|dc.description.abstract||New Zealand's forestry industry requires new information on the spatial distribution of certain target soil properties and locally significant soil classes in order to implement sustainable and site-specific forest management practices. Soil-landscape modelling has been identified as a potentially useful tool for collecting this information. However, its performance in plantation forest environments - where the impacts of forest management on soils can be considerable - has not yet been comprehensively evaluated. As a step towards such an evaluation, this study was undertaken to determine and assess the impacts of hauler-based, clear-fell, forest harvesting on the performance of soil-landscape modelling as a tool for the spatial prediction of soil classes and target soil properties in a radiata pine forest. The research was conducted within southern Mahurangi Forest, an exotic, Pinus radiata-dominated plantation forest, situated on the Northland Peninsula, North Island, New Zealand.
Three major sub-studies were conducted. In the first, the impacts of forest harvesting on the performance of the qualitative soil-landscape modelling approach to the spatial prediction of soil drainage classes (identified as locally significant) were determined. In the second, the impacts of forest harvesting on the predictive relationships used by quantitative soil-landscape modelling and class-based (including semi-quantitative soil-landscape modelling) approaches to the spatial prediction of target soil properties were determined. Harvesting impacts on the magnitude and variance of the target soil properties (topsoil pH, available Mg, available P, available K, macroporosity, and total C) were also examined. In the third sub-study, the impacts of forest harvesting on the performance of seven techniques representing the class-based, quantitative soillandscape modelling, and geostatistical approaches to the spatial prediction of target soil properties were determined and compared. Four class-based techniques (labelled 1-4 and based on soil drainage classes, landscape units, and soillandscape units), two quantitative soil-landscape modelling techniques (multilinear regression and regression kriging), and one geostatistical technique ( ordinary kriging) were investigated.
All sub-studies were undertaken primarily within two separate, but essentially adjacent 5-ha sampling plots. One plot was under first rotation mature Pinus radiata trees, i.e. the pre-harvested plot; the other plot had been harvested and was under second rotation, two-year-old trees, i.e. the post-harvested plot. Each plot consisted of 208 sample points on a 16.7-m regular grid pattern. The preharvested plot was effectively the control plot and hauler-based, clear-fell, forest harvesting was the treatment applied to the post-harvested plot.
The predictive models were developed (and predictive relationships investigated) using only 146 of the data points so that the remaining 62 points could be used to validate the predictions. A separate qualitative soil-landscape model was developed for each plot in general accordance with the land systems approach. The performances of the techniques for predicting the target soil properties were evaluated and compared using several statistical measures including mean error, root-mean-square error, goodness-of-prediction, mean rank, and standard deviation of rank. The predictive relationships between the target properties and the soil drainage classes, landscape units, and soil-landscape units were described using a least-squares-means analysis of variance whereas the relationships between the target properties and the terrain attributes ( derived from a 5-m digital elevation model) were described using the squared-multiple-correlation statistic.
The soils of southern Mahurangi Forest differ predominantly in terms of soil drainage condition. A modified version of the New Zealand soil drainage classification was defined to partition better the local variation in soil profile hydromorphology (e.g. the existing imperfectly drained class was subdivided into two new classes). The modified drainage classes were amalgamated into two broad drainage classes (Wet soils and Dry soils) to improve the practicality of the qualitative soil-landscape models. The relationships between the broad drainage classes and the landscape units were found to be fractionally weaker in the postharvest plot than in the pre-harvested. However, the weaker relationships are probably attributable to the generally drier nature of the landscape in the postharvested plot and are not likely to be due to forest harvesting. The qualitative soil-landscape models were applied to predict the spatial distribution of the broad drainage classes within their respective plots. The performance of the models was good, with both registering correct predictions at >80% of the validation points. Therefore, forest harvesting had no detrimental impact on the predictive performance of qualitative soil-landscape modelling.
Forest harvesting was found to have had a significant impact on the magnitude of all target properties. The means of some target properties (topsoil pH, available Mg, and macroporosity) were significantly decreased whereas the means of other target properties (available P, available K, and total C) were significantly increased after harvesting. The variance of some target properties was also found to have been significantly affected by forest harvesting. The variance of topsoil pH and available Mg was significantly decreased whereas the variance of total C was significantly increased after harvesting. Moreover, forest harvesting altered and weakened the relationships between most target properties and the modified soil drainage classes and landscape units. Also, the correlations between most target properties and the terrain attributes were weaker after forest harvesting. However, the relationships between most target properties and the broad drainage classes and soil-landscape units were not altered or weakened by harvesting.
Most techniques for the spatial prediction of target soil properties gave less biased and slightly less accurate predictions of most target soil properties after forest harvesting. Furthermore, most prediction techniques offered less of an improvement in accuracy over the sample mean after harvesting for most target properties, meaning that most techniques became relatively less useful after harvesting. Considering all target soil properties together, the relative performance of some prediction techniques (regression kriging and ordinary kriging) generally became poorer whereas the relative performance of other techniques (class-based 2, class-based 3, and class-based 4) generally improved after harvesting. On balance, the relative performances of the class-based 1 and multi-linear regression techniques remained the same. Ordinary kriging (the geostatistical technique) is the best predictor of target soil properties in the preharvested areas of southern Mahurangi Forest whereas the class-based 2 technique (a semi-quantitative soil-landscape model) is the best within the post-harvested areas. Furthermore, the class-based 2 technique has the potential to offer a more practical and cost-effective alternative to ordinary kriging throughout the forest. The other techniques (e.g. the quantitative soil-landscape models) either failed to perform well after harvesting or were likely to be less cost-effective, or both.||