Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

8.1 General conclusions This dissertation aimed at the development of a methodology to quantify vegetation regrowth and burn severity over large areas based on time series of satellite data. The pixel based regeneration index has been the driving force of this research, since it allows to monitor the vegetation regrowth at pixel level without the use of detailed local reference maps. The first general objective of this research was the development of this pixel based regeneration index using time series similarity measures and spatial context. The second general objective focused on the evaluation of the developed pixel based regeneration index by assessing the amount of change in post-fire vegetative development and by relating that change to burn severity and fractions of woody and herbaceous vegetation. In the first part (Chapters 3-5), the use of time series similarity and spatial context to develop the pixel based regeneration index is extensively described. Chapter 3 provides a theoretical comparison between the advantages and disadvantages of the commonly used time series similarity measures. The theoretical comparison focused on the specific properties of each time series similarity measure to provide a valuable basis for identifying, monitoring and classifying vegetation dynamics. Additionally, u numerical experiment elaborated on these theoretical differences and illustrated the utility of each of these similarity measures to discriminate between specific differences between time series (assumption I.i). The use of spatial context and hierarchical concepts as a additional information source was subsequently introduced in Chapter 4. This chapter focused on the development of an multi-temporal hierarchical image segmentation methodology that clusters adjoining pixels with similar temporal properties into hierarchical segments at various scales. Application of the methodology demonstrated the concept of MTHIS and illustrated the importance of the average and annual term to describe the majority of the spatio-temporal variability in the NDVI time series. The correspondence between MTHIS results and reference layers of vegetation characteristics at different scales revealed moreover a close relationship between segmentation output and landcover-landuse reference map (assumption I.ii). In Chapter 5, the concepts of time series similarity and spatial context were combined in a methodology for the image based selection of control pixels on regional to global scale (Chapter 5). Validation based on non-burnt pixels showed that the selection of the control pixels based on similarity of the one year VIt before fire, provides an improvement for the selection of control pixels over large areas (assumption I.iii). The time series similarity approaches specifically minimize the drawbacks of the classic approach based on reference areas,namely the within-class heterogeneity and dependence of static reference data, by selecting only pixels that effectively resemble the focal pixel based on similarity of the one year VI before fire. Comparison of the time series similarity approaches showed moreover that optimal reference pixels were obtained for the TSS-RMSD approach with x = 4 and NT = 8 due to beneficial averaging effects and minimal window size. As such, the effects of spatial heterogeneity and noise are minimized and the control pixels provide optimally the temporal profile of the focal pixel VI-focal in case the fire had not occurred. Pre-fire RMSD values of the control pixels of the TSS- approach allowed moreover to derive the quality of the control pixels before using them in pRI calculations. In the second part of this dissertation (Chapters 6-7) the developed pixel based regeneration index was used to assess the amount of change in post-fire vegetation growth. The focus of Chapter 6 was on the development of an indicator of burn severity (IBS) that quantifies the integrated change caused by vegetation fires. Comparison of the IBS with detailed estimates of burn severity showed that the IBS allows to quantify the burn severity at pixel level over large areas (assumption II.i). The IBS also showed an improvement over the traditional burn severity, since it must be collected throughout the growing season and is not dependent on the moment of Landsat image acquisition. Consequently, IBS estimates can be used as alternative input to global wildland fire emissions models to quantify the variable component related to burning efficiency. Finally, the contribution of different vegetation components to the pRI time series was studied in Chapter 7. In this chapter, the use of the STL method was evaluated to decompose the original time series of vegetation growth into separate long-term and seasonally varying components, which can be related to woody and herbaceous vegetation. Results of the methodology confirmed the assumption that temporal variations of woody and herbaceous vegetation are different, where high post-fire severities lead to a high decrease in woody vegetation. For the herbaceous component this is not necessarily true which can be explained by the general fast recovery of herbaceous vegetation, the increased amount of nutrients, and the high dependency of herbaceous vegetation on rainfall events (assumption II.ii). In general the results of this research point out the large amount of information that is contained in time series of coarse resolution satellite imagery. Standardized methodologies need however be developed to overcome the difficulties associated with these data sets. The use of relative estimates and incorporation of spatial context may provide a first steps towards such standardized methodologies. 