Mika Sulkava

Distance Matrix Based Clustering of the Self-Organizing Map

Clustering of data is one of the main applications of the Self-Organizing Map (SOM). U-matrixis a commonly used technique to cluster the SOM visually. However, in order to be really useful, clustering needs to be an automated process. There are several techniques which can be used to cluster the SOM autonomously, but the results they provide do not follow the results of U-matrixv ery well. In this paper, a clustering approach based on distance matrices is introduced which produces results which are very similar to the U-matrix. It is compared to other SOMbased clustering approaches.

Evaluation of forest nutrition based on large-scale foliar surveys: are nutrition profiles the way of the future?

This paper introduces the use of nutrition profiles as a first step in the development of a concept that is suitable for evaluating forest nutrition on the basis of large-scale foliar surveys. Nutrition profiles of a tree or stand were defined as the nutrient status, which accounts for all element concentrations, contents and interactions between two or more elements. Therefore a nutrition profile overcomes the shortcomings associated with the commonly used concepts for evaluating forest nutrition. Nutrition profiles can be calculated by means of a neural network, i.e. a self-organizing map, and an agglomerative clustering algorithm with pruning. As an example, nutrition profiles were calculated to describe the temporal variation in the mineral composition of Scots pine and Norway spruce needles in Finland between 1987 and 2000. The temporal trends in the frequency distribution of the nutrition profiles of Scots pine indicated that, between 1987 and 2000, the N, S, P, K, Ca, Mg and Al decreased, whereas the needle mass (NM) increased or remained unchanged. As there were no temporal trends in the frequency distribution of the nutrition profiles of Norway spruce, the mineral composition of the needles of Norway spruce needles subsequently did not change. Interpretation of the (lack of) temporal trends was outside the scope of this example. However, nutrition profiles prove to be a new and better concept for the evaluation of the mineral composition of large-scale surveys only when a biological interpretation of the nutrition profiles can be provided.

Assessing and improving the representativeness of monitoring networks: The European flux tower network example

It is estimated that more than 500 eddy covariance sites are operated globally, providing unique information about carbon and energy exchanges between terrestrial ecosystems and the atmosphere. These sites are often organized in regional networks like CarboEurope-IP, which has evolved over the last 15 years without following a predefined network design. Data collected by these networks are used for a wide range of applications. In this context, the representativeness of the current network is an important aspect to consider in order to correctly interpret the results and to quantify uncertainty. This paper proposes a cluster-based tool for quantitative network design, which was developed in order to suggest the best network for a defined number of sites or to assess the representativeness of an existing network to address the scientific question of interest. The paper illustrates how the tool can be used to assess the performance of the current CarboEurope-IP network and to improve its design. The tool was tested and validated with modeled European GPP data as the target variable and by using an empirical upscaling method (Artificial Neural Network (ANN)) to assess the improvements in the ANN prediction with different design scenarios and for different scientific questions, ranging from a simple average GPP of Europe to spatial, temporal, and spatiotemporal variability. The results show how quantitative network design could improve the predictive capacity of the ANN. However, the analysis also reveals a fundamental shortcoming of optimized networks, namely their poor capacity to represent the spatial variability of the fluxes.

Modeling the effects of varying data quality on trend detection in environmental monitoring

Detection of changes in ecosystem characteristics is a principal tool for identifying and understanding the effects of anthropogenic activities on the condition and functioning of ecosystems. It is widely known that temporal trends can be blurred by the imprecision of the data. Research program managers are aware of the difficulties surrounding representative sampling and therefore enforce strict sampling protocols. Standardized sampling can be so effective that the initially much smaller uncertainty in the instrumental analysis becomes substantial. However, until now the effect of the quality of the instrumental analysis on the time required for trend detection has only rarely been quantified. In this paper, we present a novel technique and theoretical computations for the detection of trends in single and combined indices. The theory is clarified with examples from the International Co-operative Programme on Assessment and Monitoring of Air Pollution on Forests (ICP Forests). Moreover, the theoretical computations were made for normalized or scaled distributions and are therefore equally valid outside the field of environmental monitoring. The results show that, when sampling protocols largely reduce the variability of representative sampling, poor quality of the instrumental analysis blurs the data such that environmental monitoring or long-term ecological research programs can lose the ability to detect trends by causing up to three decades-long delay in detecting changes. We can thus conclude that high quality of the instrumental analysis is a prerequisite for a sensitive monitoring program.

