Random Forest Regression Tree Research Articles

Predicting Student Achievement via Machine Learning: Evidence from Turkish Subset of PISA

This study seeks to identify the determinants of academic performance in mathematics, science, and reading among Turkish secondary school students. Using data from the OECD's PISA 2018 survey, which includes several student- and school-level variables as well as test scores, we employed a range of supervised machine learning methods specifically ensemble decision trees to assess their predictive performance. Our results indicate that the boosted regression tree (BRT) method outperforms other methods bagging and random forest regression trees. Notably, the BRT highlights the importance of general secondary education programs over vocational and technical (VAT) education in predicting academic achievement. Moreover, both characteristics specific to student and school environment are demonstrated to be significant predictors of academic performance in all subject areas. These findings contribute to the development of evidence-based educational policies in Turkey.

Application of machine learning algorithms for nonlinear system forecasting through analytics — A case study with mining influenced water data

Various techniques have been researched and introduced in water treatment plants to optimise treatment and management processes. This paper presents a solution that can help treatment plants to work more effectively and reach their mine water management goals. Using Python 3.7.1 programming language within an Anaconda 4.11.0 platform, neural networks and regression tree algorithms were compared to find the best performing model after the data had undergone robust data pre-processing and exploratory data analysis statistical techniques. The main aim was to use this best performing model to forecast mining influenced water (MIW) parameters. This approach will help the treatment plant operators in knowing the future MIW chemistry, and they can eventually plan ahead of time what chemicals and methods to use to treat and manage polluted MIW. Westrand mine pool water near Randfontein, South Africa is used as a case study, in which historical data (2016–2021) from shaft № 9 is used to train and test the algorithms. These algorithms included the artificial neural network (ANN), deep neural network (DNN), gradient boosting and random forest regression trees, while the multivariate long short-term memory (LSTM) was used to generate new data for the best performing algorithm. Different data pre-processing approaches were explored, including data interpolation and anomaly detection. These processes were carried out to highlight the most important part of completing a machine learning related project, which is data analytics. Finally, the random forest regression tree algorithm showed the overall best performance and was used to forecast Fe and acidity concentrations of MIW for 60 days. It could be shown that artificial intelligence techniques are capable to optimise and forecast mine water treatment plant parameters, and it is imperative to perform robust statistical analysis on the data before attempting to build forecasting models.

Using machine learning models to assess the population dynamic of the freshwater invasive snail Physa acuta Draparnaud, 1805 (Gastropoda: Physidae) in a tropical urban polluted streams-system

The poor understanding of the spread conditions and the geographical patterns of the exotic freshwater snail Physa acuta in sub-Saharan Africa urban streams makes it difficult to design and implement sustainable management tools, especially in deteriorating global and local environments. This study aimed to determine the ecological preferences and the major environmental variables associated with P. acuta abundance, occurrence and distribution in Douala urban streams-system using modelling approaches based on discriminant factor analysis, canonical Spearman correlation and random forest tree models. We inductively define four clusters for snail’s distribution along a geographical and upstream–downstream pollution gradient. Snail’s specimens were collected monthly over a 13-month period in two forest and ten urban sites using a long-handled net (30x30cm side, 400-µm mesh). Meanwhile, measurements of the physicochemical properties of water and streambed sediment were taken. In addition, ten metal contents were determined in P. acuta tissues to assess its ability of bioaccumulation of heavy metals. Results revealed that urban streams of Douala are adversely impaired, especially downstream of industrial effluents. Discriminant factor analysis and random forest regression trees model showed that dissolved oxygen, biochemical oxygen demand, alkalinity, salinity, pH, water depth, wetted width, zinc, iron, cadmium, encumbrance rate of the sediment and sediment types (high sandy mud, sandy mud and muddy sand) were the major environmental variables responsible for structuring P. acuta distribution. Abundances decline drastically following pollution gradient, with great decrease observed in juvenile size classes. High contents of metals were detected in P. acuta tissues, especially downstream of industrial effluents. This study highlights that P. acuta spread in Douala urban-streams system is structured by key environmental variables related to an upstream–downstream pollution gradient. Therefore management plans should emphasize on proper disposal ways of industrial wastes, and stakeholders should take responsible actions in water pollution control and implement the polluter-payer principle.

