Alumni of the environmental remote sensing group

Classification of fire types in savannahs of southern Africa using Sentinel data

Study

MSc Geodesy

Date of defence

20.01.2023

Supervisor

Dr. Christine Wessollek, MSc. Tichaona Tavare Mukunga (TU Wien, Department of Geodesy and Geoinformation)

Abstract

Fires are a part of the natural process in savannah areas. It is estimated that 300-1600 Teragram Carbon (TgC) per year are released by these fires. To analyse these fires regarding the land cover type and burn severity, a fire classification method was developed based on the use of land cover datasets and Sentinel-2 data. In order to do this, the ESA WorldCover 2020 land cover model was selected in an observation area in Zambia and compared with the ESA CCI LAND COVER - S2 prototype 2016. This observation area had an extent of 14.8°-15.2°S and 25.7°-26.1°E and contained all of the four main land cover types representative of savannah, tree cover, shrubland, grassland and cropland.
After selecting the first data to be used for classifying the fires, the Sentinel-2 data was processed from L1C to L2A using the processing software FORCE by David Frantz. In addition to atmospherically correcting the Sentinel-2 data, the data was topospherically corrected using the DEM_SRTM1 dataset. The difference Normalized Burn Ratio (dNBR) was then calculated from the Sentinel-2 L2A data as the difference between the Normalized Burn Ratio (NBR) before and after fires. From this, the burn severity was determined as a measure of fire effect levels for burned area based on the method of Key and Benson.
From the land cover and fire severity, sixteen fire classes were defined. These fire classes were then evaluated qualitatively by comparing the results with corresponding climate and soil moisture data. It became evident, that the yearly progression of fires was linked to the precipitation and the corresponding soil moisture of the previous months. In addition, the relationships between these fire classes and the Fire Radiative Power (FRP) were analysed. For this purpose, VIIRS 375m FRP data were compared with the individual fire classes for each year. From the data it became evident, that higher severity in the classified fires corresponds to higher FRP values. Overall, a classification of the detected fires was achieved and the results of which were found to be compatible with available weather and soil moisture data. Furthermore, a correlation between the different fire classes and the measured FRP was determined.

© David Schleehahn

Comparing the German Forest Fire Danger Index and medium-resolution satellite Products for the Quantification of Fire Occurrence in central Europe

Study

MSc Geoinformation Technologies

Date of defence

13.01.2023

Supervisor

Matthias Forkel, Dr. Dirk Pflugmacher (Humboldt-Universität zu Berlin), Christopher Marrs

Abstract

 In this work, the German Forest Fire Danger Index (WBI) was investigated and compared with two Random Forest (RF) classifier models, that aim to predict the forest fire occurrence probability. These models were trained with remote sensing data and past fire events from the year 2020. Specifically, satellite data about the Leaf Area Index (LAI), the Land Surface Temperature (LST), the Surface Soil Moisture (SSM), the Soil Water Index (SWI) as well as other static spatial data about forest type, forest height, slope, aspect, day in the year of the fire event and population density were integrated. The first model included variables which were linked to the WBI, to create comparability. The second model included additional parameters that provide information about the fire ignition.

The evaluation of the WBI, with information on past fire events, burned area, land cover and its causes showed that the WBI does not warn according to its highest fire danger class. Furthermore, the WBI danger classes do not show distinct values in the satellite variables, which means that the warning levels cannot be clearly distinguished from each other. The results of the RF models predicting fire occurrence have an accuracy of 0.67 and 0.71, which can be rated as good.

Colouring and interactive visualization of historical Earth observation data

Study

MSc Cartography

Date of defence

23.09.2022

Supervisor

Matthias Forkel, Mathias Gröbe, Lucas Kugler

Abstract

In the 1960s, the US launched a series of satellite projects aimed at building a reconnaissance system. The CORONA project was in operation between 1960 and 1972. During the operation, a large number of satellite images were taken. Until 1995, these images remained classified. Since 1995 numerous researchers have exploited the CORONA images, primarily to study surface cover changes across the globe. This study adapts the existing DeOldify model, which is dedicated to colourising images to specifically colourise CORONA satellite images through retraining the model using the CORONA image dataset. The CORONA image dataset consists of greyscale CORONA images and colour reference images covering the same region and having little or no change in the image content. A generative adversarial network is used for building this model. A U-Net-based generator network and a binary-classifier-based critic network train in an alternative approach. To improve the training efficiency, transfer learning and NoGAN training techniques are implemented. The generated images are evaluated both quantitatively, using RMSE and PSNR, and qualitatively, using a user study asking the participant if the displayed colour image looks natural. Results show that generated images with good performance using RMSE and PSNR do not necessarily have plausible colours. Compared with the original DeOldify model, the retained model can produce images with more natural and plausible colours, achieving a result of 70.1% of the participants thinking they are real, although these images have poorer reconstructing quality. To visualise the results, a web mapping application is developed with Geoserver and Leaflet. The server stores data and publishes it through WMTS or WFS. Cache service and image pyramid are used to accelerate the response on the client side. The web map client provides an interactive map interface for the users to control the displayed layers and download the selected layer in a self-defined ROI through WPS.

