Borough-level COVID-19 forecasting in London using deep learning techniques and a novel MSE-Moran's I loss function.

Following its identification in late 2019, COVID-19 has spread around the globe, and been declared a pandemic. With this in mind, modelling the spread of COVID-19 remains important for responding effectively. To date research has focused primarily on modelling the spread of COVID-19 on national and regional scales with just a few studies doing so on a city and sub-city scale. However, no attempts have yet been made to design and optimize a model explicitly for accurately forecasting the spread of COVID-19 at sub-city scale. This research aimed to address this research gap by developing an experimental LSTM-ANN deep learning model. The model is largely autoregressive in nature as it considers temporally lagged borough-level COVID-19 cases data from the last 9 days, but also considers temporally lagged (i) borough-level NO2 concentration data, (ii) government stringency data, and (iii) climatic data from the last 9 days, as well as non-temporally variable borough-level urban characteristics data when modelling and forecasting the spread of the disease. The model was also encouraged to learn the spatial relationships between boroughs with regards to the spread of COVID-19 by a novel MSE-Moran's I loss function. Overall, the model's performance appears promising and so the model represents a useful tool for assisting the decision making and interventions of governing bodies within cities. A sensitivity analysis also indicated that of the non COVID-19 variables, the government stringency is particularly important in the modelling process, with this being closely followed by the climatic variables, the NO2 concentration data, and finally the urban characteristics data. Additionally, the introduction of the novel MSE-Moran's I loss function appeared to improve the model's forecasting performance, and so this research has implications at the intersection of deep learning and disease modelling. It may also have implications within spatio-temporal forecasting more generally because such a feature may have the potential to improve forecasting in other spatio-temporal applications.

Validating the regional estimates of changes in soil organic carbon by using the data from paired-sites: the case study of Mediterranean arable lands.

BACKGROUND: Legacy data are unique occasions for estimating soil organic carbon (SOC) concentration changes and spatial variability, but their use showed limitations due to the sampling schemes adopted and improvements may be needed in the analysis methodologies. When SOC changes is estimated with legacy data, the use of soil samples collected in different plots (i.e., non-paired data) may lead to biased results. In the present work, N = 302 georeferenced soil samples were selected from a regional (Sicily, south of Italy) soil database. An operational sampling approach was developed to spot SOC concentration changes from 1994 to 2017 in the same plots at the 0-30 cm soil depth and tested. RESULTS: The measurements were conducted after computing the minimum number of samples needed to have a reliable estimate of SOC variation after 23 years. By applying an effect size based methodology, 30 out of 302 sites were resampled in 2017 to achieve a power of 80%, and an α = 0.05. A Wilcoxon test applied to the variation of SOC from 1994 to 2017 suggested that there was not a statistical difference in SOC concentration after 23 years (Z = - 0.556; 2-tailed asymptotic significance = 0.578). In particular, only 40% of resampled sites showed a higher SOC concentration than in 2017. CONCLUSIONS: This finding contrasts with a previous SOC concentration increase that was found in 2008 (75.8% increase when estimated as differences of 2 models built with non-paired data), when compared to 1994 observed data (Z = - 9.119; 2-tailed asymptotic significance < 0.001). This suggests that the use of legacy data to estimate SOC concentration dynamics requires soil resampling in the same locations to overcome the stochastic model errors. Further experiment is needed to identify the percentage of the sites to resample in order to align two legacy datasets in the same area.

New pedotransfer approaches to predict soil bulk density using WoSIS soil data and environmental covariates in Mediterranean agro-ecosystems.

For the estimation of the soil organic carbon stocks, bulk density (BD) is a fundamental parameter but measured data are usually not available especially when dealing with legacy soil data. It is possible to estimate BD by applying pedotransfer function (PTF). We applied different estimation methods with the aim to define a suitable PTF for BD of arable land for the Mediterranean Basin, which has peculiar climate features that may influence the soil carbon sequestration. To improve the existing BD estimation methods, we used a set of public climatic and topographic data along with the soil texture and organic carbon data. The present work consisted of the following steps: i) development of three PTFs models separately for top (0-0.4 m) and subsoil (0.4-1.2 m), ii) a 10-fold cross-validation, iii) model transferability using an external dataset derived from published data. The development of the new PTFs was based on the training dataset consisting of World Soil Information Service (WoSIS) soil profile data, climatic data from WorldClim at 1 km spatial resolution and Shuttle Radar Topography Mission (SRTM) digital elevation model at 30 m spatial resolution. The three PTFs models were developed using: Multiple Linear Regression stepwise (MLR-S), Multiple Linear Regression backward stepwise (MLR-BS), and Artificial Neural Network (ANN). The predictions of the newly developed PTFs were compared with the BD calculated using the PTF proposed by Manrique and Jones (MJ) and the modelled BD derived from the global SoilGrids dataset. For the topsoil training dataset (N = 129), MLR-S, MLR-BS and ANN had a R2 0.35, 0.58 and 0.86, respectively. For the model transferability, the three PTFs applied to the external topsoil dataset (N = 59), achieved R2 values of 0.06, 0.03 and 0.41. For the subsoil training dataset (N = 180), MLR-S, MLR-BS and ANN the R2 values were 0.36, 0.46 and 0.83, respectively. When applied to the external subsoil dataset (N = 29), the R2 values were 0.05, 0.06 and 0.41. The cross-validation for both top and subsoil dataset, resulted in an intermediate performance compared to calibration and validation with the external dataset. The new ANN PTF outperformed MLR-S, MLR-BS, MJ and SoilGrids approaches for estimating BD. Further improvements may be achieved by additionally considering the time of sampling, agricultural soil management and cultivation practices in predictive models.