Mapping Chinese annual gross primary productivity with eddy covariance measurements and machine learning.

Annual gross primary productivity (AGPP) is the basis for grain production and terrestrial carbon sequestration. Mapping regional AGPP from site measurements provides methodological support for analysing AGPP spatiotemporal variations thereby ensures regional food security and mitigates climate change. Based on 641 site-year eddy covariance measuring AGPP from China, we built an AGPP mapping scheme based on its formation and selected the optimal mapping way, which was conducted through analysing the predicting performances of divergent mapping tools, variable combinations, and mapping approaches in predicting observed AGPP variations. The reasonability of the selected optimal scheme was confirmed by assessing the consistency between its generating AGPP and previous products in spatiotemporal variations and total amount. Random forest regression tree explained 85 % of observed AGPP variations, outperforming other machine learning algorithms and classical statistical methods. Variable combinations containing climate, soil, and biological factors showed superior performance to other variable combinations. Mapping AGPP through predicting AGPP per leaf area (PAGPP) explained 86 % of AGPP variations, which was superior to other approaches. The optimal scheme was thus using a random forest regression tree, combining climate, soil, and biological variables, and predicting PAGPP. The optimal scheme generating AGPP of Chinese terrestrial ecosystems decreased from southeast to northwest, which was highly consistent with previous products. The interannual trend and interannual variation of our generating AGPP showed a decreasing trend from east to west and from southeast to northwest, respectively, which was consistent with data-oriented products. The mean total amount of generated AGPP was 7.03 ± 0.45 PgC yr-1 falling into the range of previous works. Considering the consistency between the generated AGPP and previous products, our optimal mapping way was suitable for mapping AGPP from site measurements. Our results provided a methodological support for mapping regional AGPP and other fluxes.

A dataset of annual gross primary productivity in China’s terrestrial ecosystems during 2000-2020

The annual gross primary productivity (AGPP) is the basis of food production and carbon sequestration in terrestrial ecosystems. An accurate assessment of regional AGPP can provide a theoretical basis for analyzing the spatiotemporal variation of AGPP and ensuring regional food security and mitigating climate change trends. Based on Chinese Flux Observation and Research Network (ChinaFLUX) measurements and public datasets, we produced a dataset of annual gross primary productivity over China’s terrestrial ecosystems was constructed. In combination with biological, climatic, and soil factors, we used the random forest regression tree to construct the assessment model of China AGPP by simulating the AGPP of unit leaf area. The dataset of annual gross primary productivity over China’s terrestrial ecosystems during 2000-2020 was generated with a spatial resolution of 30arcsecond and a data format of tiff. The dataset can provide validation data for model simulation, as well as data support for regional productivity, ecological quality, and assessment and management of terrestrial carbon sinks.

Modelling important areas for breeding waders as a tool to target conservation and minimise conflicts with land use change

Breeding waders are high profile species of conservation concern because of their declining populations. Their future depends on the outcomes of land-use policy and local management decisions. However, the low spatial resolution of extensive data on wader occurrence is poorly suited to directing conservation initiatives or to minimising detrimental impacts arising from land-use changes such as forest expansion. We used statistical models to produce high-resolution maps of predicted wader abundance in Britain and tested whether these were sufficiently accurate to be used for decision making. Random forest regression trees were developed using Bird Atlas data modelled with a range of environmental data sets to predict the relative abundances of ten species of breeding wader across Britain at 1-km square resolution. Similar analytical frameworks could usefully be applied in other geographical areas.Correlations with Bird Atlas metrics and other independent data indicate that these predictive models worked best for Curlew Numenius arquata and Oystercatcher Haematopus ostralegus. Model performance was poorer for species with more restricted distributions and/or habitat requirements, such as Greenshank Tringa nebularia and Ringed Plover Charadrius hiaticula. The precision of model predictions was also limited by the rescaling of data to 1-km square resolution.To facilitate practical and consistent interpretation by stakeholders influencing the status of breeding waders, model outputs were categorised into five discrete strata of relative abundance. These strata were used to produce sensitivity maps (available at https://app.bto.org/wader-map/), primarily to inform high-level policy decisions. The model outputs were also used to investigate broad-scale associations of breeding wader distribution with landscapes and land uses. These show that high proportions of populations (relative to their extent) are supported by areas with comparatively low predation risk (e.g. islands and grouse moors) and also by nature reserves. At more local scales, the sensitivity maps can be used to help decide on the levels of scrutiny required to assess development proposals (e.g. for establishment of new tree planting).