Differentiation of Calluna vulgaris in seed germination and stock sprouting in the Kyritz-Ruppiner Heide in relation with maintenance activities

Study

MSc Spatial Development and Natural Resource Management

Date of defence

14.04.2022

Supervisor

Matthias Forkel, Dr. Carsten Neumann (GFZ Potsdam)

Abstract

The loss of biodiversity is evident in many areas. It is becoming increasingly visible and has many faces, with human influence playing a significant role. Humans are reshaping the planet through immense changes in land use. This influence also contributes to the preservation and creation of new ecosystems. Heathlands are an example of the (un)conscious action of humans, as this habitat was created and shaped by traditional human management. However, this habitat is now also under threat, as traditional agriculture has given way to industrial agriculture and management techniques have changed as a result. It was precisely these that ensured the resilience and thus the preservation of the habitat. Meanwhile, many of the heathland areas are over-aged and in poor to very poor condition. The 2019 FHH report additionally states that their condition will continue to deteriorate (BFN, 2019). The preservation of these habitats is thus dependent on human care.

The promotion of maintenance measures to rejuvenate the heather (Calluna vulgaris) stands is essential to build up a better structure and thus resilience of the entire habitat. The distinction between seed germination and re-sprouting is relevant here (Watt, 1955; Whittaker & Gimingham, 1962; Miller & Miles, 1970; Webb, 1986).
To distinguish between these vegetation types, it was hypothesised that there are differences in growth and flower formation. To validate this hypothesis, areas in the Kyritz-Ruppiner Heide were maintained by burning and mowing. The re-sprouting was recorded by means of regular drone recordings, which serve as data basis for this work. By using different indices, a classification of the individuals heather plants took place, which allowed conclusions to be drawn about the variables: area size, growth difference and flower formation. Based on the assumption that Calluna seedlings have a lower growth rate than re-sprouts, differences in area size and growth height can be assumed. Due to these dynamics, a later and shorter flowering period can also be assumed. Threshold values in the growth dynamics could be identified by means of stepwise adjustment of the variables and subsequent data analysis. In addition, links between flowering periods, growth periods and maintenance measures were identified. Natural areas created by humans develop through and with the influence of humans. To protect and maintain the biodiversity of these habitats, various measures are needed.

Model-based estimation of Live Fuel Moisture Content with Sentinel-1 radar data

Study

MSc Geoinformation Technologies

Date of defence

04.04.2022

Supervisor

Matthias Forkel, Dipl.-Ing. Ruxandra-Maria Zotta (TU Wien, Department of Geodesy and Geoinformation)

Abstract

The assessment of wildfire risk is of great importance worldwide, not least because of climate change. One important risk indicator is the live fuel moisture content (LFMC) defined as the ratio between plant water content and dry biomass, since a higher water content of vegetation decreases the probability of fire ignition and the speed of fire propagation.

Several studies demonstrated the potential of optical satellite data to estimate LFMC at the regional and global scale. However, since the use of optical sensors is limited by atmospheric influences and cloud coverage, microwave satellite data offers a potential alternative. This thesis focused on the potential of Sentinel-1 radar data for the retrieval of LFMC by modifying the approach of Wang et al. (2019), which combines the water cloud model (WCM) and a linear bare soil backscatter model. Since the above-mentioned approach does not consider influences of changes in biomass, vegetation structure and soil moisture on the radar backscatter, biophysical variables of vegetation and soil-related parameters were introduced as additional input in the WCM with the aim to assess their impact on the model’s accuracy. The model was first trained and validated with in-situ measurements available for study sites in the United States of America and subsequently transferred to forest locations in Europe, where LFMC measurements are few.

