Intrinsic and extrinsic techniques for quantification uncertainty of an interpretable GRU deep learning model used to predict atmospheric total suspended particulates (TSP) in Zabol, Iran during the dusty period of 120-days wind

Total suspended particulates (TSP), as a key pollutant, is a serious threat for air quality, climate, ecosystems and human health. Therefore, measurements, prediction and forecasting of TSP concentrations are necessary to mitigate their negative effects. This study applies the gated recurrent unit (GRU) deep learning model to predict TSP concentrations in Zabol, Iran, during the dust period of the 120-day wind (3 June - 4 October 2014). Three uncertainty quantification (UQ) techniques consisting of the blackbox metamodel, heteroscedastic regression and infinitesimal jackknife were applied to quantify the uncertainty associated with GRU model. Permutation feature importance measure (PFIM), based on the game theory, was employed for the interpretability of the predictive model's outputs. A total of 80 TSP samples were collected and were randomly divided as training (70%) and validation (30%) datasets, while eight variables were used in the TSP prediction model. Our findings showed that GRU performed very well for TSP prediction (with r and Nash Sutcliffe coefficient (NSC) values above 0.99 for both datasets, and RMSE of 57 μg m-3 and 73 μg m-3 for training and validation datasets, respectively). Among the three UQ techniques, the infinitesimal jackknife was the most accurate one, while all the observed and predicted TSP values fell within the continence limitation estimated by the model. PFIM plots showed that wind speed and air humidity were the most and least important variables, respectively, impacting the predictive model's outputs. This is the first attempt of using an interpretable DL model for TSP prediction modelling, recommending that future research should involve aspects of uncertainty and interpretability of the predictive models. Overall, UQ and interpretability techniques have a key role in reducing the impact of uncertainties during optimization and decision making, resulting in better understanding of sophisticated mechanisms related to the predictive model.

Interpretability of simple RNN and GRU deep learning models used to map land susceptibility to gully erosion

Gully erosion possess a serious hazard to critical resources such as soil, water, and vegetation cover within watersheds. Therefore, spatial maps of gully erosion hazards can be instrumental in mitigating its negative consequences. Among the various methods used to explore and map gully erosion, advanced learning techniques, especially deep learning (DL) models, are highly capable of spatial mapping and can provide accurate predictions for generating spatial maps of gully erosion at different scales (e.g., local, regional, continental, and global). In this paper, we applied two DL models, namely a simple recurrent neural network (RNN) and a gated recurrent unit (GRU), to map land susceptibility to gully erosion in the Shamil-Minab plain, Hormozgan province, southern Iran. To address the inherent black box nature of DL models, we applied three novel interpretability methods consisting of SHaply Additive explanation (SHAP), ceteris paribus and partial dependence (CP-PD) profiles and permutation feature importance (PFI). Using the Boruta algorithm, we identified seven important features that control gully erosion: soil bulk density, clay content, elevation, land use type, vegetation cover, sand content, and silt content. These features, along with an inventory map of gully erosion (based on a 70 % training dataset and 30 % test dataset), were used to generate spatial maps of gully erosion using DL models. According to the Kolmogorov-Smirnov (KS) statistic performance assessment measure, the simple RNN model (with KS = 91.6) outperformed the GRU model (with KS = 66.6). Based on the results from the simple RNN model, 7.4 %, 14.5 %, 18.9 %, 31.2 % and 28 % of total area of the plain were classified as very-low, low, moderate, high and very-high hazard classes, respectively. According to SHAP plots, CP-PD profiles, and PFI measures, soil silt content, vegetation cover (NDVI) and land use type had the highest impact on the model's output. Overall, the DL modelling techniques and interpretation methods used in this study proved to be helpful in generating spatial maps of soil erosion hazard, especially gully erosion. Their interpretability can support watershed sustainable management.

Novel integrated modelling based on multiplicative long short-term memory (mLSTM) deep learning model and ensemble multi-criteria decision making (MCDM) models for mapping flood risk