Abstract 13075: Initial Physical Activity Level is a Strong Predictor of Adherence to the Physical Activity Goal: A Secondary Analysis of the SMARTER Weight Loss Trial

Background: Over a third of excess coronary heart disease mortality in the US is related to inadequate physical activity (PA). Use of wearable activity trackers is beneficial for PA improvement in many populations. However, predictors of long-term adherence to PA remain unknown. Hypothesis: The 1 st week PA levels will be the strongest predictors of long-term adherence to the PA goal in adults with overweight or obesity in a mHealth weight loss intervention trial. Methods: A secondary analysis of a 12-month behavioral weight loss trial (SMARTER) that randomized 502 participants to either self-monitoring with tailored feedback messages (SM+FB, n=251) or self-monitoring alone (SM, n=251) was conducted. At baseline, each participant received a Fitbit Charge 2 and PA goal of ≥150 minutes of moderate or vigorous activity (MVPA) per week. The outcome of interest for this analysis was long-term PA adherence, i.e., the number of weeks adherent to PA goal over 12 months. Twenty-six explanatory variables were considered, including treatment assignment; baseline socio-demographics; health history; body mass index (<30 vs. ≥30 kg/m 2 ); psychosocial factors; and1 st week PA levels (<150 vs. ≥150 MVPA min/week). Machine learning methods (i.e., random forest regression trees and lasso regression) were used to identify predictors of long-term PA adherence. Results: The sample ranged in age from 19 to 75 years (57% were <50 years old), and was mostly Caucasian (83%), female (80%) and had a mean BMI 33.7 ± 4.0 kg/m 2 . The sample was randomly split 4:1 into training (n=394) and validation subsamples (n=104) which were comparable by treatment assignment. Both random forest and lasso regression identified 1 st week PA levels as the most important predictor of long-term PA adherence. The findings were confirmed in the validation cohort with an out-of-sample R-square of 40.2% from random forest and a R-square of 47.9% from lasso regression, suggesting good predictive abilities. Conclusion: This analysis confirmed higher initial PA levels as the strongest predictor of long-term PA adherence. The findings support screening for PA at study entry and developing tailored interventions for those with lower initial PA levels.

Predicting energy cost of public buildings by artificial neural networks, CART, and random forest

The paper deals with modeling the cost of energy consumed in public buildings by leveraging three machine learning methods: artificial neural networks, CART, and random forest regression trees. Energy consumption is one of the major issues in global and national policies, therefore scientific efforts in creating prediction models of energy consumption and cost are highly important. One of the largest energy consumers in every state is its public sector, consisting of educational, health, public administration, military, and other types of public buildings. Recent technologies based on sensor networks and Big data platforms enable collection of large amounts of data that could be used to analyze energy consumption and cost. A real data from Croatian public sector is used in this paper including a large number of constructional, energetic, occupational, climate and other attributes. The algorithms for data pre-processing and modeling by optimizing parameters are suggested. Three strategies were tested: (1) with all available variables, (2) with a filter-based variable selection, and (3) with a wrapper-based variable selection which integrates Boruta algorithm and random forest. Prediction models of energy cost are created using two approaches: (a) comparative usage of artificial neural networks and two types of regression trees, CART and random forest, and (b) integration of RF-Boruta variable selection and machine learning methods for prediction. A cross-validation procedure was used to optimize the artificial neural network and regression tree topology, as well to select the most appropriate activation function. Along with creating a prediction model, the aim of the paper was also to extract the relevant predictors of energy cost in public buildings which are important in planning the construction or renovation of buildings. The results have shown that the second approach which integrates machine learning with Boruta method, where the random forest algorithm is used for both variable reduction and prediction modeling, has produced a higher accuracy of prediction than the individual usage of three machine learning methods. Such findings confirm the potential of hybrid machine learning methods which are suggested in previous research, but in favor of random forest method over CART and artificial neural networks. Regarding the variable selection, the model has extracted heating and occupational data as the most important, followed by constructional, cooling, electricity, and lighting attributes. The model could be implemented in public buildings information systems and their IoT networks within the concept of smart buildings and smart cities.

