Machine Learning and Feature Selection for soil spectroscopy. An evaluation of Random Forest wrappers to predict soil organic matter, clay, and carbonates

Soil spectroscopy estimates soil properties using the absorption features in soil spectra. However, modelling soil properties with soil spectroscopy is challenging due to the high dimensionality of spectral data. Feature Selection wrapper methods are promising approaches to reduce the dimensionality but are barely used in soil spectroscopy. The aim of this study is to evaluate the performance of two feature selection wrapper methods, Sequential Forward Selection (SFS) and Sequential Flotant Forward Selection (SFFS) built using the Random Forest (RF) algorithm, for dimensionality reduction of spectral data and predictive modelling of modelling soil organic matter (SOM), clay and carbonates. The reflectance of 100 soil samples, acquired from Sierra de las Nieves (Spain), was measured under laboratory conditions using ASD FieldSpec Pro JR. Four different datasets were obtained after applying two spectral preprocessing methods to raw spectra: raw spectra, Continuum Removal (CR), Multiplicative Scatter Correction (MSC), and a so-called "Global" dataset composed of raw, CR and MSC features. The performance of RF models built with feature selection methods was compared to that of Partial Least Squares Regression (PLSR) and RF (alone). RF models built with SFS and SFFS outperformed PLSR and RF alone models: The best RF models with feature selection had a respective ratio of performance to interquartile distance of 1.93, 0.38 and 2.56. PLSR models had an accuracy of 1.41, 0.29 and 1.81 for SOM, carbonates, and clay, respectively. RF alone had a respective performance of 1.29, 0.29 and 1.81. The application of feature selection wrapper methods reduced the number of features to less than 1 % of the starting features. Features were selected across all spectra for SOM and clay, and around 900 nm, 1900 nm, and 2350 nm for carbonates. However, feature selection highlighted features around 1100 nm in SOM modelling, as well as other features around 2200 nm, which is considered a main absorption feature of clay. The application of feature selection with Random Forest was very important in improving modelling accuracy, reducing the redundant features and avoiding the curse of dimensionality or Hughes effect. Thus, this research showed an alternative to dimensionality reduction approaches that have been applied to date to model soil properties with spectroscopy and paves the way for further scientific investigation based on feature selection methods and machine learning.

Analysis of landslide explicative factors and susceptibility mapping in an andean context: The case of Azuay province (Ecuador)

Landslides are one of the natural phenomena with more negative impacts on landscape, natural resources, and human health worldwide. Andean geomorphology, urbanization, poverty, and inequality make it more vulnerable to landslides. This research focuses on understanding explanatory landslide factors and promoting quantitative susceptibility mapping. Both tasks supply valuable knowledge for the Andean region, focusing on territorial planning and risk management support. This work addresses the following questions using the province of Azuay-Ecuador as a study area: (i) How do EFA and LR assess the significance of landslide occurrence factors? (ii) Which are the most significant landslide occurrence factors for susceptibility analysis in an Andean context? (iii) What is the landslide susceptibility map for the study area? The methodological framework uses quantitative techniques to describe landslide behavior. EFA and LR models are based on a historical inventory of 665 records. Both identified NDVI, NDWI, altitude, fault density, road density, and PC2 as the most significant factors. The latter factor represents the standard deviation, maximum value of precipitation, and rainfall in the wet season (January, February, and March). The EFA model was built from 7 latent factors, which explained 55% of the accumulated variance, with a medium item complexity of 1.5, a RMSR of 0.02, and a TLI of 0.89. This technique also identified TWI, fault distance, plane curvature, and road distance as important factors. LR's model, with AIC of 964.63, residual deviance of 924.63, AUC of 0.92, accuracy of 0.84, and Kappa of 0.68, also shows statistical significance for slope, roads density, geology, and land cover factors. This research encompasses a time-series analysis of NDVI, NDWI, and precipitation, including vegetation and weather dynamism for landslide occurrence. Finally, this methodological framework replaces traditional qualitative models based on expert knowledge, for quantitative approaches for the study area and the Andean region.

Evaluating the BFAST method to detect and characterise changing trends in water time series: A case study on the impact of droughts on the Mediterranean climate

Mediterranean climate regions are facing increased aridity conditions and water scarcity, thus needing integrated management of water resources. Detecting and characterising changes in water resources over time is the natural first step towards identifying the drivers of these changes and understanding the mechanism of change. The aim of this study is to evaluate the potential of Breaks For Additive Seasonal and Trend (BFAST) method to identify gradual (trend) and abrupt (step- change) changes in the freshwater resources time series over a long-term period. This research shows an alternative to the Pettitt's test, LOESS (locally estimated scatterplot smoothing) filter, Mann-Kendall trend test among other common methods for change detection in hydrological data, and paves the way for further scientific investigation related to climate variability and its influence on water resources. We used the monthly accumulated stored water in three reservoirs, the monthly groundwater levels of three hydrological settings and a standardized precipitation index to show BFAST performance. BFAST was successfully applied, enabling: (1) assessment of the suitability of past management decisions when tackling drought events; (2) detection of recovery and drawdown periods (duration and magnitude values) of accumulated stored water in reservoirs and groundwater bodies after wet and dry periods; 3) measurement of resilience to drought conditions; (4) establishment of similarities/differences in trends between different reservoirs and groundwater bodies with regard to drought events.