Using the additional variables, LFMC ground measurements, leaf area index (LAI) and volumetric soil moisture (VSW), the modified WCM was able to accurately simulate the contributions of soil and vegetation to the total radar backscatter. Compared to VSW, LAI exerted a stronger influence on model performance (R²) during the calibration process. To invert the model for the estimation of LFMC, an optimised look-up table (LUT) was built by running the WCM forward using all possible combinations of a predetermined range of values for the input variables. The inclusion of LAI and VSW in the minimisation of a cost function, used to retrieve the optimal LFMC from the LUT, enabled the transfer of the WCM to Europe, since the model accuracy depended primarily on relations between backscatter, LAI, LFMC and VSW. The relations were found to vary with the heterogeneity in prevalent vegetation species and the fraction of vegetation coverage. The derived LFMC values were classified into coniferous, broadleaved and mixed forest, and then validated with the MODIS LFMC product based on the algorithm by Yebra et al. (2018). Promising results for root mean square error (RMSE = 25 %, 26%) and coefficients of determination (R² = 0.50, 0.42) were obtained for deciduous and mixed forests, while the WCM resulted in higher inaccuracies for coniferous forests (RMSE = 45 %, R² = 0.17).

© Johanna Kranz

Predicting spatial patterns of forest fire using random forest machine learning algorithm in the Terai Arc Landscape of Nepal

Study

MSc Tropical Forestry and Management

Date of defence

14.01.2022

Supervisor

Matthias Forkel, Mir A. Matin

Abstract

Increasing forest fire events present novel challenges to forest ecosystems and human well-being. Early detection of forest fires and understanding interrelationship between forest fires and forest fire driving factors are considered important for effective fire management. Unerstanding the perception of decision-makers of existing fire management infrastructure and initiatives is equally important in order to implement preventive measures. We used VIIRS fire dataset (2012-2021), climatic, topographic, and infrastructure predictor variables aggregated over time to generate a fire susceptibility map from random forest (RF) machine learning algorithm. A key-informant questionnaire survey was also undertaken with the forestry professionals working in the TAL region to understand their perception of existing fire management initiatives. RF model achieved an overall accuracy of 72.8% and a sensitivity value of 88% in the evaluation dataset indicating a good predictive ability of fire occurrence. Most parts of the TAL region were susceptible to fire. Our findings of increased fire risk in some districts located at the south-west and north-east part of the TAL were unique to the findings of previous fire-related studies in TAL. Fire occurrence was lowest in 2020 (0.42%) followed by sharp increase in 2021 (19%). Broad-leaved forests were highly susceptible to fire. An increased number of fire events were observed with high temperatures. Fire events were higher in proximity to roads and settlements. Respondents were of the opinion that there is a need for effective management initiatives to improve existing quality and condition of fuel management infrastructure and fire management initiatives. These findings will assist in in the effective decision-making process of fire prevention and suppression reducing the potential future damage due to forest fires in the TAL region of Nepal.

Comparison of deep learning model structures for land cover classification using historical Corona satellite imagery

Study

MSc Geodesy

Date of defence

05.01.2022

Supervisor

Matthias Forkel, Philipp Körner

Abstract

Climate is strongly influenced by the characteristics of land cover, where water, energy and trace gases interact directly with the atmosphere. It is therefore not surprising that a change in land cover also causes a change in local climate and material cycles. Remote sensing has played a fundamental role in continuous earth observation and land cover change detection — at least since the launch of the Landsat satellite program (1972). In 1995, the Central Intelligence Agency from the United States of America released previously secret spy satellite images. These images are pan-chromatic satellite images taken from 1960 onwards with near-global coverage. The satellite images are particularly interesting because they contain information about land cover from 10 years before the beginning of civilian Earth observation. Nevertheless, there are a scarcely studied data source to date. Building on initial approaches, the present work investigates applying different neuronal networks to automatically extract land cover from CORONA data. The particular focus of the work is on the transferability of the approaches to unknown CORONA data. In the long term, the landscape cover is to be extracted from CORONA data on a large scale. To investigate the transferability, seven spatially and temporally distributed test areas in the the federal state Saxony, Germany were selected for cross-validation.