Using Natural Language Processing to Assess Text Usefulness to Readers: The Case of Conference Calls and Earnings Prediction

We examine whether support vector regressions (SVR), supervised LDA (sLDA), random forest regression trees (RF), and ‘tone’ extract narrative content from conference calls that correlates with useful information that a human reader would identify. We find that each narrative-content measure (along with a composite measure) explains a portion of analyst-forecast revisions for quarter q 1 issued after the conference call in quarter q. Correlation with analyst-forecast revisions improves when the composite measure adapts to context (positive/negative returns; high variance/low variance) and ignores sparse words. The correlation is comparable and incremental to that of financial signals (cash-flow changes, earnings surprises, and management forecasts), which suggests that the narrative content of conference calls as extracted by readers is economically significant. Our results suggest that models of narrative content have reasonable construct validity and that this validity is likely to be improved by further thought on the unique characteristics of text.

Constructing a Finer-Resolution Forest Height in China Using ICESat/GLAS, Landsat and ALOS PALSAR Data and Height Patterns of Natural Forests and Plantations

Monitoring forest height is crucial to determine the structure and biodiversity of forest ecosystems. However, detailed spatial patterns of forest height from 30 m resolution remotely sensed data are currently unavailable. In this study, we present a new method for mapping forest height by combining spaceborne Light Detection and Ranging (LiDAR) with imagery from multiple remote sensing sources, including the Landsat 5 Thematic Mapper (TM), the Phased Array L-band Synthetic Aperture Radars (PALSAR), and topographic data. The nationwide forest heights agree well with results obtained from 525 independent forest height field measurements, yielding correlation coefficient, root mean square error (RMSE), and mean absolute error (MAE) values of 0.92, 4.31 m, and 3.87 m, respectively. Forest heights derived from remotely sensed data range from 1.41 m to 38.94 m, with an average forest height of 16.08 ± 3.34 m. Mean forest heights differ only slightly among different forest types. In natural forests, conifer forests have the greatest mean forest heights, whereas in plantations, bamboo forests have the greatest mean forest heights. Important predictors for modeling forest height using the random forest regression tree method include slope, surface reflectance of red bands and HV backscatter. The uncertainty caused by the uneven distribution of Geoscience Laser Altimeter System (GLAS) footprints is estimated to be 0.64 m. After integrating PALSAR data into the model, the uncertainty associated with forest height estimation was reduced by 4.58%. Our finer-resolution forest height could serve as a benchmark to estimate forest carbon storage and would greatly contribute to better understanding the roles of ecological engineering projects in China.

In-Situ and Remote Sensing Platforms for Mapping Fine-Fuels and Fuel-Types in Sonoran Semi-Desert Grasslands

