Ruth Depth Prediction In Boreal Forest Based On Open-Source Data and Machine Learning

One of the major concerns in modern forestry revolves around the environmental impacts stemming from machine trafficking, especially in terms of rutting and compaction. Extensive research efforts in recent years have explored the relationship between soil properties and bearing capacity of forest soils across various types, including fine-grained forest, mid-grained forest, and peatland forests. While significant strides have been made in hydrological modelling of forest soils, models that overlook soil properties tend to poorly predict water content and trafficability. Despite a general understanding of the causes of rutting, the development of suitable predictive models and operational procedures to anticipate and visualize the probability of rutting for forest machine operators prior to operations remains a significant challenge. The Forest Centre of Finland (Metsäkeskus) conducts annual inventories at roughly 100 thinning sites with a specialized post-harvest quality control method. The method involves inventorying all trees within 5-10 circular plots. Additionally, rut depth measurements are taken along the closest logging trail. With precise knowledge of the of the tree plots and rut inventory lines, we complemented the field data with all relevant information sourced from various open-source portals. These data include topography factors, permanent soil attributes (soil texture, etc.), non-permanent soil factors (including the depth-to-water index, timing of operations, snowfall, precipitation, etc.), as well as the distance to ditches or ditch density. We selected 40 sites with a normal distribution across Finnish forest, comprising 420 sample plots and 425 rut measurement points for our investigation. Subsequently, we overlaid a hexagonal tessellation, each cell measuring 100 sq.m, onto the selected sites. Rut depth (the target variable) values and predictor variables (e.g., topography, soil characteristics, non-permanent soil attributes, and hydrology) were aggregated within the cells of the tessellation. We then utilized approximately 420 cell-samples, encompassing all variables including the target and predictors, to train a Random Forest regression model with optimal parameters for predicting rut depth. Our findings demonstrate the efficacy of the trained Random Forest model in predicting soil rut depth, emphasizing the importance of comprehending the dynamic interplay between soil rutting and its determining factors for fostering sustainable forest management practices. In this presentation we will show how different data structures were created and combined, along with the introduction of our trained machine learning-based models designed to predicts probabilities of rut depths in both mineral and peaty forest soils. We will discuss the appropriateness of the modes developed in assisting forest machine operators in planning and execution of daily operations.