In the present study, a total of four different U-Net architectures were used. A U-Net is a deep neural network for semantic segmentation of image data. All four architectures have different capabilities and are optimised for different land cover classification challenges. As input three different image compositions. The first compostion is the original pixel values (one channel), the second is the GLCM texture measures (ten channels). The third composition is a simulated RGB image (three channels). These three compositions were combined into one image with 14 channels. In addition, semantic masks were created for all test areas to serve as reference data. For this purpose, the landscape coverage was divided into six classes. From these input data, seven different training datasets and test
datasets were derived for cross-validation. Based on the training datasets — four archi-
tectures were trained for each test area and then applied to the corresponding test data.   
Resulting from the training it could first be deduced that a classification of land cover from
CORONA data with the used architectures is possible. In all test areas, all architectures
achieved a training IOU above 85%. Poor to moderate IOU values were achieved for the
individual test areas. The best mean segmentation accuracy is 30.50%, while the worst
value for a test area is 12.71%. Strong accuracy variations were detected among the classes. Classes that have a lot of sample data achieve significantly better IOU values than
underrepresented classes. The best segmentation performance, with an average IOU of
54.51%, was obtained for the agricultural land class, closely followed by forest with 52.80%. For the classes shrub vegetation, open areas and water bodies mostly only IOU
values below 10% were achieved. For classes with many sample data, moderate to good segmentation accuracy was achieved for both spatial and temporal transferability. Since
the data set does not cover all factors for generalised transferability, transferability could
only be ensured under certain framework conditions. For example, the study does not
cover handling of clouds or CORONA data of poor quality.

As a conclusion it was recognised that a combination of results from different architectures might improve the outcome. As a outcome, it was recommended to consider this issue for further developments. 

Assessing the Spatial Variability of Above-ground Forest Carbon using Sentinel-1, Sentinel-2 and Field Inventory Data of the Miombo Woodlands in Songwe District, Tanzania.

Study

MSc Tropical Forestry

Date of defence

29.11.2021

Supervisor

Matthias Forkel, Hamidu Seki ( University of Dar es Salaam - Mkwawa University College of Education, Tanzania)

Abstract

Accurate estimates of above-ground biomass and carbon (AGC) are needed to understand the role of forests in the global carbon cycle. To achieve this, recent forest monitoring and assessment techniques now rely on integrating remote sensing and field inventory data to map and monitor carbon changes in tropical forests. More importantly, AGC spatial distribution is intricately linked to climate, soil, topography, and anthropogenic  disturbances, whose influences should be understood so that their spatial effects are utilised to sustain forest carbon. Hence, this study combined Sentinel-1 and Sentinel-2 variables with field inventory data in a generalised additive model (GAM) to map the Tanzanian Miombo AGC while explaining its spatial variability. Fifteen (15) models were trained and tested using the leave-one-out cross-validation in GAM. Model 6, containing selected predictors – B11 (Shortwave Infrared 1), NDVI (normalized difference vegetation index) and tree canopy cover (%tcc), indicated best prediction accuracy (R²  = 0.87) with the least error (RMSE = 5.24 tC/ha) and an upper bound prediction of 114.87 tC/ha. A combination of Sentinel-2 variables (B11 & NDVI) and %tcc are better predictors for mapping AGC in the study area. Overall, the spatial drivers explained 65.5 % of AGC spatial variability with elevation having the highest predictive power (21.65 %), followed by mean annual temperature (MAT; 21.10 %), the interaction effect of elevation and distance to settlement (20.99 %), distance to settlement (18.24 %) and silt content (18.02 %). The study also showed that AGC increases with increasing elevations and silt content and decreases  with increasing MAT and distance to settlement. More so, AGC increases accordingly with increasing interaction effect between elevation and distance to settlement. Therefore, understanding the impact of these factors on the distribution of AGC will be necessary for adequate forest planning and management.

Predicting, understanding, and visualizing fire dynamics with neural networks

Study

MSc Cartography

Date of defence

22.11.2021

Supervisor

Matthias Forkel, Tichaona Tavare Mukunga MSc. (TU Wien)

