Using an interpretable deep learning model for the prediction of riverine suspended sediment load.

The prediction of suspended sediment load (SSL) within riverine systems is critical to understanding the watershed's hydrology. Therefore, the novelty of our research is developing an interpretable (explainable) model based on deep learning (DL) and Shapley Additive ExPlanations (SHAP) interpretation technique for prediction of SSL in the riverine systems. This paper investigates the abilities of four DL models, including dense deep neural networks (DDNN), long short-term memory (LSTM), gated recurrent unit (GRU), and simple recurrent neural network (RNN) models for the prediction of daily SSL using river discharge and rainfall data at a daily time scale in the Taleghan River watershed, northwestern Tehran, Iran. The performance of models was evaluated by using several quantitative and graphical criteria. The effect of parameter settings on the performance of deep models on SSL prediction was also investigated. The optimal optimization algorithms, maximum iteration (MI), and batch size (BC) were obtained for modeling daily SSL, and structure of the model impact on prediction remarkably. The comparison of prediction accuracy of the models illustrated that DDNN (with R2 = 0.96, RMSE = 333.46) outperformed LSTM (R2 = 0.75, RMSE = 786.20), GRU (R2 = 0.73, RMSE = 825.67), and simple RNN (R2 = 0.78, RMSE = 741.45). Furthermore, the Taylor diagram confirmed that DDNN has the highest performance among other models. Interpretation techniques can address the black-box nature of models, and here, SHAP was applied to develop an interpretable DL model to interpret of DL model's output. The results of SHAP showed that river discharge has the strongest impact on the model's output in estimating SSL. Overall, we conclude that DL models have great potential in watersheds to predict SSL. Therefore, different interpretation techniques as tools to interpret DL model's output (DL model is as black-box model) are recommended in future research.

Mapping groundwater potential zone in the subarnarekha basin, India, using a novel hybrid multi-criteria approach in Google earth Engine.

Assessing groundwater potential for sustainable resource management is critically important. In addressing this concern, this study aims to advance the field by developing an innovative approach for Groundwater potential zone (GWPZ) mapping using advanced techniques, such as FuzzyAHP, FuzzyDEMATEL, and Logistic regression (LR) models. GWPZ was carried out by integrating various primary factors, such as hydrologic, soil permeability, morphometric, terrain distribution, and anthropogenic influences, incorporating twenty-seven individual criteria using multi-criteria decision models along with a hybrid approach for the Subarnarekha River basin, India, in Google earth engine (GEE). The predictive capability of the model was evaluated using a Multi-Collinearity test (VIF <10.0), followed by applying a random forest model, considering the weighted impact of the five primary factors. The hybrid model for GWPZ classification showed that 21.97 % (4256.3 km2) of the area exhibited very high potential, while 11.37 % (2202.1 km2) indicated very low potential for GW in this area. Validation of the groundwater level data from 72 observation wells, performed by the Area under receiver operating characteristic (AUROC) curve technique, yielded values ranging between 75 % and 78 % for different models, underscoring the robust predictability of GWPZ. The hybrid and LR-FuzzyAHP models demonstrated remarkable effectiveness in GWPZ mapping, indicating that the downstream and southern regions boast substantial groundwater potential attributed to alluvial soil and favorable recharge conditions. Conversely, the central part grapples with a scarcity of groundwater. It holds the potential to assist planners and managers in formulating strategies for managing groundwater levels and alleviating the impacts of future droughts.

HDTO-DeepAR: A novel hybrid approach to forecast surface water quality indicators.

Water is a vital resource supporting a broad spectrum of ecosystems and human activities. The quality of river water has declined in recent years due to the discharge of hazardous materials and toxins. Deep learning and machine learning have gained significant attention for analysing time-series data. However, these methods often suffer from high complexity and significant forecasting errors, primarily due to non-linear datasets and hyperparameter settings. To address these challenges, we have developed an innovative HDTO-DeepAR approach for predicting water quality indicators. This proposed approach is compared with standalone algorithms, including DeepAR, BiLSTM, GRU and XGBoost, using performance metrics such as MAE, MSE, MAPE, and NSE. The NSE of the hybrid approach ranges between 0.8 to 0.96. Given the value's proximity to 1, the model appears to be efficient. The PICP values (ranging from 95% to 98%) indicate that the model is highly reliable in forecasting water quality indicators. Experimental results reveal a close resemblance between the model's predictions and actual values, providing valuable insights for predicting future trends. The comparative study shows that the suggested model surpasses all existing, well-known models.