Photoperiod controls vegetation phenology across Africa

Vegetation phenology is driven by environmental factors such as photoperiod, precipitation, temperature, insolation, and nutrient availability. However, across Africa, there's ambiguity about these drivers, which can lead to uncertainty in the predictions of global warming impacts on terrestrial ecosystems and their representation in dynamic vegetation models. Using satellite data, we undertook a systematic analysis of the relationship between phenological parameters and these drivers. The analysis across different regions consistently revealed photoperiod as the dominant factor controlling the onset and end of vegetation growing season. Moreover, the results suggest that not one, but a combination of drivers control phenological events. Consequently, to enhance our predictions of climate change impacts, the role of photoperiod should be incorporated into vegetation-climate and ecosystem modelling. Furthermore, it is necessary to define clearly the responses of vegetation to interactions between a consistent photoperiod cue and inter-annual variation in other drivers, especially under a changing climate.

Spatiotemporal analysis of the housing bubble's contribution to the proliferation of illegal landfills - The case of Gran Canaria

Illegal landfills are the source of many impacts that can alter the environment and represent a public health risk. This study investigates their spatiotemporal distribution in two representative areas of Gran Canaria: northwest (Zone A) and east (Zone B). Illegal landfill occurrence was simulated between 2000 and 2018, to estimate and spatially locate the surface growth of illegal landfills based on cellular automata, cellular automata-Markov and multiobjective land allocation models. The proliferation of illegal landfills in 2018 was simulated following the calibration and validation of the proposed models. Models' accuracy was assessed using Kappa index and landscape metrics. The cellular automata-Markov model had the best performance. The model simulations predicted an increase of 52.3 ha and 81.5 ha affected by illegal landfills in Zone A and Zone B for 2018, respectively. The interannual growth rate of surfaces affected by illegal landfills for the period between 2000 and 2006 was 4.5% and 9.5% and between 2006 and 2012 it was 6.6% and 6.7%, for Zone A and Zone B respectively. The growth of illegal landfills between 2000 and 2006 was higher in urban areas, construction sites, and industrial zones, and may be closely related to the process of urban expansion linked to the real estate boom. The latter would have a deep impact on the landscape due to the proliferation of illegal construction and demolition waste. The growth rate of illegal landfills in urban environments fell during the later period of urban expansion. Overall, simulation outputs showed the model's ability to correctly reproduce the distribution patterns for illegal landfill proliferation. Even though the simulated spatial location of illegal landfills was not highly accurate, the models built in this study provide an informative tool to policy makers to aid the process creating policies for environmental protection as well as territorial planning.

Characterization and mapping of illegal landfill potential occurrence in the Canary Islands

The proliferation of illegal landfills (IL) has a negative impact, especially for ecologically sensitive areas or those attractive for tourists. This research focuses on the drivers of the IL spatial distribution in archipelagic environments for mapping the IL potential occurrence. 286 and 153 illegal landfills localizations were identified through fieldwork in the islands of Gran Canaria (GC) and La Palma (LP), respectively. The characterization of IL was carried out from a set of features (177) such as: waste type, control and surveillance, socioeconomic, accessibility, distance to elements of interest, visibility and physiographic. Feature selection was performed using the Discriminant Analysis technique (DA). The DA model selected 10 and 9 features for GC and LP, respectively. The GC IL potential occurrence was mainly related to the greenhouse density, type of cadastral plot and distance to the coast. For the case of LP, the following features were selected: population density, distance to natural protected areas, distance to urban areas, slope and Normalised Difference Vegetation Index (NDVI). Different potential illegal landfill occurrence maps were obtained: (i) likelihood of occurrence of IL; and (ii) areas potentially affected by IL, based on the application of ROC (Receiver Operating Characteristics) curves and success rate. ROC was equal to 0.973 and 0.979 in LP and GC, respectively. Success rate was equal to 81.58% considering an affected area of 21.95% in LP, whereas success rate was equal to 87.32% in GC considering 20.10% affected area.

Identifying the main physical and socioeconomic drivers of illegal landfills in the Canary Islands

The management of disposed waste in illegal landfills (ILs) is a significant problem in contemporary societies due to respective hazards for the environment and human health. This paper presents a characterisation of ILs on the islands of La Palma (LP) and Gran Canaria (GC) based on multivariable statistical analysis. Inspection of numerous sites on both islands revealed a total of 153 and 286 ILs on LP and GC, respectively. A geospatial database was created composed of different potentially explanatory features of different typology (177): waste type, control and vigilance, socioeconomic, accessibility, distance to elements of interest, visibility and physical. The degree of association between the explanatory features and the occurrence of ILs was analysed with the support of exploratory statistics and the multivariable analysis techniques of principal component analysis (PCA) and binary logistic regression (LR). PCA explained 82.34% and 81.83% of total data variance in LP and GC, respectively, considering 7 and 6 components (Kaiser-Mayer-Olkin; LP: 0.715; GC: 0.711). The LR models for LP and GC had an overall accuracy of 93.5% and 92.5%. In LP and GC, 6 of 23 features and 9 of 21 features were, respectively, selected. The features most associated with the occurrence of ILs were: in LP, building density, distance to agricultural spaces and distance to green zones; in GC, the industrial activity indicator, density of ground use transition to artificial covers, density of greenhouses and distance to communication routes.