Fire has historically played an important role in shaping the structure and composition of Sonoran semi-desert grassland vegetation. Yet, human use and land management activities have significantly altered arid grassland ecosystems over the last century, often producing novel fuel conditions. The variety of continuously updated satellite remote sensing systems provide opportunities for efficiently mapping combustible fine-fuels and fuel-types (e.g., grass, shrub, or tree cover) over large landscapes that are helpful for evaluating fire hazard and risk. For this study, we compared field ceptometer leaf area index (LAI) measurements to conventional means for estimating fine-fuel biomass on 20, 50 m × 20 m plots and 431, 0.5 m × 0.5 m quadrats on the Buenos Aires National Wildlife Refuge (BANWR) in southern Arizona. LAI explained 65% of the variance in fine-fuel biomass using simple linear regression. An additional 19% of variance was explained from Random Forest regression tree models that included herbaceous plant height and cover as predictors. Field biomass and vegetation measurements were used to map fine-fuel and vegetation cover (fuel-type) from plots on BANWR comparing outcomes from multi-date (peak green and dormant period) Worldview-3 (WV3) and Landsat Operational Land Imager (OLI) imagery. Fine-fuel biomass predicted from WV3 imagery combined with terrain information from a digital elevation model explained greater variance using regression tree models (65%) as compared to OLI models (58%). Vegetation indices developed using red-edge bands as well as modeled bare ground and herbaceous cover were important to improve WV3 biomass estimates. Land cover classification for 11 cover categories with high spatial resolution WV3 imagery showed 80% overall accuracy and highlighted areas dominated by non-native grasses with 87% user’s class accuracy. Mixed native and non-native grass and shrublands showed 59% accuracy and less common areas dominated by native grasses on plots showed low class accuracy (23%). Digital data layers from WV3 models showed a significantly positive relationship (r2 = 0.68, F = 119.2, p < 0.001) between non-native grass cover (e.g., Eragrostis lehmanniana) and average fine-fuel biomass within refuge fire management units. Overall, both WV3 and OLI produced similar fine-fuel biomass estimates although WV3 showed better model performance and helped characterized fine-scale changes in fuel-type and continuity across the study area.

Seasonal forecasting of hydrological drought in the Limpopo Basin: a comparison of statistical methods

Abstract. The Limpopo Basin in southern Africa is prone to droughts which affect the livelihood of millions of people in South Africa, Botswana, Zimbabwe and Mozambique. Seasonal drought early warning is thus vital for the whole region. In this study, the predictability of hydrological droughts during the main runoff period from December to May is assessed using statistical approaches. Three methods (multiple linear models, artificial neural networks, random forest regression trees) are compared in terms of their ability to forecast streamflow with up to 12 months of lead time. The following four main findings result from the study. 1. There are stations in the basin at which standardised streamflow is predictable with lead times up to 12 months. The results show high inter-station differences of forecast skill but reach a coefficient of determination as high as 0.73 (cross validated). 2. A large range of potential predictors is considered in this study, comprising well-established climate indices, customised teleconnection indices derived from sea surface temperatures and antecedent streamflow as a proxy of catchment conditions. El Niño and customised indices, representing sea surface temperature in the Atlantic and Indian oceans, prove to be important teleconnection predictors for the region. Antecedent streamflow is a strong predictor in small catchments (with median 42 % explained variance), whereas teleconnections exert a stronger influence in large catchments. 3. Multiple linear models show the best forecast skill in this study and the greatest robustness compared to artificial neural networks and random forest regression trees, despite their capabilities to represent nonlinear relationships. 4. Employed in early warning, the models can be used to forecast a specific drought level. Even if the coefficient of determination is low, the forecast models have a skill better than a climatological forecast, which is shown by analysis of receiver operating characteristics (ROCs). Seasonal statistical forecasts in the Limpopo show promising results, and thus it is recommended to employ them as complementary to existing forecasts in order to strengthen preparedness for droughts.