Enhancing subsurface contamination assessment via ensemble prediction of ground electrical property: A Colorado AMD-impacted wetland case study.

Acid mine drainage (AMD) is recognized as a major environmental challenge in the Western United States, particularly in Colorado, leading to extreme subsurface contamination issue. Given Colorado's arid climate and dependence on groundwater, an accurate assessment of AMD-induced contamination is deemed crucial. While in past, machine learning (ML)-based inversion algorithms were used to reconstruct ground electrical properties (GEP) such as relative dielectric permittivity (RDP) from ground penetrating radar (GPR) data for contamination assessment, their inherent non-linear nature can introduce significant uncertainty and non-uniqueness into the reconstructed models. This is a challenge that traditional ML methods are not explicitly designed to address. In this study, a probabilistic hybrid technique has been introduced that combines the DeepLabv3+ architecture-based deep convolutional neural network (DCNN) with an ensemble prediction-based Monte Carlo (MC) dropout method. Different MC dropout rates (1%, 5%, and 10%) were initially evaluated using 1D and 2D synthetic GPR data for accurate and reliable RDP model prediction. The optimal rate was chosen based on minimal prediction uncertainty and the closest alignment of the mean or median model with the true RDP model. Notably, with the optimal MC dropout rate, prediction accuracy of over 95% for the 1D and 2D cases was achieved. Motivated by these results, the hybrid technique was applied to field GPR data collected over an AMD-impacted wetland near Silverton, Colorado. The field results underscored the hybrid technique's ability to predict an accurate subsurface RDP distribution for estimating the spatial extent of AMD-induced contamination. Notably, this technique not only provides a precise assessment of subsurface contamination but also ensures consistent interpretations of subsurface condition by different environmentalists examining the same GPR data. In conclusion, the hybrid technique presents a promising avenue for future environmental studies in regions affected by AMD or other contaminants that alter the natural distribution of GEP.

Extraction of Roads Using the Archimedes Tuning Process with the Quantum Dilated Convolutional Neural Network.

Road network extraction is a significant challenge in remote sensing (RS). Automated techniques for interpreting RS imagery offer a cost-effective solution for obtaining road network data quickly, surpassing traditional visual interpretation methods. However, the diverse characteristics of road networks, such as varying lengths, widths, materials, and geometries across different regions, pose a formidable obstacle for road extraction from RS imagery. The issue of road extraction can be defined as a task that involves capturing contextual and complex elements while also preserving boundary information and producing high-resolution road segmentation maps for RS data. The objective of the proposed Archimedes tuning process quantum dilated convolutional neural network for road Extraction (ATP-QDCNNRE) technology is to tackle the aforementioned issues by enhancing the efficacy of image segmentation outcomes that exploit remote sensing imagery, coupled with Archimedes optimization algorithm methods (AOA). The findings of this study demonstrate the enhanced road-extraction capabilities achieved by the ATP-QDCNNRE method when used with remote sensing imagery. The ATP-QDCNNRE method employs DL and a hyperparameter tuning process to generate high-resolution road segmentation maps. The basis of this approach lies in the QDCNN model, which incorporates quantum computing (QC) concepts and dilated convolutions to enhance the network's ability to capture both local and global contextual information. Dilated convolutions also enhance the receptive field while maintaining spatial resolution, allowing fine road features to be extracted. ATP-based hyperparameter modifications improve QDCNNRE road extraction. To evaluate the effectiveness of the ATP-QDCNNRE system, benchmark databases are used to assess its simulation results. The experimental results show that ATP-QDCNNRE performed with an intersection over union (IoU) of 75.28%, mean intersection over union (MIoU) of 95.19%, F1 of 90.85%, precision of 87.54%, and recall of 94.41% in the Massachusetts road dataset. These findings demonstrate the superior efficiency of this technique compared to more recent methods.

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.

Development of Debiasing Technique for Lung Nodule Chest X-ray Datasets to Generalize Deep Learning Models.

