Abstract
Agricultural land suitability assessments are important tools to ensure our natural resources are being used effectively and efficiently. The goal of the assessment is to find homogeneous regions using environmental variables and classify the regions depending on how they rank for a specific process (e.g. crop yield, soil erosion, pesticide leaching). The environmental variables and methods used to generate homogeneous regions for agricultural land suitability maps drastically differ from study to study. Some studies only use soil while others only use climate. Nevertheless, improving yield and implementing sustainable practices in these homogeneous regions require detailed information on soil properties and other abiotic factors, such as climate and topography. However, the typical output is on a small-scale and vegetation specific. Currently, large-scale polygon maps have been used for many decades for agricultural land suitability, but with the increase in computation power and available data, these methods might be outdated. A more recent approach to create homogeneous regions is terron mapping, which analyzes spatial variation by clustering soil, climate, and landscape data into similar regions.
While the word terron was derived from terroir, terrons have been proposed to be used for a range of uses, not just with vegetation suitability mapping. Terron mapping can be helpful in predicting environmental risk and reducing the amount of soil management needed. For example, erosion risk mapping assesses the interaction between slope, soil, land cover, and climate variables. Thus, mapping terrons constitutes an easy method for erosion risk assessment and facilitates improved soil management since terron classes consists of soil, climate, landscape, and their interactions. Studies that need to create different zones where management should change locally, such as in precision agriculture, or predict yield according to different variables in various fields can also utilize terron classes and mapping.
Two methods exist for generating terrons: starting from points or continuous, gridded data (i.e. rasters). The different methods have not been compared previously. In this thesis, the two terron methods are compared and the strengths and weaknesses of each method are determined. Through the creation of national scale maps on production (i.e. hard grain values, winter wheat yield, and forest growing conditions), validation of the terrons can be achieved. The data spans from 1688 to present day and represent different spatial patterns. Accomplishing the thesis consisted of three parts: create point-based and raster-based terrons for Denmark, generate large-scale maps on production, and compare the terron maps and methods based on the production data and different aspects of the methods.
The creation of national scale terron maps using both methods resulted in maps that show many similarities and some differences. The point-based terrons included 27 environmental variables and utilized both unsupervised (i.e. fuzzy C-means) and supervised (i.e. Cubist regression trees) machine learning algorithms. The nine terron classes each have an associated terron membership map. The raster-based terrons yielded two maps: a coarse national scale terron map and a detailed regional scale terron map. The three national terrons were created using six low-resolution environmental variables and fuzzy C-means, which also generated a corresponding membership map. The nine regional terrons within each national terron (for a total of 27 regional terrons) were created using six high-resolution environmental variables and K-means. All regional terrons were combined into seven terron groups for easier interpretation and comparison. Both terron methods indicated that the soils in the west are the worst and the soils in the east are the best for most crops. The difference between the maps arise in how many properties are associated with the terron classes. The point-based terron classes include many more environmental variables compared to the raster-based terrons.
The data created for validation consisted of four datasets that span three time periods and three different measures of suitability. The first two datasets are hard grain values from 1688 and 1844. Historical data was important to incorporate since current farming practices allow most crops to be grown across all of Denmark. Using historical data constitutes a way to look at natural growth potential before technology drastically altered the soil (e.g. drainage pipes, irrigation, fertilizer). The two present-day datasets are winter wheat yield and tree growing conditions. The winter wheat yield map utilized more environmental covariates and a sophisticated machine learning algorithm (i.e. Random Forest) to update the previous winter wheat yield map. The tree growing conditions are based on ecograms for 23 species. The number of optimal trees that can grow in each location was used for this validation dataset. The validation datasets represented a range of spatial and temporal variation to assess the terron maps from many different aspects.
The comparison of methods revealed that the raster-based terrons take much longer to accomplish but the method has been automated; thus, this method is ideal for a user who does not want to change parameters and has limited knowledge of computer programming and machine learning algorithms. The time required to run the machine learning algorithms for the point-based terrons was much quicker; however, many more small steps are required, increasing the amount of time needed. My recommendation for which terron map should be used for Denmark would be the raster-based terron map. This map worked better with the validation data, the steps are automated, and easily repeatable for updating the terrons.