Automatic detection of onset and cessation of tree stem radius increase using dendrometer data

Dendrometers are devices, which measure continuously the stem radius of a tree. In this work, we studied the use of cumulative sum (CUSUM) charts for automatically and, thus, objectively determining the onset and cessation dates of radial increase based on dendrometer data. We used data measured in two forest stands in southern Finland to demonstrate the idea and to test the performance of the CUSUM chart. In order to produce reliable results, one has to choose suitable parameter values for the chart. Once configured properly, the method produced results similar to those determined by an expert.

EnvSOM: a SOM algorithm conditioned on the environment for clustering and visualization

In this paper, we present a new approach suitable for analysis of large data sets, conditioned on the environment. Mainly, the envSOM algorithm consists of two consecutive trainings of the self-organizing map. In the first phase, a SOM is trained using every available variable, but only those which characterize the environment are used to compute the winner unit. Therefore, this phase produces an accurate model of the environment. In the second phase, a new SOM is initialized appropriately with information from the codebooks of the first SOM. The new SOM uses all the variables for winner selection. However, in this case the environmental variables are kept fixed and only the remaining ones are involved in the update process. A model of the whole data set influenced by the environmental conditions is obtained in this second phase. The result of this algorithm represents a probability function of a data set, given the environment information. Therefore, it could be very useful in the analysis of processes which have close dependencies on environmental conditions.

Reliability of temperature signal in various climate indicators from northern Europe

We collected relevant observational and measured annual-resolution time series dealing with climate in northern Europe, focusing in Finland. We analysed these series for the reliability of their temperature signal at annual and seasonal resolutions. Importantly, we analysed all of the indicators within the same statistical framework, which allows for their meaningful comparison. In this framework, we employed a cross-validation procedure designed to reduce the adverse effects of estimation bias that may inflate the reliability of various temperature indicators, especially when several indicators are used in a multiple regression model. In our data sets, timing of phenological observations and ice break-up were connected with spring, tree ring characteristics (width, density, carbon isotopic composition) with summer and ice formation with autumn temperatures. Baltic Sea ice extent and the duration of ice cover in different watercourses were good indicators of winter temperatures. Using combinations of various temperature indicator series resulted in reliable temperature signals for each of the four seasons, as well as a reliable annual temperature signal. The results hence demonstrated that we can obtain reliable temperature information over different seasons, using a careful selection of indicators, combining the results with regression analysis, and by determining the reliability of the obtained indicator.

Clustering of the self-organizing map reveals profiles of farm profitability and upscaling weights

Abstract Profitability and other economic aspects of farming in Finland are analyzed using clustering of the self-organizing map. The analysis of profitability bookkeeping data reveals several interesting relationships between the monitored financial variables. Economic profiles of farms are presented based on the clustering, and the findings are confirmed with statistical tests. A weight optimization system is proposed for upscaling financial figures of the sample of profitability bookkeeping farms to the whole country level. The system output is analyzed, and it is confirmed that the most important large and medium-sized enterprises are represented well by the sample. Furthermore, it seems that the utilized arable area is the key factor in guiding the weight optimization process. These findings may turn out to be useful in developing the sampling of bookkeeping farms in the future.

Comparative analysis of power consumption in university buildings using envSOM

Analyzing power consumption is important for economic and environmental reasons. Through the analysis of electrical variables, power could be saved and, therefore, better energy efficiency could be reached in buildings. The application of advanced data analysis helps to provide a better understanding, especially if it enables a joint and comparative analysis of different buildings which are influenced by common environmental conditions. In this paper, we present an approach to monitor and compare electrical consumption profiles of several buildings from the Campus of the University of Leon. The envSOM algorithm, a modification of the self-organizing map (SOM), is used to reduce the dimension of data and capture their electrical behaviors conditioned on the environment. After that, a Sammons mapping is used to visualize global, component-wise or environmentally conditioned similarities among the buildings. Finally, a clustering step based on k-means algorithm is performed to discover groups of buildings with similar electrical behavior.

Combining measurement quality into monitoring trends in foliar nutrient concentrations

Quality of measurements is an important factor affecting the reliability of analyses in environmental sciences. In this paper we combine foliar measurement data from Finland and results of multiple measurement quality tests from different sources in order to study the effect of measurement quality on the reliability of foliar nutrient analysis. In particular, we study the use of weighted linear regression models in detecting trends in foliar time series data and show that the development of measurement quality has a clear effect on the significance of results.