Screening programs for early lung cancer diagnosis are uncommon, primarily due to the challenge of reaching at-risk patients located in rural areas far from medical facilities. To overcome this obstacle, a comprehensive approach is needed that combines mobility, low cost, speed, accuracy, and privacy. One potential solution lies in combining the chest X-ray imaging mode with federated deep learning, ensuring that no single data source can bias the model adversely. This study presents a pre-processing pipeline designed to debias chest X-ray images, thereby enhancing internal classification and external generalization. The pipeline employs a pruning mechanism to train a deep learning model for nodule detection, utilizing the most informative images from a publicly available lung nodule X-ray dataset. Histogram equalization is used to remove systematic differences in image brightness and contrast. Model training is then performed using combinations of lung field segmentation, close cropping, and rib/bone suppression. The resulting deep learning models, generated through this pre-processing pipeline, demonstrate successful generalization on an independent lung nodule dataset. By eliminating confounding variables in chest X-ray images and suppressing signal noise from the bone structures, the proposed deep learning lung nodule detection algorithm achieves an external generalization accuracy of 89%. This approach paves the way for the development of a low-cost and accessible deep learning-based clinical system for lung cancer screening.

Retraction Note: Weather indicators and improving air quality in association with COVID-19 pandemic in India.

[This retracts the article DOI: 10.1007/s00500-021-06012-9.].

Explainable artificial intelligence (XAI) for interpreting the contributing factors feed into the wildfire susceptibility prediction model.

One of the worst environmental catastrophes that endanger the Australian community is wildfire. To lessen potential fire threats, it is helpful to recognize fire occurrence patterns and identify fire susceptibility in wildfire-prone regions. The use of machine learning (ML) algorithms is acknowledged as one of the most well-known methods for addressing non-linear issues like wildfire hazards. It has always been difficult to analyze these multivariate environmental disasters because modeling can be influenced by a variety of sources of uncertainty, including the quantity and quality of training procedures and input variables. Moreover, although ML techniques show promise in this field, they are unstable for a number of reasons, including the usage of irrelevant descriptor characteristics when developing the models. Explainable AI (XAI) can assist us in acquiring insights into these constraints and, consequently, modifying the modeling approach and training data necessary. In this research, we describe how a Shapley additive explanations (SHAP) model can be utilized to interpret the results of a deep learning (DL) model that is developed for wildfire susceptibility prediction. Different contributing factors such as topographical, landcover/vegetation, and meteorological factors are fed into the model and various SHAP plots are used to identify which parameters are impacting the prediction model, their relative importance, and the reasoning behind specific decisions. The findings drawn from SHAP plots show the significant contributions made by factors such as humidity, wind speed, rainfall, elevation, slope, and normalized difference moisture index (NDMI) to the suggested model's output for wildfire susceptibility mapping. We infer that developing an explainable model would aid in comprehending the model's decision to map wildfire susceptibility, pinpoint high-contributing components in the prediction model, and consequently control fire hazards effectively.

Weather indicators and improving air quality in association with COVID-19 pandemic in India.

The COVID-19 pandemic enforced nationwide lockdown, which has restricted human activities from March 24 to May 3, 2020, resulted in an improved air quality across India. The present research investigates the connection between COVID-19 pandemic-imposed lockdown and its relation to the present air quality in India; besides, relationship between climate variables and daily new affected cases of Coronavirus and mortality in India during the this period has also been examined. The selected seven air quality pollutant parameters (PM10, PM2.5, CO, NO2, SO2, NH3, and O3) at 223 monitoring stations and temperature recorded in New Delhi were used to investigate the spatial pattern of air quality throughout the lockdown. The results showed that the air quality has improved across the country and average temperature and maximum temperature were connected to the outbreak of the COVID-19 pandemic. This outcomes indicates that there is no such relation between climatic parameters and outbreak and its associated mortality. This study will assist the policy maker, researcher, urban planner, and health expert to make suitable strategies against the spreading of COVID-19 in India and abroad. Supplementary Information: The online version contains supplementary material available at 10.1007/s00500-021-06012-9.

Multi-air pollution risk assessment in Southeast Asia region using integrated remote sensing and socio-economic data products.

Air pollution has massive impacts on human life and poor air quality results in three million deaths annually. Air pollution can result from natural causes, including volcanic eruptions and extreme droughts, or human activities, including motor vehicle emissions, industry, and the burning of farmland and forests. Emission sources emit multiple pollutant types with diverse characteristics and impacts. However, there has been little research on the risk of multiple air pollutants; thus, it is difficult to identify multi-pollutant mitigation processes, particularly in Southeast Asia, where air pollution moves dynamically across national borders. In this study, the main objective was to develop a multi-air pollution risk index product for CO, NO2, and SO2 based on Sentinel-5P remote sensing data from 2019 to 2020. The risk index was developed by integrating hazard, vulnerability, and exposure analyses. Hazard analysis considers air pollution data from remote sensing, vulnerability analysis considers the air pollution sources, and exposure analysis considers the population density. The novelty of this study lies in its development of a multi-risk model that considers the weights obtained from the relationship between the hazard and vulnerability parameters. The highest air pollution risk index values were observed in urban areas, with a high exposure index that originates from pollution caused by human activity. Multi-risk analysis of the three air pollutants revealed that Singapore, Vietnam, and the Philippines had the largest percentages of high-risk areas, while Indonesia had the largest total high-risk area (4361 km2). Using the findings of this study, the patterns and characteristics of the risk distribution of multiple air pollutants in Southeast Asia can be identified, which can be used to mitigate multi-pollutant sources, particularly with respect to supporting the clean air targets in the Sustainable Development Goals.