Variational-LSTM autoencoder to forecast the spread of coronavirus across the globe.

Modelling the spread of coronavirus globally while learning trends at global and country levels remains crucial for tackling the pandemic. We introduce a novel variational-LSTM Autoencoder model to predict the spread of coronavirus for each country across the globe. This deep Spatio-temporal model does not only rely on historical data of the virus spread but also includes factors related to urban characteristics represented in locational and demographic data (such as population density, urban population, and fertility rate), an index that represents the governmental measures and response amid toward mitigating the outbreak (includes 13 measures such as: 1) school closing, 2) workplace closing, 3) cancelling public events, 4) close public transport, 5) public information campaigns, 6) restrictions on internal movements, 7) international travel controls, 8) fiscal measures, 9) monetary measures, 10) emergency investment in health care, 11) investment in vaccines, 12) virus testing framework, and 13) contact tracing). In addition, the introduced method learns to generate a graph to adjust the spatial dependences among different countries while forecasting the spread. We trained two models for short and long-term forecasts. The first one is trained to output one step in future with three previous timestamps of all features across the globe, whereas the second model is trained to output 10 steps in future. Overall, the trained models show high validation for forecasting the spread for each country for short and long-term forecasts, which makes the introduce method a useful tool to assist decision and policymaking for the different corners of the globe.

A systematic approach to normalization in probabilistic models.

Every information retrieval (IR) model embeds in its scoring function a form of term frequency (TF) quantification. The contribution of the term frequency is determined by the properties of the function of the chosen TF quantification, and by its TF normalization. The first defines how independent the occurrences of multiple terms are, while the second acts on mitigating the a priori probability of having a high term frequency in a document (estimation usually based on the document length). New test collections, coming from different domains (e.g. medical, legal), give evidence that not only document length, but in addition, verboseness of documents should be explicitly considered. Therefore we propose and investigate a systematic combination of document verboseness and length. To theoretically justify the combination, we show the duality between document verboseness and length. In addition, we investigate the duality between verboseness and other components of IR models. We test these new TF normalizations on four suitable test collections. We do this on a well defined spectrum of TF quantifications. Finally, based on the theoretical and experimental observations, we show how the two components of this new normalization, document verboseness and length, interact with each other. Our experiments demonstrate that the new models never underperform existing models, while sometimes introducing statistically significantly better results, at no additional computational cost.

Spatio-temporal topsoil organic carbon mapping of a semi-arid Mediterranean region: The role of land use, soil texture, topographic indices and the influence of remote sensing data to modelling.

SOC is the most important indicator of soil fertility and monitoring its space-time changes is a prerequisite to establish strategies to reduce soil loss and preserve its quality. Here we modelled the topsoil (0-0.3m) SOC concentration of the cultivated area of Sicily in 1993 and 2008. Sicily is an extremely variable region with a high number of ecosystems, soils, and microclimates. We studied the role of time and land use in the modelling of SOC, and assessed the role of remote sensing (RS) covariates in the boosted regression trees modelling. The models obtained showed a high pseudo-R2 (0.63-0.69) and low uncertainty (s.d.<0.76gCkg-1 with RS, and <1.25gCkg-1 without RS). These outputs allowed depicting a time variation of SOC at 1arcsec. SOC estimation strongly depended on the soil texture, land use, rainfall and topographic indices related to erosion and deposition. RS indices captured one fifth of the total variance explained, slightly changed the ranking of variance explained by the non-RS predictors, and reduced the variability of the model replicates. During the study period, SOC decreased in the areas with relatively high initial SOC, and increased in the area with high temperature and low rainfall, dominated by arables. This was likely due to the compulsory application of some Good Agricultural and Environmental practices. These results confirm that the importance of texture and land use in short-term SOC variation is comparable to climate. The present results call for agronomic and policy intervention at the district level to maintain fertility and yield potential. In addition, the present results suggest that the application of RS covariates enhanced the modelling performance.