Forest Information Technology M.Sc.
Refine
Document Type
- Master's Thesis (3)
Language
- English (3)
Has Fulltext
- yes (3)
Is part of the Bibliography
- no (3)
Keywords
- Dead trees (1)
- Deep learning (1)
- Remote sensing (1)
- Semantic segmentation (1)
- Soil Organic Carbon; Remote Sensing; Proximal Sensing; Soil Sampling; Random Forest (1)
- U-Net (1)
- UAV (1)
- gestość punktów (1)
- lidar (1)
- point density (1)
Institute
Soil organic carbon (SOC) is an important environmental factor that impacts soil quality and function, global food security, and efforts to mitigate climate change. It is essential to accurately estimate and predict SOC levels on a large scale. While spectroscopy through proximal sensing is effective in accurately predicting SOC levels, its limitation in estimating SOC on a large spatial scale is a concern. Hence, there is a need for faster and more cost-effective techniques for quantifying SOC content. Recent research has proposed the use of remote sensing (RS) methods as a potential solution. The main objective of this research was to assess and compare the proximal soil sensing (field spectroscopy measurements) in monitoring and estimating SOC levels with data obtained from spaceborne Unmanned Aerial Vehicle (UAV) and Sentinel-2A on an agricultural field. To improve the accuracy of the RS methods (UAV and Sentinel-2A) in predicting SOC levels, nine spectral indices were created. The modelling process involved the use of different bands and wavelengths, specifically four bands for UAV and eight for Sentinel-2A. In addition, the computed spectral indices were used as independent variables to create prediction models for soil content, Random Forest (RF) model is trained with 90 soil samplings collected from the field, and validated by ten-fold cross-validation. Prior to conducting the SOC predictions, the study investigated the covariate importance. The models created from proximal sensing data had better accuracy in making predictions with the help of RF compared to the other two methods. The study demonstrated that proximal and remote sensing technologies can be exploited efficiently for SOC prediction.
Forests are an important part of the ecosphere and have long been at the center of extensive research. For quantifying forest metrics on larger scales, remote sensing technologies have become important tools over the last decades. One of these tools is airborne lidar, which has enabled detailed analyses of canopy structures by providing three-dimensional point clouds of forests. Within such data sets, individual crown bodies can be detected and delineated, which is referred to as tree crown segmentation. However, among the few available algorithms the most sophisticated ones are also the most resource hungry. Another general issue with lidar point clouds is that their density can vary strongly between adjacent areas which also influences segmentation results. In this study a segmentation workflow from point cloud preparation to the evaluation of results was developed. Part of this workflow was a novel approach to removing undesired density patterns from point clouds, using detailed information on the measurement procedure. Also, a scalable segmentation algorithm, based on 3-dimensional mean shift clustering was implemented. It reached single-core runtimes of about 50 seconds ha -1 on point clouds with a density of 10 points m -2 . The segmentation results allowed for the estimation of stem diameter distributions for a temperate mixed forest stand. Overall, the presented workflow provides a solid basis for individual tree segmentation and can be further extended and scaled with relative ease.
The area covered with forests and trees is an important indicator of the state of the environment. There are multiple difficulties that forests face which cause them to decline. It is possible to track the physiological stress caused by biotic or abiotic stimuli in forests. At present, one of the major problems is the large number of dead trees that directly damage neighbouring trees. If a thorough overview of the affected area and the number of damaged trees can be quickly recorded and incorporated into the planning of forest management measures. Dead trees can be identified appropriately using Remote Sensing (RS) and Artificial Intelligence (AI) approaches. AI has developed a number of image segmentation algorithms that can classify dead trees from RS data, such as unmanned aerial vehicles (UAVs) images. One of the Machine Learning (ML) algorithm, Deep Learning (DL), is becoming increasingly popular due to its outstanding image segmentation performance and various image processing methods. From this perspective, this research aims to utilize one of the DL models, U-net, to segment the dead spruce (Picea abies) trees in the UAV orthophotos. The network was trained using several experiments as preliminary tests by altering the pixel size, loss functions, the number of parameters, etc. As an outcome, the semantic image segmentation using U-net architecture and the combination of an appropriate training strategy for dead spruce tree detection on UAV orthophotos were successful. However, the prediction model revealed a minor overfitting, which may be fixed by adding more training data sets and modifying the U-net architecture. Even though the constructed model and the available data were used to produce a very effective visual explanation of the classification.