Explainable AI in Scene Understanding for Autonomous Vehicles in Unstructured Traffic Environments on Indian Roads Using the Inception U-Net Model with Grad-CAM Visualization.

The intelligent transportation system, especially autonomous vehicles, has seen a lot of interest among researchers owing to the tremendous work in modern artificial intelligence (AI) techniques, especially deep neural learning. As a result of increased road accidents over the last few decades, significant industries are moving to design and develop autonomous vehicles. Understanding the surrounding environment is essential for understanding the behavior of nearby vehicles to enable the safe navigation of autonomous vehicles in crowded traffic environments. Several datasets are available for autonomous vehicles focusing only on structured driving environments. To develop an intelligent vehicle that drives in real-world traffic environments, which are unstructured by nature, there should be an availability of a dataset for an autonomous vehicle that focuses on unstructured traffic environments. Indian Driving Lite dataset (IDD-Lite), focused on an unstructured driving environment, was released as an online competition in NCPPRIPG 2019. This study proposed an explainable inception-based U-Net model with Grad-CAM visualization for semantic segmentation that combines an inception-based module as an encoder for automatic extraction of features and passes to a decoder for the reconstruction of the segmentation feature map. The black-box nature of deep neural networks failed to build trust within consumers. Grad-CAM is used to interpret the deep-learning-based inception U-Net model to increase consumer trust. The proposed inception U-net with Grad-CAM model achieves 0.622 intersection over union (IoU) on the Indian Driving Dataset (IDD-Lite), outperforming the state-of-the-art (SOTA) deep neural-network-based segmentation models.

Effects of the COVID-19 lockdown and recovery on People's mobility and air quality in the United Arab Emirates using satellite and ground observations.

The stringent COVID-19 lockdown measures in 2020 significantly impacted people's mobility and air quality worldwide. This study presents an assessment of the impacts of the lockdown and the subsequent reopening on air quality and people's mobility in the United Arab Emirates (UAE). Google's community mobility reports and UAE's government lockdown measures were used to assess the changes in the mobility patterns. Time-series and statistical analyses of various air pollutants levels (NO2, O3, SO2, PM10, and aerosol optical depth-AOD) obtained from satellite images and ground monitoring stations were used to assess air quality. The levels of pollutants during the initial lockdown (March to June 2020) and the subsequent gradual reopening in 2020 and 2021 were compared with their average levels during 2015-2019. During the lockdown, people's mobility in the workplace, parks, shops and pharmacies, transit stations, and retail and recreation sectors decreased by about 34%-79%. However, the mobility in the residential sector increased by up to 29%. The satellite-based data indicated significant reductions in NO2 (up to 22%), SO2 (up to 17%), and AOD (up to 40%) with small changes in O3 (up to 5%) during the lockdown. Similarly, data from the ground monitoring stations showed significant reductions in NO2 (49% - 57%) and PM10 (19% - 64%); however, the SO2 and O3 levels showed inconsistent trends. The ground and satellite-based air quality levels were positively correlated for NO2, PM10, and AOD. The data also demonstrated significant correlations between the mobility and NO2 and AOD levels during the lockdown and recovery periods. The study documents the impacts of the lockdown on people's mobility and air quality and provides useful data and analyses for researchers, planners, and policymakers relevant to managing risk, mobility, and air quality.

A new application of deep neural network (LSTM) and RUSLE models in soil erosion prediction.

