes the risk that the scales that are important for soil formation are not available in the model. In our extended implementation of the Gaussian pyramid, we tested four intermediate scales between any two consecutive octaves of the Gaussian pyramid and modelled the data with Deep Learning and Random Forests. We performed the experiments using three different datasets and show that mixed scaling with the extended Gaussian pyramid produced the best performing set of covariates and that modelling with Deep Learning produced the most accurate predictions, which on average were 4–7% more accurate compared to modelling with Random Forests.

Mapping the Topographic Features of Mining-Related Valley Fills Using Mask R-CNN Deep Learning and Digital Elevation Data

“…Since no prior studies have attempted to map VFFs using automated methods, we are not able to relate our findings to any prior studies that have explored this specific task; however, our findings do reinforce those of Tier et al [6] and Behrens et al [48], which note the value of CNNs for extracting features from digital terrain data. More broadly, this study supports prior findings that CNNs in general and Mask R-CNN specifically are of great value for mapping features with a unique spatial, contextual, or textural signature and that may not be spectrally separable from other classes or features [51,52,84,85].…”

“…They reported mixed results, with some areas predicted well and other areas suffering from many false positives. Behrens et al [48] explored digital soil mapping using DTM raster data and DL and obtained a more accurate output than that produced by RF.…”

Modeling Depth of the Redox Interface at High Resolution at National Scale Using Random Forest and Residual Gaussian Simulation

“…Very recent advances in machine learning in the hydrological sciences promote deep learning, a machine learning method based on artificial neural networks, as a versatile scientific tool (Shen, ). While the adaption of deep learning in hydrology is slow, it already advanced related fields, such as soil sciences (Behrens et al, ). The increase in machine learning applications in hydrology and related fields is facilitated by readily available programming packages and a continual growth in availability, detail, and wealth of environmental data.…”

A hybrid approach for predictive soil property mapping using conventional soil survey data

“…Potential limitations or shortfalls of PSPM studies completed to date include (a) difficulty finding sufficient quantities of training and covariate data; (b) omission of many soil properties generally found in a conventional soil survey, like the soil survey geographic (SSURGO) database of the United States (e.g., salinity, erodibility, sodium adsorption ratio, gypsum, carbonates, sand fractions, and others) (Soil Survey Staff, 2018); and (c) inconsistent representation of uncertainty (Nauman & Duniway, 2019). Emerging trends show promise in improving PSPM, including the use of many more covariates, such as remotely sensed imagery summarized over large time series (Hengl et al., 2017; Maynard & Levi, 2017; Ramcharan et al., 2018), covariates and model learners that account for varying spatial scales of inference (Behrens et al., 2014; Behrens, Schmidt, MacMillan, & Viscarra Rossel, 2018; Behrens, Zhu, Schmidt, & Scholten, 2010b), and approaches that modify random forests (RFs) to better deal with inherent averaging bias (Hengl, Nussbaum, Wright, Heuvelink, & Gräler, 2018; Nauman et al., 2019; Nguyen, Huang, & Nguyen, 2015; Zhang & Lu, 2012). Because many studies rely on previously collected data that may not be a representative sample (Brus, Kempen, & Heuvelink, 2011), further exploration of patterns in model uncertainty may help fill sampling gaps.…”