Multi-scale assessment of human-induced changes to Amazonian instream habitats

Land use change and forest degradation have myriad effects on tropical ecosystems. Yet their consequences for low-order streams remain very poorly understood, including in the world´s largest freshwater basin, the Amazon. Determine the degree to which physical and chemical characteristics of the instream habitat of low-order Amazonian streams change in response to past local- and catchment-level anthropogenic disturbances. To do so, we collected field instream habitat (i.e., physical habitat and water quality) and landscape data from 99 stream sites in two eastern Brazilian Amazon regions. We used random forest regression trees to assess the relative importance of different predictor variables in determining changes in instream habitat response variables. Multiple drivers, operating at multiple spatial scales, were important in determining changes in the physical habitat and water quality of the sites. Although we found few similarities in modelled relationships between the two regions, we observed non-linear responses of specific instream characteristics to landscape change; for example 20 % of catchment deforestation resulted in consistently warmer streams. Our results highlight the importance of local riparian and catchment-scale forest cover in shaping instream physical environments, but also underscore the importance of other land use changes and activities, such as road crossings and upstream agriculture intensification. In contrast to the property-scale focus of the Brazilian Forest code, which governs environmental regulations on private land, our results reinforce the importance of catchment-wide management strategies to protect stream ecosystem integrity.

Mapping Annual Precipitation across Mainland China in the Period 2001–2010 from TRMM3B43 Product Using Spatial Downscaling Approach

Spatially explicit precipitation data is often responsible for the prediction accuracy of hydrological and ecological models. Several statistical downscaling approaches have been developed to map precipitation at a high spatial resolution, which are mainly based on the valid conjugations between satellite-driven precipitation data and geospatial predictors. Performance of the existing approaches should be first evaluated before applying them to larger spatial extents with a complex terrain across different climate zones. In this paper, we investigate the statistical downscaling algorithms to derive the high spatial resolution maps of precipitation over continental China using satellite datasets, including the Normalized Distribution Vegetation Index (NDVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS), the Global Digital Elevation Model (GDEM) from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), and the rainfall product from the Tropical Rainfall Monitoring Mission (TRMM). We compare three statistical techniques (multiple linear regression, exponential regression, and Random Forest regression trees) for modeling precipitation to better understand how the selected model types affect the prediction accuracy. Then, those models are implemented to downscale the original TRMM product (3B43; 0.25° resolution) onto the finer grids (1 × 1 km2) of precipitation. Finally we validate the downscaled annual precipitation (a wet year 2001 and a dry year 2010) against the ground rainfall observations from 596 rain gauge stations over continental China. The result indicates that the downscaling algorithm based on the Random Forest regression outperforms, when compared to the linear regression and the exponential regression. It also shows that the addition of the residual terms does not significantly improve the accuracy of results for the RF model. The analysis of the variable importance reveals the NDVI related predictors, latitude, and longitude, elevation are key elements for statistical downscaling, and their weights vary across different climate zones. In particular, the NDVI, which is generally considered as a powerful geospatial predictor for precipitation, correlates weakly with precipitation in humid regions.

Climate- and soil-based models of site productivity in eastern US tree species

As concerns rise over potential effects of greenhouse gas related climate change on terrestrial ecosystems, forest managers require growth and yield modeling capabilities responsive to changing climate conditions. Our goal was to develop prediction models of site index for eastern US forest tree species with climate and soil properties as predictors for use in predicting potential responses of forest productivity to climate change. Species-specific site index data from the USDA Forest Service Forest Inventory and Analysis (FIA) program were linked to contemporary climate data and soil properties mapped in the USDA Soil Survey Geographic (SSURGO) database. Random forest regression tree based ensemble prediction models of site index were constructed based on 37 climate-related and 15 soil attributes. In addition to a species-specific site index, aggregate models were developed for species grouped into two broad categories: conifer (softwood) and hardwood (broadleaved) species groups. Species-specific models based on climate and soil predictors explained the most variation in site index of any models tested (R2= 62.5%, RMSE = 3.2 m). Comparable results were found when grouping species into conifer and hardwood groups (R2= 63.9%, RMSE = 4.6 m for conifers; R2= 35.9%, RMSE = 4.2 m for hardwoods). Model predictions based on multiple global circulation models (GCMs) and Intergovernmental Panel on Climate Change (IPCC) development scenarios were tested for statistical significance using bootstrap resampling methods. Results showed significant increases over the 21st century in mean site index for conifers between +0.5 and +2.4 m. Over the same time period, mean hardwood site index showed decreases of as much as −1.7 m for the scenarios tested. The results demonstrate the utility of using climate and soils data in predicting site index across a large geographic region, and the potential of climate change to alter forest productivity in the eastern US. Additional investigation is needed to interpret spatial patterns and ecological relationships related to predictions from this type of model.