Rainfall variation causes frequent unexpected disasters all over the world. Increasing rainfall intensity significantly escalates soil erosion and soil erosion related hazards. Forecasting accurate rainfall helps early detection of soil erosion vulnerability and can minimise the damages by taking appropriate measures caused by severe storms, droughts and floods. This study aims to predict soil erosion probability using the deep learning approach: long short-term memory neural network model (LSTM) and revised universal soil loss equation (RUSLE) model. Daily rainfall data were gathered from five agro-meteorological stations in the Central Highlands of Sri Lanka from 1990 to 2021 and fed into the LSTM model simulation. The LSTM model was forecasted with the time-series monthly rainfall data for a long lead time period, rainfall values for next 36 months in each station. Geo-informatics tools were used to create the rainfall erosivity map layer for the year 2024. The RUSLE model prediction indicates the average annual soil erosion over the Highlands will be 11.92 t/ha/yr. Soil erosion susceptibility map suggests around 30 % of the land area will be categorised as moderate to very-high soil erosion susceptible classes. The resulted map layer was validated using past soil erosion map layers developed for 2000, 2010 and 2019. The soil erosion susceptibility map indicates an accuracy of 0.93 with the area under the receiver operator characteristic curve (AUC-ROC), showing a satisfactory prediction performance. These findings will be helpful in policy-level decision making and researchers can further tested different deep learning models with the RUSLE model to enhance the prediction capability of soil erosion probability.

Spatial based drought assessment: Where are we heading? A review on the current status and future.

Droughts are the most spatially complex natural hazards that exert global impacts and are further aggravated by climate change. The investigation of drought events is challenging as it involves numerous factors ranging from detection and assessment to modelling, management and mitigation. The analysis of these factors and their quantitative assessments have significantly evolved in recent times. In this paper, we review recent methods used to examine and model droughts from a spatial viewpoint. Our analysis was conducted at three spatial scales (point-wise, regional and global) and we evaluated how recent spatial methods have advanced our understanding of drought through case study examples. Further, we also examine and provide a broad overview of relevant case studies related to future drought occurrences under climate change. This study is a comprehensive synthesis of the various quantitative techniques used to assess the spatial characteristics of droughts at different spatial scales, and not an exhaustive review of all drought aspects. However, this serves as a basis for understanding the key milestones and advances accomplished through new spatial concepts relative to the traditional approaches to study drought. This work also aims to address the gaps in knowledge that are in need of further attention and provides recommendations to improve our understanding of droughts.

Unboxing Deep Learning Model of Food Delivery Service Reviews Using Explainable Artificial Intelligence (XAI) Technique.

The demand for food delivery services (FDSs) during the COVID-19 crisis has been fuelled by consumers who prefer to order meals online and have it delivered to their door than to wait at a restaurant. Since many restaurants moved online and joined FDSs such as Uber Eats, Menulog, and Deliveroo, customer reviews on internet platforms have become a valuable source of information about a company's performance. FDS organisations strive to collect customer complaints and effectively utilise the information to identify improvements needed to enhance customer satisfaction. However, only a few customer opinions are addressed because of the large amount of customer feedback data and lack of customer service consultants. Organisations can use artificial intelligence (AI) instead of relying on customer service experts and find solutions on their own to save money as opposed to reading each review. Based on the literature, deep learning (DL) methods have shown remarkable results in obtaining better accuracy when working with large datasets in other domains, but lack explainability in their model. Rapid research on explainable AI (XAI) to explain predictions made by opaque models looks promising but remains to be explored in the FDS domain. This study conducted a sentiment analysis by comparing simple and hybrid DL techniques (LSTM, Bi-LSTM, Bi-GRU-LSTM-CNN) in the FDS domain and explained the predictions using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME). The DL models were trained and tested on the customer review dataset extracted from the ProductReview website. Results showed that the LSTM, Bi-LSTM and Bi-GRU-LSTM-CNN models achieved an accuracy of 96.07%, 95.85% and 96.33%, respectively. The model should exhibit fewer false negatives because FDS organisations aim to identify and address each and every customer complaint. The LSTM model was chosen over the other two DL models, Bi-LSTM and Bi-GRU-LSTM-CNN, due to its lower rate of false negatives. XAI techniques, such as SHAP and LIME, revealed the feature contribution of the words used towards positive and negative sentiments, which were used to validate the model.

Sentiment Analysis of Customer Reviews of Food Delivery Services Using Deep Learning and Explainable Artificial Intelligence: Systematic Review.