8.2 Relevance and potential applications Assessment of the impact of fire emissions on atmospheric chemistry and understanding, in terms of ecological processes, the changes of terrestrial ecosystems after fire at various spatial and temporal scales are major challenges for a wide range of researchers, ecologists and resource managers worldwide. The developed methodologies and obtained results of this dissertation can contribute to this assessment and understanding at different scale levels. The TSS methodology, for example, allows the calculation of pRI time series based on any multi-temporal data set. As such, it can be used to examine the global effects of fire disturbances from regional to continental scales and can contribute to measure how fire effects vary within burnt areas (e.g., by using the IBS). This within-burn variability is crucial to accurately estimate the impact of fire on vegetation, but also is essential for ecologists and resource managers who want to understand fire effects on ecosystems and want to plan post-fire rehabilitation. Moreover, it is essential to reduce current uncertainties in the current emission models that quantify the impact of fire on atmospheric chemistry. Moreover, the developed methodologies are not restricted to burnt areas. The proposed methodologies are generalized approaches that can be applied to any type of disturbance or change based on any multi-temporal data set. As such, they can contribute to obtain a better understanding of factors affecting land-use and land-cover resulting in potential climate warming, land degradation and bio-diversity loss. 8.3 Further research suggestions and perspectives Several new research possibilities arise from this work. Broadly, without claiming the completeness of the list, these research possibilities can be grouped into (i) potential improvements of the current approach and (ii) completely new research pathways. 8.3.1 Potential improvements In the framework of this research, emphasis was on NDVI time series of VGT S10 time series. This data set has however its limitations, which lead to the following suggestions that can enhance the performance of the methodology: BRDF corrected data sets may provide u useful improvement to overcome the noise associated with the high variability due to variation in illumination and observation angles. The use of BRDF corrected data sets (e.g., SPOT-Vegetation D10, MODIS corrected BRDF data) should therefore certainly considered towards future research. Adapted vegetation indices may provide a useful alternative to the current use of NDVI time series. NDVI time series were not designed to capture specific vegetation variation after fire and consequently are rather insensitive to changes induced by the fire on the vegetation cover (Lasaponara, 2006). The use of adapted wavelengths and spectral design may therefore provide an improvement to the proposed approaches. Improved compositing criteria are necessary when studying vegetation regrowth. The S10 data set is a maximum NDVI value composite. As such, it selects pixel measurements of the date when the NDVI is maximum and tends to select pixels that are minimally affected by fire occurrence due to the expected NDVI decrease after fire. The use of alternative compositing criteria (e.g., minimum NIR composites) should therefore certainly be studied. Improved spatial resolution will allow to obtain similar characteristics at a more detailed scale level. The use of other data sets with higher spatial resolution (e.g., MODIS and MERIS) should therefore be considered to improve accuracy at more detailed scales. Improved validation is essential before applying the methodologies on global scale levels. This requires however a large sample of in situ measurements and fine spatial resolution

Academic collection --- 574.9 --- 504 --- 630*231 --- 630*43 --- 528.8 --- Biogeography in general. Geographical distribution of organisms --- Environment. Environmental science --- Natural regeneration --- Forest fires. Prevention and control. Salvage. Revegetation etc. Use of fire in silviculture --- Remote sensing. Teledetection --- Theses --- 528.8 Remote sensing. Teledetection --- 630*43 Forest fires. Prevention and control. Salvage. Revegetation etc. Use of fire in silviculture --- 630*231 Natural regeneration --- 504 Environment. Environmental science --- 574.9 Biogeography in general. Geographical distribution of organisms

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Seasonal forecasts of the Earth’s landscape and vegetation can be very valuable information to overcome modern-day problems such as climate change, hunger, and natural disasters. Satellite images, which are frames of the Earth’s surface, can provide a lot of information for making such seasonal forecasts. Moreover, it is highly available data accessible over an extended range of 8 years back in time. Having large-scale and long-term historical data simplifies the process of training and testing artificial neural networks for landscape predictions. Forecasting those images, with the help of additional parameters, can help predict the future conditions of the Earth’s surface. EarthNet2021 is an open-source challenge that provides a large dataset suitable for training deep neural networks on this task. It contains 32,000 samples of satellite imagery covering 2.56 x 2.56 km in 128 x 128 pixels, which are already pre-processed and ready to be used as input data for prediction systems. The existing literature, which includes baseline models