Abstract

As machine learning techniques are contributing to scientific research and advancement, the interpretability and visualisation of these algorithms grow in importance. These techniques have introduced many improvements to advance our understanding of fire regime dynamics outperforming process-based approaches. Neural networks have achieved great accuracy with fire modelling, however, challenges arise with unbalanced time series. In this thesis, LSTM neural networks, which are designed for sequence modeling and handling unbalanced data, are investigated to explore their ability to predict fire ignition points. The research is conducted for a small area in western Africa using monthly meteorological variables and fAPAR as an indicator for vegetation for a period spanning from 2003 to 2016. The chosen methodology is based on training one LSTM for each pixel independently. Datasets are pre-processed, structured as a multivariate time series and then arranged to fit LSTM 3D data format. The network architecture was chosen by conducting multiple experiments. The pixel-based LSTM was able to capture the seasonal and spatial varieties with RMSE value computed at 3.333. However, it underestimated the high values of ignitions during the peak of fire season and was not able to record sudden events. To better understand LSTM behavior, multiple interpretation techniques were investigated to evaluate their abilities to determine the most important features and visualise their dependencies. Permutation feature importance gave an overview of overall feature importance while variance-based feature importance was able to map the spatial distribution of each feature. SHAP summary plots gave a detailed interpretation of feature importance of precedent time steps. The most important features to predict fire ignitions were found to be fAPAR, precipitation and maximum temperature. Recent conditions were found more important north of the study area, whereas, in the middle and southern regions, precedent year conditions were of higher importance. SHAP dependence plots were able to depict feature-output relationships. Using these plots, it was observed that LSTM represented the fire-predictor relationship correctly only for a few variables. For feature interactions, a 3D extension of SHAP dependence plot with added color visual variable was found to be the best visualisation technique. Visualisation of LSTM helped with understanding how the model is learning and which variables were modelled correctly. From here, further improvements could be applied leading to increasing trust in machine learning approaches.

Increasing the accuracy of tree species classification in Sentinel-2 data through better informed feature selection

Study

MSc Geoinformation Technologies

Date of defence

15.10.2021

Supervisor

Matthias Forkel, Dzhaner Emin (iABG mbH)

Abstract

The tree species composition of forests plays a central role in sustainable forestry and nature conservation. With the help of multispectral remote sensing data and machine learning methods, tree species can be classified objectively and over a wide area with good results, as previous studies have shown. The aim of this master thesis is to increase the classification accuracy for tree species based on Sentinel-2 data using informed feature selection. The classification is based on inventory data of Thuringia, and the study area is located in western Thuringia and the five most common tree species there are classified. To increase the classification accuracy of the tree species, their spectral behaviour, phenology and topography are included in the classification. For this purpose, the Sentinel-2 scenes were first preprocessed with the free processing software FORCE and the spectral and phenological features were then derived from the Sentinel-2 time series and the topographic features from the SRTM-DEM. These were then used to conduct an exploratory data analysis to better understand the relationships between tree species and their species-specific characteristics, especially to gain insights for their classification. Finally, the tree species were classified in different random forest models with their spectral, phenological and topographic features and the importance of these features for their differentiation was determined using the "Mean Decrease Gini". The investigations show that the integration of the phenological and topographical features increased the classification accuracy by about 5 percent to 90 percent. The most important input features include the phenological features of the beginning and peak of the growing season, as well as the elevation and spectral bands of the SWIR. By applying a "second majority vote" on the different values of the spectral features, the classification accuracy could even be increased by several percent. The developed workflow for tree species classification and the gained knowledge should serve as inspiration for future tree species classifications and help to better understand tree species characteristics.

© Johannes Albert

Evaluation of the real land use in ATKIS with digital orthophotos and convolutional neural networks

Study

MSc Geoinformation Technologies

Date of defence

08.10.2021

Supervisor

Matthias Forkel, Gotthard Meinel (Leibniz Institute of Ecological Urban and Regional Development)

Abstract

The ATKIS-Basis-DLM is one of the most important databases in Germany regarding the modelling of the earth's surface. Based on the attribute "Real land use", numerous sustainability indicators are calculated in the IÖR-Monitor of the Leibniz Institute of Ecological Urban and Regional Development. However, there are still no reliable data on the quality of the Real land use.

The aim of this thesis was therefore to determine the quality of the recording of the Real land use in ATKIS. Regarding a possible automation of the process, it was tested in parallel whether and to what extent machine learning methods are suitable to take over this task. For this purpose, an artificial neural network was programmed to determine the Real land use based on high-resolution ortho-air images.

The starting point of the investigations was a manual classification of the Real land use in a randomly selected sample of 1,000 aerial images from the years 2012 and 2019. The basic selection of the sample was based on four different types of land use from the categories of non-urban and settlement. Both ATKIS and the neural network were compared with the manually created reference. It was found that ATKIS had the correct Real land use in 93.98% of cases examined. The neural network was able to detect the correct usage in 93.7% of cases. The general weaknesses of the two products are partially complementary to each other, so in the future a combined procedure could be developed to improve the quality of the Real land use in ATKIS.

