Wild boar damage mapping in agricultural grass and wheatlands using Unmanned Aerial Vehicle (UAV) data

Abstract

The growth of the wild boar (Sus Scrofa) population in Sweden results in increasing wild boar damage on agricultural grounds, thereby influencing the livelihoods of farmers. Finding mitigation strategies that are positive for both wildlife and human society is challenging. It is therefore essential to enable the creation and investigation of precise and differentiated wild boar management strategies. This necessitates the development of semi-automatized and trusted methods that quantitatively and spatially assess wild boar damages on agricultural lands at the landscape scale (i.e., cm-level). Current methods cannot answer this need as they consist of manual field surveys that are time-consuming, subjective, and have a too-coarse scale. Unmanned Aerial Vehicle (UAV) can be a solution to systematically assess damages in agricultural fields on this landscape scale, as it provides data with high temporal and spatial scale, flexibility, and affordability. This study developed three methods that utilized UAV data to automatically map (wild boar) damages in agricultural grass and wheatlands in Boo, Hjortkvarn Municipality, Sweden. Two methods, pixel-based and object-based classifications, focused on performing an image classification on 2-dimensional (2D) multispectral UAV data by applying the machine learning algorithms Random Forest (RF) and Support Vector Machine (SVM). The object and pixel-based RF and SVM classification retrieved overall accuracies of 85% and above for the wheatlands and 91% and above for the grassland, whereby the best overall performance was achieved by the object-based SVM method for both wheat and grassland. The two classification methods created damage maps with similar damage locations but the pixel-based classification mapped the damage extents smaller compared to the object-based classification. The third method utilized the 3-dimensional (3D) UAV photogrammetry-derived point clouds of the wheatlands. This method extracted the normalized height values of the point cloud and applied a height threshold to create the damage mapping. The method was not able to capture the damage mapping in wheatlands due to the too-low density point clouds but has the potential to aid the 2D classifications by providing the extracted normalized height values as additional information for the machine learning classifiers.

Popular Abstract

Wild boar has been damaging agricultural fields in Sweden. By collecting drone data from damaged agricultural fields, damage mappings can be created that can aid the creation of precise mitigation strategies for this conflict. This master’s thesis developed three methods that can automatically transform drone data from agricultural lands into wild boar damage mappings with centimeter-level spatial resolution in agricultural grass and wheatlands.

There has been an increase in wild boar (Sus Scrofa) population in Sweden (Swedish Association for Hunting and Wildlife Management). The coexistence of humans and wild boar has been challenging and has led to conflicts. So does the wild boar damage agricultural lands in Sweden of which the negative consequences were estimated to be around 17% of the Swedish net farm income in 2015 (Gren et al., 2019).
The creation of mitigation strategies for this conflict can be aided by having precise wild boar damage mappings. Currently in Sweden, field surveys are done to collect data on wild boar damage. However, these are time-consuming and often lead to inaccurate and too coarse data. Studies have been looking into alternatives and found drone data to have potential. This data has a centimeter resolution and can be collected with high efficiency and flexibility in usage.
This master’s thesis focused on creating automated methods that transform this drone data into wild boar damage mappings for grass and wheatlands in Boo, Hjortkvarn Municipality, Sweden. Three methods were developed and tested for performance. Two methods, the object and pixel-based classifications, used two-dimensional (2D) multispectral drone data and used the machine learning classifier Random Forest (RF) and Support Vector Machine (SVM) to perform the damage mapping for grass and wheatlands. The third method used three-dimensional (3D) drone data to create wild boar damage mappings for wheatlands.
The results showed that the best overall performance in wild boar damage mapping was achieved by the object-based SVM for both grass (overall accuracy 91%) and wheatland (85%). This method relied mostly on the texture-related values to create the damage mappings. The object and pixel-based classifications mapped similar damage locations but the pixel-based mapped them less dense. The 3D method failed to map a lot of the damage.

Popular Abstract

There has been an increase in wild boar population growth in Sweden leading to more wild boar damage to agricultural lands. To aid the creation of differentiated and precise mitigation strategies for this conflict, this master thesis developed three methods that automatically transform Unmanned Aerial Vehicle (UAV) data into high spatial scale (centimeter-level) damage mappings using data from agricultural grass and wheatlands in Boo, Hjortkvarn Municipality, Sweden.

Two developed methods, called object and pixel-based classification, focused on using two-dimensional UAV data to create wild boar damage mapping in grass and wheatlands. The object-based method has been found useful by related studies for this specific application but has the disadvantage of being computationally heavy. It was therefore interesting to also investigate the pixel-based classification which is a computational lighter approach. The input data for these methods consisted of ortho-mosaics containing wavelength bands Red, Green, Near Infrared, Red Edge, and five derived bands which were the Normalized Difference Vegetation Index and four texture-based bands. The methods applied the machine learning classifiers Random Forest (RF) and Support Vector Machine (SVM). The best overall performance was achieved by the object-based SVM classification with an overall accuracy of 91% and 85% for grass and wheatlands respectively. The SVM had put the most importance on the texture values during the classification. There was an agreement of about 80% between the object and pixel-based damage mapping where similar damage locations were mapped but the pixel-based mapped the damages in those areas less dense.
The third method developed in this thesis was a point cloud height threshold classification. This method used a UAV photogrammetry-derived point cloud to create damage mappings for wheatlands. It is based on applying a threshold on the elevation difference between non-damaged (i.e. wheat plants) and damaged areas (i.e. damage pits). The performance showed an average detection of the validation data of only 19% where it failed to map most of the damages. The performance could be explained by the too-low-density point cloud that was used for the damage mapping.
Prospects can be related to scaling up this application to the national level using Airborne data and to creating an automatic damage type discrimination that can classify the mapped damages to their exact type of damage (e.g., wild boar, machinery, drought, deer, etc.)

• author: Kleijkers, Yrsa
• supervisor: Ulrik Mårtensson; Florent Rumiano; Petter Kjellander
• organization: Dept of Physical Geography and Ecosystem Science
• alternative title: Wild boar damage mapping in agricultural grass and wheatlands using drone data
• course: NGEM01 20241
• year: 2024
• type: H2 - Master's Degree (Two Years)
• subject: Earth and Environmental Sciences
• keywords: Physical Geography and Ecosystem analysis, Unmanned Aerial Vehicle, Wild Boar, Agriculture, Machine Learning, Random Forest, Support Vector Machine, Geo-information Science, Remote Sensing, RStudio
• publication/series: Student thesis series INES
• report number: 668
• language: English
• additional info: This master's thesis was supported by Dr. Florent Rumiano, Prof. Petter Kjellander and Arvid Norström from SLU (The Swedish University of Agricultural Sciences) Grimsö Wildlife Research Station. Dr. Florent Rumiano supervised all aspects of this master’s thesis. The methodology used in this thesis is an adaptation of his developed methodology and is part of a project from Prof. Petter Kjellander at the SLU Grimsö Wildlife Research Station, Department of Ecology, funded by Naturvårdsverket (EPA). This thesis’s developed scripts in the programming language R, are my transcription of the methodology based on a collaboration between Dr. Florent Rumiano and me. The data used in this study was collected by Arvid Norström and Dr. Florent Rumiano.
• id: 9167486
• date added to LUP: 2024-06-24 09:34:40
• date last changed: 2024-06-24 09:34:40