Remotely sensed vegetation phenology for describing and predicting the biomes of South Africa

Questions: What are the patterns of remotely sensed vegetation phenology, including their inter-annual variability, across South Africa? What are the phenological attributes that contribute most to distinguishing the different biomes? How well can the distribution of the recently redefined biomes be predicted based on remotely sensed, phenology and productivity metrics? Location: South Africa. Method: Ten-day, 1 km, NDVI AVHRR were analysed for the period 1985 to 2000. Phenological metrics such as start, end and length of the growing season and estimates of productivity, based on small and large integral (SI, LI) of NDVI curve, were extracted and long-term means calculated. A random forest regression tree was run using the metrics as the input variables and the biomes as the dependent variable. A map of the predicted biomes was reproduced and the differentiating importance of each metric assessed. Results: The phenology metrics (e.g. start of growing season) showed a clear relationship with the seasonality of rainfall, i.e. winter and summer growing seasons. The distribution of the productivity metrics, LI and SI were significantly correlated with mean annual precipitation. The regression tree initially split the biomes based on vegetation production and then by the seasonality of growth. A regression tree was used to produce a predicted biome map with a high level of accuracy (73%). Main conclusion: Regression tree analysis based on remotely sensed metrics performed as good as, or better than, previous climate-based predictors of biome distribution. The results confirm that the remotely sensed metrics capture sufficient functional diversity to classify and map biome level vegetation patterns and function.

Using VLAD scores to have a look insight ICU performance: towards a modelling of the errors

Mortality prediction models using logistic regression analysis play a pivotal role in intensive care quality evaluation, allowing a hospital's performance to be compared with a standard. However, when a difference between predicted and observed mortality exists, that is, the numerator of the Variable Life Adjusted Display (VLAD) score, the investigation for a possible explanation could be arduous. In this article we tested the ability of Bayesian Network (BN) to identify factors determining the negative discrepancy between expected and actual outcomes recorded in four Italian intensive care units (ICUs). A BN was implemented to predict the extent of the expected-observed distance quantified by the VLAD score. BN performance was compared with those of a set of tools including Linear Model, Random Forest Regression Tree analysis, Artificial Neural Networks and Support Vector Machine. BN allows the identification of critical areas responsible for bad performance. Compared with other techniques, BN always explains a higher variance percentage and it shows similar or superior discrimination ability. BN, being able to guide interpretation of covariates role by means of a graphic representation of relationships, confirms its utility particularly where many interactions between predictors exist and when a coherent set of theories regarding which variables are related and how is not available.

What limits fire? An examination of drivers of burnt area in Southern Africa

AbstractThe factors controlling the extent of fire in Africa south of the equator were investigated using moderate resolution (500 m) satellite‐derived burned area maps and spatial data on the environmental factors thought to affect burnt area. A random forest regression tree procedure was used to determine the relative importance of each factor in explaining the burned area fraction and to address hypotheses concerned with human and climatic influences on the drivers of burnt area. The model explained 68% of the variance in burnt area. Tree cover, rainfall in the previous 2 years, and rainfall seasonality were the most important predictors. Human activities – represented by grazing, roads per unit area, population density, and cultivation fraction – were also shown to affect burnt area, but only in parts of the continent with specific climatic conditions, and often in ways counter to the prevailing wisdom that more human activity leads to more fire. The analysis found no indication that ignitions were limiting total burnt area on the continent, and most of the spatial variation was due to variation in fuel load and moisture. Split conditions from the regression tree identified (i) low rainfall regions, where fire is rare; (ii) regions where fire is under human control; and (iii) higher rainfall regions where burnt area is determined by rainfall seasonality. This study provides insights into the physical, climatic, and human drivers of fire and their relative importance across southern Africa, and represents the beginnings of a predictive framework for burnt area.