© Denis Reiter

Estimating tree heights from mangrove forests using machine learning algorithms on combined Sentinel-1 and Sentinel-2 data

Study

MSc Geoinformation Technologies

Date of defence

30.09.2021

Supervisor

Matthias Forkel,  Uday Pimple (King Mongkuts University of Technology, Thonburi, Thailand)

Abstract

Mangrove forests are highly productive ecosystems that also offer coastal and inland protection against natural hazards such as tsunamis or hurricanes. As they are also one of the most threatened ecosystems, several projects now focus on natural regeneration and rehabilitation of formerly destroyed mangrove forests. With mangrove tree height, monitoring, regeneration and rehabilitation efforts can be observed and optimised as well supporting the estimation of above-ground biomass and analysing health status and effectiveness of CO₂ sequestration. Furthermore, tree heights can be used as a level of protection against natural hazards. Field measurements in Trat, Thailand, resulted in three transect lines that run perpendicular to the coast and consist in sum of 59, 10 m × 10 m, plots containing the mean tree height per plot. With two Sentinel-2 scenes, the spectral vegetation indices Normalised Difference Vegetation Index (NDVI) and Chlorophyll Vegetation Index (CVI) were calculated as well as the red-edge band "B8A" was extracted for the analysis. Mean Vertical-Horizontal (VH)-backscatter coefficients of four low tide ground range detected Sentinel-1 scenes support the Sentinel-2 variables as independent observations for a Random Forest Regression (RFR) and a Support Vector Regression (SVR). A site-specific division into training and test plots resulted in meaningful validation results with Root Mean Squared Errors (RMSE) of approx. 1.60 m and two models that can be used to estimate mangrove tree heights of the entire mangrove forest in Trat. A tree height distribution map shows that vertical structural diversity varies along the intertidal zones. Additionally, a three-fold cross validation gave an overview of possible accuracies that a model can achieve, depending on the selection of training and test samples. When calculating the average tree height from the estimations of all three models from each fold, a standard deviation (SD) map can reveal prediction errors. Especially the RFR-SD map was able to detect uncertainties because of clouds in a used Sentinel-2 scene. A comparison of the mangrove tree height distribution map with a species distribution map revealed that regeneration efforts in Trat were successful.

© Nadja Krause

Estimating high-resolution land surface temperature by integrating Landsat-8, Sentinel-1 and Sentinel-2 using a Random Forest algorithm

Study

MSc Geodesy

Date of defence

20.09.2021

Supervisor

Matthias Forkel, Anette Eltner

Abstract

Land Surface Temperature (LST) observations enable a better understanding of the ecosystem since LST is connected to water, energy and carbon fluxes. A couple of satellite systems provide relevant information for the calculation of LST using thermal infrared (TIR) bands. However, the data is available either at high spatial or temporal resolution which disables monitoring approaches at small scale. In order to fill the gap between the spatial and temporal resolution, downscaling algorithms have been developed, reaching a spatial resolution of 30m. The aim of this master thesis is to develop an algorithm that is able to contribute higher spatial resolution of 10m getting closer to monitoring approaches at small scale. The combination of both Sentinel-1 (S1) and Sentinel-2 (S2) data is a promising benefit since both contribute high spatial resolution. Further, S1 data is sensitive to the water content in vegetation and therefore correlates with LST. Together with a digital elevation model (DEM), features derived from S1 and S2 serve as predictors in order to train a Random Forest model. Four strategies were tested, each consisting in a set of features, an evaluation method and a specified training and target resolution. The first strategy uses the same satellite scene at different spatial resolution. Therefore, a Random Forest model trains features against Sentinel-3 (S3) LST at 1000m, while it is tested against Landsat-8 (L8) LST at 30m. Spatial Cross Validation (SCV) serves as evaluation method in all other strategies, separating the study area into four subsets. In doing so, they train a Random Forest model where the training dataset holds three subsets while the remaining one acts as testing dataset. These strategies are conducted with data at 1000m on the one hand and on the other with data at 30m spatial resolution. Since the datasets at 30m are trained against L8 LST, S3 LST is included as feature in one of the strategies. Best results were achieved with spatial cross validation at 30m resolution containing B12, DEM, NDVI, NDWI, SRWI and S3 LST as feature. Comparing predicted and observed LST, the Root Mean Square Error (RMSE) is 1.62K while the coefficient of determination (R²) is 0.73 and the correlation coefficient amounts to 0.86. In general, the trained Random Forest model tends to predict lower temperatures for urban areas and higher ones for forests. Features derived from S1 do not improve the results. The reason  could be found in different acquisition times of the data or that other research use soil moisture derived from S1 backscatter instead of S1 observations themselves. However, spatial cross validation was not found to be well suited as evaluation method and could have an impact on the results making them less reliable. The outcomes also present that some information is missing to fully explain the relationship between LST and features. To solve this issue, other combinations of features and the addition of new ones, e.g. soil moisture, fractional green vegetation cover and air temperature, should be considered.