Flood risk assessment is a key step in flood management and mitigation, and flood risk maps provide a quantitative measure of flood risk. Therefore, integration of deep learning - an updated version of machine learning techniques - and multi-criteria decision making (MCDM) models can generate high-resolution flood risk maps. In this study, a novel integrated approach has been developed based on multiplicative long short-term memory (mLSTM) deep learning models and an MCDM ensemble model to map flood risk in the Minab-Shamil plain, southern Iran. A flood hazard map generated by the mLSTM model is based on nine critical features selected by GrootCV (distance to the river, vegetation cover, variables extracted from DEM (digital elevation model) and river density) and a flood inventory map (70% and 30% data were randomly selected as training and test datasets, respectively). The values of all criteria used to assess model accuracy performance (except Cohens kappa for train dataset = 86, and for test dataset = 84) achieved values greater than 90, which indicates that the mLSTM model performed very well for the generation of a spatial flood hazard map. According to the spatial flood hazard map produced by mLSTM, the very low, low, moderate, high and very high classes cover 26%, 35.3%, 20.5%, 11.2% and 7% of the total area, respectively. Flood vulnerability maps were produced by the combinative distance-based assessment (CODAS), the evaluation based on distance from average solution (EDAS), and the multi-objective optimization on the basis of simple ratio analysis (MOOSRA), and then validated by Spearman's rank correlation coefficients (SRC). Based on the SRC, the three models CODAS, EDAS, and MOOSRA showed high-ranking correlations with each other, and all three models were then used in the ensemble process. According to the CODAS-EDAS-MOOSRA ensemble model, 21.5%, 34.2%, 23.7%, 13%, and 7.6% of the total area were classified as having a very low to very high flood vulnerability, respectively. Finally, a flood risk map was generated by the combination of flood hazard and vulnerability maps produced by the mLSTM and MCDM ensemble model. According to the flood risk map, 27.4%, 34.3%, 14.8%, 15.7%, and 7.8% of the total area were classified as having a very low, low, moderate, high, and very high flood risk, respectively. Overall, the integration of mLSTM and the MCDM ensemble is a promising tool for generating precise flood risk maps and provides a useful reference for flood risk management.

Stacking- and voting-based ensemble deep learning models (SEDL and VEDL) and active learning (AL) for mapping land subsidence

This contribution presents a novel methodology based on the feature selection, ensemble deep learning (EDL) models, and active learning (AL) approach for prediction of land subsidence (LS) hazard and rate, and its uncertainty in an area involving two important plains - the Minab and Shamil-Nian plains - in the Hormozgan province, southern Iran. The important features controlling LS hazard were identified by ridge regression. Then, two EDL models were constructed by stacking (SEDL) and voting (VEDL) five dense deep learning (DL) models (model 1 to model 5) for mapping LS hazard. Thereafter, the predictive model performance was assessed by a precision-recall curve and Kolmogorov-Smirnov (KS) plot. A partial dependence plot (PDP), individual conditional expectation plots (ICEP), game theory, and a sensitivity analysis were used for the interpretability of the predictive DL model. According to SEDL - a model with higher accuracy - 34% (1624 km²), 14.7% (698 km²), and 19.2% (912 km²) of the total area were classified as being of very low, low, and moderate hazards, whereas 17.7% (845 km²) and 14.4% (683 km²) of area were classified as being of high and very high hazards, respectively. Based on all interpretability techniques, aquifer loss or groundwater drawdown is the most important feature controlling LS hazard, and it having the greatest impact on the SEDL model output. Based on a Taylor diagram and R² as model performance assessment indicators, SEDL-AL (with R² > 95% for training and test datasets) performed better than SEDL for quantify LS rate, the rate of LS ranging between 0 and 48.1 cm. The highest rate of LS occurred in the Minab plain - an area located downstream of the Minab Esteghlal dam. SEDL-AL was used to quantify the uncertainty associated with the LS rate. The observed values fell within predictions provided by SEDL-AL, which indicates a high accuracy of our predictive model. Overall, our newly developed modeling techniques are helpful tools for the spatial mapping of LS susceptibility and rate, and its uncertainty.

Novel deep learning hybrid models (CNN-GRU and DLDL-RF) for the susceptibility classification of dust sources in the Middle East: a global source

Dust storms have many negative consequences, and affect all kinds of ecosystems, as well as climate and weather conditions. Therefore, classification of dust storm sources into different susceptibility categories can help us mitigate its negative effects. This study aimed to classify the susceptibility of dust sources in the Middle East (ME) by developing two novel deep learning (DL) hybrid models based on the convolutional neural network-gated recurrent unit (CNN-GRU) model, and the dense layer deep learning-random forest (DLDL-RF) model. The Dragonfly algorithm (DA) was used to identify the critical features controlling dust sources. Game theory was used for the interpretability of the DL model's output. Predictive DL models were constructed by dividing datasets randomly into train (70%) and test (30%) groups, six statistical indicators being then applied to assess the DL hybrid model performance for both datasets (train and test). Among 13 potential features (or variables) controlling dust sources, seven variables were selected as important and six as non-important by DA, respectively. Based on the DLDL-RF hybrid model - a model with higher accuracy in comparison with CNN-GRU-23.1, 22.8, and 22.2% of the study area were classified as being of very low, low and moderate susceptibility, whereas 20.2 and 11.7% of the area were classified as representing high and very high susceptibility classes, respectively. Among seven important features selected by DA, clay content, silt content, and precipitation were identified as the three most important by game theory through permutation values. Overall, DL hybrid models were found to be efficient methods for prediction purposes on large spatial scales with no or incomplete datasets from ground-based measurements.

Spatial modelling of soil salinity: deep or shallow learning models?

