Soil sampling strategy optimization to support predictive soil mapping

点击翻译按钮获取中文摘要

Policy makers, planners and producers have recognized that high-quality soil information is essential for effective decision making, but they are often faced with a shortage of sufficiently high-resolution information. Predictive soil mapping is increasingly used to address this shortage, leveraging the power of machine-learning based approaches to overcome a national shortage of field-trained soils specialists. However, predictive mapping still requires in-field sample collection. In the case of producers, sample collection is commonly carried out by their agronomy team and these data are supplemented via equipment-mounted sensors, but the specific sampling needs for effective predictive soil mapping are not the same as for traditional soil testing or yield monitoring to inform nutrient recommendations. Currently, one of the key methodology gaps is determining the amount of sample points required to support predictive soil mapping and assessments. While there has been research conducted on this subject, most of this work has focused on sampling strategies to completely characterize the statistical variability in the landscape. There is still a gap in the literature regarding how much sampling is economically optimal and practical, specifically in the context of the Canadian Prairies. The goal of this project will be to develop a sampling strategy based on a modified conditioned Latin hypercube design that incorporates economic factors. This work will be undertaken with our research partner, CropPro Consulting. CropPro is instrumental for this project as they will be providing both essential data for the research sites along with valuable agronomic expertise. This missing piece in the literature is one of the key obstacles preventing widespread application of predictive soil mapping, as the accuracy and utility of spatial information generated from predictive soil mapping is directly influenced by the soil data collected for use as training data for predictive models.