Application of machine learning for soil survey updates: A case study in southeastern Ohio.

摘要

Future activities of the National Cooperative Soil Survey include SSURGO (Soil Survey Geographic Database) updates on a Major Land Resource Area (MLRA) basis. To further soil survey updating, machine learning techniques were used to build predictive soil-landscape models for two counties (Monroe and Noble) within the central Allegheny plateau (MLRA 126) in southeastern Ohio. Characterized by their similar soil forming factors but a varied survey intensity and vintage, these two counties provided a setting ideal for building and testing the application of the models. Twenty five different environmental correlates including 10m resolution raster coverages of terrain and its derivatives, climate, geology, and historic vegetation were used as predictor variables for soil class.; Randomly sampled points proportionate to the area of the different soil classes from the published soil survey of Monroe County (SSURGO) were used to train the soil-landscape model. Since map units can contain more than one component soil series, each sample point within a map unit can possibly belong to any one of them. Hence there is ambiguity in labeling of the training instances with appropriate soil series. A kNN-based heuristic approach was used to disambiguate the training set labels. The training sets were further preprocessed for removal of outliers (to reduce noise) and for selection of fewer attributes (to avoid redundancy and irrelevancy) in the data. Modeling was performed using two learning algorithms namely J48 classification tree and Random Forest (RF). The map models were then evaluated for the quality of prediction using two prediction rate measures (one based on the dominant soil series and the other based on component series probabilities) and two landscape fragmentation statistics (contiguity index and aggregation index).; Generally, Random Forest recorded a higher prediction rate and greater contiguity when compared to J48. However, Random Forest over-predicted soils such as Gilpin, Guernsey, Zanesville and Captina Series which occupy large areas, at the cost of prediction accuracy of soils which occurred in smaller proportions. The results showed that the highest prediction rate based on the dominant soil series (>0.5) and higher values of contiguity index (0.83) and aggregation index (84.2) for RF was observed in the model built using the training set preprocessed for disambiguation. This suggests an improvement in the quality of predicted maps as a result of disambiguation of training set labels. The model predictions were helpful in locating many individual component series in soil consociations and associations. The maps were useful in identifying areas of uncertainty such as misplacement of polygon boundaries, incorrect labeling and disparity along the county edges, which could serve as a guide for further field investigations. The predicted models also provided valuable information for rationalizing the mapping intensity for adjacent SSURGO maps.