provided by EarthNet and new extensions of these, provides a starting point for the analysis in this thesis. The thesis first trains the various models and chooses the currently best suitable model for landscape prediction. It then analyses the source of the limited prediction accuracy of the chosen model. The analysis shows that the prediction accuracy varies according to the location of the satellite frames, the greenness variation for each image time series, and the land usage of each area. Central Europe exhibits the lowest prediction accuracy among the studied regions; regions with moderate variation in greenness have a minimal EarthNet prediction score; and landscapes classified as croplands also display poor prediction performance. These findings lead to the recommendation to tune the training and testing of artificial neural networks for landscape prediction on specific landscapes. To complete this thesis, a use case is developed to demonstrate how the landscape is changing according to manipulated weather variables. The model was implemented and analyzed with the use of the programming language Python and expanded with Deep Learning libraries such as PyTorch.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Gone with the wind :Wind-blown snow in Antarctica from a modelling and observational perspectiveRecent findings point out the importance of ice shelf dynamics to control the mass balance of the Antarctic Ice Sheet (AIS). Ice shelves exert a buttressing effect on the ice sheet and their disappearance can trigger enhanced ice discharge into the oceans, in turn increasing the sea level rise. Currently, the main concern is directed towards basal mass balance of the ice shelves. Yet, not much is known about the effect of wind-induced processes on the health of the ice shelves: the surface energy balance and surface melt of the ice shelves is affected by processes such as drifting snow, frequent at the surface of the AIS. Regional climate models are often used to investigate current and future evolution of the AIS surface mass balance (SMB) but sill lack an accurate drifting snow parametrization. We propose here to use unprecedented long-term ground observations to validate satellite retrievals. Then, we investigate the representation of accumulation by regional climate models and reanalyses (driving models for regional climate modelling) over the AIS. Finally, we include a blowing snow routine in a regional climate model recently adapted for Antarctic conditions.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The overall goal is to identify the role of precipitation and clouds in the East Antarctica, where little is known about the interaction between clouds, precipitation and their effect on the surface mass balance. Moreover there is a large uncertainty about relating local surface mass balance measurements to the mass gain or loss of the ice sheet due to the process of snowdrift. This is unfortunate, as the Antarctic ice sheet is expected to become a dominant contributor to sea level rise in the 21st century. The candidate will focus on the cloud and precipitation processes by using both observational data and a regional climate model.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

One of the last strongholds of old-growth forests in Europe can be found in the Romanian Carpathians. These forests provide a myriad of ecosystems services and therefore require strict protection, yet Romania’s old-growth forests are currently threatened due to inadequate policy. Mapping old-growth forests is a first step towards their protection. Since current mapping efforts are based on incomplete inventories with different focuses, omission errors are likely high. Therefore, an approach based on remote sensing is proposed. The high structural variation of old-growth forests is a common element across the many existing definitions. This variation is exploited in the proposed method, which aims to answer (i) how well old-growth forests can be distinguished from other forests based on their horizontal variation in forest structure, (ii) what the added value is of vertical variation derived from space-borne LiDAR (GEDI) in the identification of old-growth forests, and (iii) what the necessary minimal area is to identify old-growth forest patches as such. Sentinel-2 data was employed to assess the horizontal variation using the coefficient of variation of the NDVI, NDWI and SI, and using the mean, contrast and entropy of the grey-level co-occurrence matrix for the green, NIR and SWIR bands. Vertical variation was assessed using metrics derived from the Global Ecosystem Dynamics Investigation (GEDI) spaceborne LiDAR. As the magnitude of variation depends on the measurement scale, horizontal variation was assessed for windows of increasing size and vertical variation was assessed in transects of increasing length. Support vector machine models based on both types of variation separately gave unsatisfactory results because of the limited gradient in non-old-growth forests and class imbalance in the available training data. The optimal support vector machine model based on horizontal variation used a minimum area of 1 ha, which is in accordance with the minimum area for set-aside old-growth patches to be ecologically functional when combined with larger old-growth forests in a network. Although field inventories remain necessary to identify specific indicators, such as the presence of habitat trees, remote sensing can provide a first delineation of old-growth forests based on their structural variation, if the proposed improvements to the method are taken into account.