The terron maps provide valuable information for large-scale land suitability assessments, especially if developing a terroir is the final goal. The point-based and raster-based methods each have their strengths and weaknesses and can be utilized for different purposes. The creation of these maps incorporates an array of environmental information and will be used in future digital soil mapping projects.
While the word terron was derived from terroir, terrons have been proposed to be used for a range of uses, not just with vegetation suitability mapping. Terron mapping can be helpful in predicting environmental risk and reducing the amount of soil management needed. For example, erosion risk mapping assesses the interaction between slope, soil, land cover, and climate variables. Thus, mapping terrons constitutes an easy method for erosion risk assessment and facilitates improved soil management since terron classes consists of soil, climate, landscape, and their interactions. Studies that need to create different zones where management should change locally, such as in precision agriculture, or predict yield according to different variables in various fields can also utilize terron classes and mapping.
Two methods exist for generating terrons: starting from points or continuous, gridded data (i.e. rasters). The different methods have not been compared previously. In this thesis, the two terron methods are compared and the strengths and weaknesses of each method are determined. Through the creation of national scale maps on production (i.e. hard grain values, winter wheat yield, and forest growing conditions), validation of the terrons can be achieved. The data spans from 1688 to present day and represent different spatial patterns. Accomplishing the thesis consisted of three parts: create point-based and raster-based terrons for Denmark, generate large-scale maps on production, and compare the terron maps and methods based on the production data and different aspects of the methods.
The creation of national scale terron maps using both methods resulted in maps that show many similarities and some differences. The point-based terrons included 27 environmental variables and utilized both unsupervised (i.e. fuzzy C-means) and supervised (i.e. Cubist regression trees) machine learning algorithms. The nine terron classes each have an associated terron membership map. The raster-based terrons yielded two maps: a coarse national scale terron map and a detailed regional scale terron map. The three national terrons were created using six low-resolution environmental variables and fuzzy C-means, which also generated a corresponding membership map. The nine regional terrons within each national terron (for a total of 27 regional terrons) were created using six high-resolution environmental variables and K-means. All regional terrons were combined into seven terron groups for easier interpretation and comparison. Both terron methods indicated that the soils in the west are the worst and the soils in the east are the best for most crops. The difference between the maps arise in how many properties are associated with the terron classes. The point-based terron classes include many more environmental variables compared to the raster-based terrons.
The data created for validation consisted of four datasets that span three time periods and three different measures of suitability. The first two datasets are hard grain values from 1688 and 1844. Historical data was important to incorporate since current farming practices allow most crops to be grown across all of Denmark. Using historical data constitutes a way to look at natural growth potential before technology drastically altered the soil (e.g. drainage pipes, irrigation, fertilizer). The two present-day datasets are winter wheat yield and tree growing conditions. The winter wheat yield map utilized more environmental covariates and a sophisticated machine learning algorithm (i.e. Random Forest) to update the previous winter wheat yield map. The tree growing conditions are based on ecograms for 23 species. The number of optimal trees that can grow in each location was used for this validation dataset. The validation datasets represented a range of spatial and temporal variation to assess the terron maps from many different aspects.
The comparison of methods revealed that the raster-based terrons take much longer to accomplish but the method has been automated; thus, this method is ideal for a user who does not want to change parameters and has limited knowledge of computer programming and machine learning algorithms. The time required to run the machine learning algorithms for the point-based terrons was much quicker; however, many more small steps are required, increasing the amount of time needed. My recommendation for which terron map should be used for Denmark would be the raster-based terron map. This map worked better with the validation data, the steps are automated, and easily repeatable for updating the terrons.
The terron maps provide valuable information for large-scale land suitability assessments, especially if developing a terroir is the final goal. The point-based and raster-based methods each have their strengths and weaknesses and can be utilized for different purposes. The creation of these maps incorporates an array of environmental information and will be used in future digital soil mapping projects.
| Originalsprog | Engelsk |
|---|
| Forlag | Århus Universitet |
|---|---|
| Antal sider | 192 |
| Status | Udgivet - mar. 2021 |