Assessing environmental controls on phenology with dense Landsat and Sentinel-2 time series in Andean Araucaria forests of Chile

Study

MSc Geodesy

Date of defence

19.08.2021

Supervisor

Matthias Forkel,  Jaime Hernández (Universidad de Chile)

Abstract

Araucaria araucana, commonly called “monkey puzzle tree”, is an endemic, evergreen conifer native to the temperate woodlands of the South American Andes and has a high cultural importance for indigenous tribes and Chilean identity. It typically occurs together with several southern beech species (Nothofagus spp.), forming the characteristic Araucaria-Nothofagus forests. Human exploitation and lack of preservation have led to widespread landscape fragmentation in south-central Chile and Argentina, their only natural habitat. The effects of ongoing climate change are another major threat to the forest ecosystem. Analysing the phenological traits of terrestrial vegetation is one of the key factors in assessing long- and short-term trends of plant growth. However, no detailed study has ever been conducted that investigates the phenology of Araucaria forest stands. In this master’s thesis, the natural phenological variability of undisturbed Araucaria forests in a Chilean national park was analysed between 2016 and 2020 using high-resolution satellite imagery. Dense NDVI and EVI time series of Landsat 8 and Sentinel-2 observations were smoothed and interpolated and phenological parameters, such as the onsets of greenness and senescence were extracted after carefully examining various suitable techniques. These metrics were then closely evaluated and compared with variations in topography, climate, species composition and tree height using a Random Forest model in order to identify the drivers of phenology changes. The results showed high dependencies of start and end of season on elevation with forest stands over 1400 m experiencing an increasingly delayed reaction. The maximum value of vegetation greenness throughout the year was also positively correlated with the timing of phenological events. Furthermore, it was identified that growth started earlier on average in areas where the dominant species is Nothofagus dombeyi and later where Nothofagus antarctica dominates the forest. A possibly related finding could be made for tree height with taller trees experiencing an advanced start of season and vice versa. End-of-season events showed much fewer spatial differences. Temporally, no explicit trend was detected in the study period, but smaller connections to climate could be revealed for most years. It was also discovered that EVI serves as a much better indicator of plant productivity in the region than NDVI. The results prove that several phenological differences exist within the complex Araucaria forest ecosystem. These characteristics and their influence factors help to better understand the dynamics of untouched forest stands in the region, especially with regard to future challenges caused by climate change and can be a basis for local authorities in developing better forest management and conservation strategies within and outside of protected areas. As part of a joint project between TU Dresden and Universidad de Chile, this master’s thesis also delivers important contributions to ongoing research activities about the impact of human influence on the Araucaria forest ecosystem.

© Eric Koszor

Capability of CNNs for analyzing land cover using historical CORONA data

Study

MSc Geography

Date of defence

15.10.2020

Supervisor

Matthias Forkel,   Pierre Karrasch

Abstract

Land cover and the climate system are closely interlinked components. This interlinkage means that the decades-long change in land cover also plays a significant role in climate change. With the help of remote sensing and the civilian Landsat satellites, it has been possible to monitor land cover since 1972 continuously. There have also been earlier efforts to monitor the earth with satellites. The military satellite program CORONA was able to take single-channel images of the earth's surface from 1959 to1972. These images were not made available to the public until 1995. Until today they are a barely used data source for land cover change worldwide. This work aims to develop a Convolutional Neural Network (CNN), which can automatically extract land cover from CORONA images. The  CNN  was trained and evaluated based on the study areas Saxon Switzerland and Zittau Mountains. Germany in 1965. CNN is a method of machine learning, which is particularly suitable for the processing of image data. In remote sensing, CNNs have already been successfully used for object recognition and classification of multispectral data. So far, however, there are no detailed studies on the use of single-channel images for land cover classification.