A methodology for assessing public health risk associated with groundwater nitrate contamination: a case study in an agricultural setting (southern Spain)

Groundwater nitrate contamination from agriculture is of paramount environmental interest. A continuous consumption of polluted water as drinking water or for culinary purposes is by no means a minor hazard for people's health that must be studied. This research presents a new methodology for the spatial analysis of health risk rate from intake of nitrate-polluted groundwater. The method is illustrated through its application to a water quality sampling campaign performed in the south of Spain in 2003. The probability risk model used by the US Environmental Protection Agency has been applied, considering a residential intake framework and three representative population age groups (10, 40 and 65 years).The method was based upon coupling Monte Carlo simulations and geostatistics, which allowed mapping of the health risk coefficient (RC). The maps obtained were interpreted in the framework of water resources management and user's health protection (municipalities). The results showed waterborne health risk caused by nitrate-polluted water is moderately low for the region. The observed risk was larger for the elderly and children, although no significant differences were found among the three age groups (RC average values of 95th percentile for age of 0.37, 0.33 and 0.37, respectively). Significant risk values of RC > 1 were obtained for 10 % of the surface in the NW site of the study area, where the municipalities with the highest contamination thresholds are located (agricultural activity). Nitrate concentration and intake rate stood out as the main explanatory variables of the RC.

Predictive modeling of groundwater nitrate pollution using Random Forest and multisource variables related to intrinsic and specific vulnerability: a case study in an agricultural setting (Southern Spain)

Watershed management decisions need robust methods, which allow an accurate predictive modeling of pollutant occurrences. Random Forest (RF) is a powerful machine learning data driven method that is rarely used in water resources studies, and thus has not been evaluated thoroughly in this field, when compared to more conventional pattern recognition techniques key advantages of RF include: its non-parametric nature; high predictive accuracy; and capability to determine variable importance. This last characteristic can be used to better understand the individual role and the combined effect of explanatory variables in both protecting and exposing groundwater from and to a pollutant. In this paper, the performance of the RF regression for predictive modeling of nitrate pollution is explored, based on intrinsic and specific vulnerability assessment of the Vega de Granada aquifer. The applicability of this new machine learning technique is demonstrated in an agriculture-dominated area where nitrate concentrations in groundwater can exceed the trigger value of 50 mg/L, at many locations. A comprehensive GIS database of twenty-four parameters related to intrinsic hydrogeologic proprieties, driving forces, remotely sensed variables and physical-chemical variables measured in "situ", were used as inputs to build different predictive models of nitrate pollution. RF measures of importance were also used to define the most significant predictors of nitrate pollution in groundwater, allowing the establishment of the pollution sources (pressures). The potential of RF for generating a vulnerability map to nitrate pollution is assessed considering multiple criteria related to variations in the algorithm parameters and the accuracy of the maps. The performance of the RF is also evaluated in comparison to the logistic regression (LR) method using different efficiency measures to ensure their generalization ability. Prediction results show the ability of RF to build accurate models with strong predictive capabilities.

Categorical Indicator Kriging for assessing the risk of groundwater nitrate pollution: the case of Vega de Granada aquifer (SE Spain)

Groundwater nitrate pollution associated with agricultural activity is an important environmental problem in the management of this natural resource, as acknowledged by the European Water Framework Directive. Therefore, specific measures aimed to control the risk of water pollution by nitrates must be implemented to minimise its impact on the environment and potential risk to human health. The spatial probability distribution of nitrate contents exceeding a threshold or limit value, established within the quality standard, will be helpful to managers and decision-makers. A methodology based on non-parametric and non-linear methods of Indicator Kriging was used in the elaboration of a nitrate pollution categorical map for the aquifer of Vega de Granada (SE Spain). The map has been obtained from the local estimation of the probability that a nitrate content in an unsampled location belongs to one of the three categories established by the European Water Framework Directive: CL. 1 good quality [Min - 37.5 ppm], CL. 2 intermediate quality [37.5-50 ppm] and CL. 3 poor quality [50 ppm - Max]. The obtained results show that the areas exceeding nitrate concentrations of 50 ppm, poor quality waters, occupy more than 50% of the aquifer area. A great proportion of the area's municipalities are located in these poor quality water areas. The intermediate quality and good quality areas correspond to 21% and 28%, respectively, but with the highest population density. These results are coherent with the experimental data, which show an average nitrate concentration value of 72 ppm, significantly higher than the quality standard limit of 50 ppm. Consequently, the results suggest the importance of planning actions in order to control and monitor aquifer nitrate pollution.