During the COVID-19 crisis, customers' preference in having food delivered to their doorstep instead of waiting in a restaurant has propelled the growth of food delivery services (FDSs). With all restaurants going online and bringing FDSs onboard, such as UberEATS, Menulog or Deliveroo, customer reviews on online platforms have become an important source of information about the company's performance. FDS organisations aim to gather complaints from customer feedback and effectively use the data to determine the areas for improvement to enhance customer satisfaction. This work aimed to review machine learning (ML) and deep learning (DL) models and explainable artificial intelligence (XAI) methods to predict customer sentiments in the FDS domain. A literature review revealed the wide usage of lexicon-based and ML techniques for predicting sentiments through customer reviews in FDS. However, limited studies applying DL techniques were found due to the lack of the model interpretability and explainability of the decisions made. The key findings of this systematic review are as follows: 77% of the models are non-interpretable in nature, and organisations can argue for the explainability and trust in the system. DL models in other domains perform well in terms of accuracy but lack explainability, which can be achieved with XAI implementation. Future research should focus on implementing DL models for sentiment analysis in the FDS domain and incorporating XAI techniques to bring out the explainability of the models.

Solving transparency in drought forecasting using attention models.

Droughts are one of the most devastating and recurring natural disaster due to a multitude of reasons. Among the different drought studies, drought forecasting is one of the key aspects of effective drought management. The occurrence of droughts is related to a multitude of factors which is a combination of hydro-meteorological and climatic factors. These variables are non-linear in nature, and neural networks have been found to effectively forecast drought. However, classical neural nets often succumb to over-fitting due to various lag components among the variables and therefore, the emergence of new deep learning and explainable models can effectively solve this problem. The present study uses an Attention-based model to forecast meteorological droughts (Standard Precipitation Index) at short-term forecast range (1-3 months) for five sites situated in Eastern Australia. The main aim of the work is to interpret the model outcomes and examine how a deep neural network achieves the forecasting results. The plots show the importance of the variables along with its short-term and long-term dependencies at different lead times. The results indicate the importance of large-scale climatic indices at different sequence dependencies specific to the study site, thus providing an example of the necessity to build a spatio-temporal explainable AI model for drought forecasting. The use of such interpretable models would help the decision-makers and planners to use data-driven models as an effective measure to forecast droughts as they provide transparency and trust while using these models.

Generalizability assessment of COVID-19 3D CT data for deep learning-based disease detection.

BACKGROUND: Artificial intelligence technologies in classification/detection of COVID-19 positive cases suffer from generalizability. Moreover, accessing and preparing another large dataset is not always feasible and time-consuming. Several studies have combined smaller COVID-19 CT datasets into "supersets" to maximize the number of training samples. This study aims to assess generalizability by splitting datasets into different portions based on 3D CT images using deep learning. METHOD: Two large datasets, including 1110 3D CT images, were split into five segments of 20% each. Each dataset's first 20% segment was separated as a holdout test set. 3D-CNN training was performed with the remaining 80% from each dataset. Two small external datasets were also used to independently evaluate the trained models. RESULTS: The total combination of 80% of each dataset has an accuracy of 91% on Iranmehr and 83% on Moscow holdout test datasets. Results indicated that 80% of the primary datasets are adequate for fully training a model. The additional fine-tuning using 40% of a secondary dataset helps the model generalize to a third, unseen dataset. The highest accuracy achieved through transfer learning was 85% on LDCT dataset and 83% on Iranmehr holdout test sets when retrained on 80% of Iranmehr dataset. CONCLUSION: While the total combination of both datasets produced the best results, different combinations and transfer learning still produced generalizable results. Adopting the proposed methodology may help to obtain satisfactory results in the case of limited external datasets.

BreaCNet: A high-accuracy breast thermogram classifier based on mobile convolutional neural network.

The presence of a well-trained, mobile CNN model with a high accuracy rate is imperative to build a mobile-based early breast cancer detector. In this study, we propose a mobile neural network model breast cancer mobile network (BreaCNet) and its implementation framework. BreaCNet consists of an effective segmentation algorithm for breast thermograms and a classifier based on the mobile CNN model. The segmentation algorithm employing edge detection and second-order polynomial curve fitting techniques can effectively capture the thermograms' region of interest (ROI), thereby facilitating efficient feature extraction. The classifier was developed based on ShuffleNet by adding one block consisting of a convolutional layer with 1028 filters. The modified Shufflenet demonstrated a good fit learning with 6.1 million parameters and 22 MB size. Simulation results showed that modified ShuffleNet alone resulted in a 72% accuracy rate, but the performance excelled to a 100% accuracy rate when integrated with the proposed segmentation algorithm. In terms of diagnostic accuracy of the normal and abnormal test, BreaCNet significantly improves the sensitivity rate from 43% to 100% and specificity of 100%. We confirmed that feeding only the ROI of the input dataset to the network can improve the classifier's performance. On the implementation aspect of BreaCNet, the on-device inference is recommended to ensure users' data privacy and handle an unreliable network connection.