For this study, three different input data were generated from the CORONA data: the first is the unchanged, original pixel values of the CORONA images; the second is calculated GLCM texture measures, which include the neighbouring pixels for calculation; the last are the original pixel values together with the landscape metrics of a previous segmentation. Based on these input data and existing information about the land cover of the study area in 1965, image sections for the individual land cover classes were extracted. The learning and training of CNN needed these image sections. In total, six different training options were investigated. The first finding from the training was the very low accuracy of the CNNs that trained on the landscape metrics. With an accuracy of less than 50%, they were far below the results of the other two input data of 85% to 90%. As a result, the landscape metrics were not considered for the rest of the study. The training accuracies were compared, and the best three with accuracies of 83% were applied to the two study areas. One model classified on the Original Pixel Values and the other two on the Texture Measures.

Using an iterating search window, the CNN models were deployed to the study areas. Classification accuracies between 69% and 75% were achieved for the study area Saxon Switzerland. The application to the study area Zittau Mountains did not achieve such high values. The Original Pixel Values attained an accuracy of 61% while the Texture Measures only yielded an accuracy of 1.8% and 3.5%. These results indicate an overtraining of the model since drastically reduced accuracies were achieved in the Zittau Mountains region. Reasons for this could be an unequal distribution of the training data for the individual land cover classes, differences between the two CORONA images, and homogeneity of the image values between the individual land cover classes. A concluding analysis of the land cover change between 1965 and 2012 showed an increase in urban and agricultural areas. The open-cast mining and the subsequent natural conversion characterises the Zittau Mountains area. The forested area has not changed significantly in the 47 years.

© Susanne Otto

Potential of multi-temporal Sentinel-2 data to retrieve leaf live-fuel moisture content

Study

BSc Geodesy and Geoinformation

Date of defence

08.10.2020

Supervisor

Matthias Forkel, Luisa Schmidt

Abstract

The objectives of this thesis are the evaluation of already proven spectral indices, development of new approaches using the red-edge and the development of
multi-variate regression models in terms of Live Fuel Moisture Content (LFMC)
approximation. The data used in this thesis is multi-spectral bottom of the atmosphere Sentinel-2 data and LFMC in-situ data from the National Fuel Moisture Database (NFMD). The results of other studies regarding correlations between short wave infrared (SWIR) indices and LFMC cannot be confirmed. However the results of indices using bands of the red-edge or near red-edge bands are promising in that regard. In addition to this, when developing new models based on Sentinel-2 data for LFMC approximation, a new method is developed. For the purpose of model development, a new method is being developed that uses explorative factor analysis (EFA) to derive latent factors of strongly correlated
indices in order to solve the problem of multi-variant models with multicollinearity. In this context, a workflow for the development with this method is proposed. Furthermore, these derived latent factors are combined with other indices in multi-variate regression models. These are developed for the land cover types bushland and grassland and yield both similar results with R² = 0.610 for shrubland and R² = 0.666 for grassland. The developed models will also be tested individually on individual vegetation species and locations. These tests show the robustness of the developed latent factors and the developed LFMC models in comparison to the individual indices.

Developing Data Cylinders to map changes and feedbacks between Arctic Sea Ice and Vegetation

Study

MSc Cartography

Date of defence

28.09.2020

Supervisor

Matthias Forkel, Dr. Paulo Raposo (University of Twente)

Abstract

Arctic sea ice hits second-lowest level on record in 2020 and a “new Arctic” climate era is commencing (Landrum & Holland, 2020). Living on a planet which is facing a global climate change, the understanding and the communication of those vast environmental changes have great importance. The objective of this thesis is the development of a three-dimensional spatio-temporal visual, in a cylindrical shape due to the circularity of the Arctic region, depicting the analysis of the changes of the sea ice and vegetation in the Arctic. Level4 remote sensing data is used for the vegetation and sea ice concentration in a pan-arctic scale for the needs of this research. Its main outcomes are the changes themselves mirrored through their anomalies, and the creation of one cartographic product which illustrates these phenomena in the best suitable way.The representation of time and space in a single visual has been viewed as a challenge in the field of modern cartography for many decades. While data-cubes have been proposed for regional or even global geographical data visualisation, the polar regions seem not suitable for such cubes. The proposed cartographic representation of the Arctic space and time into “Data-Cylinders” responds to the challenging befitting of physical phenomena along space and time.The context of the “Data-Cylinders” includes the visualisation of the changes of climatic and physical phenomena in the form of time-series. A cylinder in this case could be described as many circular maps on top of the other in chronological order. “Space” in this scenario is placed across cycles, while “time” is given along the height of the cylinder.