Understanding the spatial distribution of soil salinity is required to conserve land against degradation and desertification. Against this background, this study is the first attempt to predict soil salinity in the Jaghin basin, in southern Iran, by applying and comparing the performance of four deep learning (DL) models (deep convolutional neural networks-DCNNs, dense connected deep neural networks-DenseDNNs, recurrent neural networks-long short-term memory-RNN-LSTM and recurrent neural networks-gated recurrent unit-RNN-GRU) and six shallow machine learning (ML) models (bagged classification and regression tree-BCART, cforest, cubist, quantile regression with LASSO penalty-QR-LASSO, ridge regression-RR and support vectore machine-SVM). To do this, 49 environmental landsat8-derived variables including digital elevation model (DEM)-extracted covariates, soil-salinity indices, and other variables (e.g., soil order, lithology, land use) were mapped spatially. For assessing the relationships between soil salinity (EC) and factors controlling EC, we collected 319 surficial (0-5 cm depth) soil samples for measuring soil salinity on the basis of electrical conductivity (EC). We then selected the most important features (covariates) controlling soil salinity by applying a MARS model. The performance of the DL and shallow ML models for generating soil salinity spatial maps (SSSMs) was assessed using a Taylor diagram and the Nash Sutcliff coefficient (NSE). Among all 10 predictive models, DL models with NSE ≥ 0.9 (DCNNs was the most accurate model with NSE = 0.96) were selected as the four best models, and performed better than the six shallow ML models with NSE ≤ 0.83 (QR-LASSO was the weakest predictive model with NSE = 0.50). Based on DCNNs-, the values of the EC ranged between 0.67 and 14.73 dS/m, whereas for QR-LASSO the corresponding EC values were 0.37 to 19.6 dS/m. Overall, DL models performed better than shallow ML models for production of the SSSMs and therefore we recommend applying DL models for prediction purposes in environmental sciences.

Mapping wind erosion hazard with regression-based machine learning algorithms

Land susceptibility to wind erosion hazard in Isfahan province, Iran, was mapped by testing 16 advanced regression-based machine learning methods: Robust linear regression (RLR), Cforest, Non-convex penalized quantile regression (NCPQR), Neural network with feature extraction (NNFE), Monotone multi-layer perception neural network (MMLPNN), Ridge regression (RR), Boosting generalized linear model (BGLM), Negative binomial generalized linear model (NBGLM), Boosting generalized additive model (BGAM), Spline generalized additive model (SGAM), Spike and slab regression (SSR), Stochastic gradient boosting (SGB), support vector machine (SVM), Relevance vector machine (RVM) and the Cubist and Adaptive network-based fuzzy inference system (ANFIS). Thirteen factors controlling wind erosion were mapped, and multicollinearity among these factors was quantified using the tolerance coefficient (TC) and variance inflation factor (VIF). Model performance was assessed by RMSE, MAE, MBE, and a Taylor diagram using both training and validation datasets. The result showed that five models (MMLPNN, SGAM, Cforest, BGAM and SGB) are capable of delivering a high prediction accuracy for land susceptibility to wind erosion hazard. DEM, precipitation, and vegetation (NDVI) are the most critical factors controlling wind erosion in the study area. Overall, regression-based machine learning models are efficient techniques for mapping land susceptibility to wind erosion hazards.

A new integrated data mining model to map spatial variation in the susceptibility of land to act as a source of aeolian dust

This research developed a more efficient integrated model (IM) based on combining the Nash-Sutcliffe efficiency coefficient (NSEC) and individual data mining (DM) algorithms for the spatial mapping of dust provenance in the Hamoun-e-Hirmand Basin, southeastern Iran. This region experiences severe wind erosion and includes the Sistan plain which is one of the most PM_2.5-polluted regions in the world. Due to a prolonged drought over the last two decades, the frequency of dust storms in the study area is increasing remarkably. Herein, 14 factors controlling dust emissions (FCDEs) including soil characteristics, climatic variables, digital elevation map, normalized difference vegetation index, land use and geology were mapped. Correlation and collinearity among the FCDEs were examined by the Pearson test, tolerance coefficient (TC) and variance inflation factor (VIF), with the results suggesting a lack of collinearity between FCDEs. A tree-based genetic algorithm was applied to prioritize and quantify the importance weights of the FCDEs. Thirteen individual data mining models were applied for mapping dust provenance. The model performance was assessed using root mean square error, mean absolute error and NSEC. Based on clustering analysis, the 13 DM models were grouped into five clusters and then the cluster with the highest NSEC values used in an integrated modelling process. Based on the results, the IM (NSEC = 93%) outperformed the individual DM models (the NSEC values range between 51 and 92%). Using the IM, 11, 5, 7 and 77% of the total study area were classified into low, moderate, high and very high susceptibility classes for dust provenance, respectively. Overall, the results illustrate the benefits of an IM for mapping spatial variation in the susceptibility of catchment areas to act as dust sources.