Explainable artificial intelligence (XAI) for exploring spatial variability of lung and bronchus cancer (LBC) mortality rates in the contiguous USA.

Machine learning (ML) has demonstrated promise in predicting mortality; however, understanding spatial variation in risk factor contributions to mortality rate requires explainability. We applied explainable artificial intelligence (XAI) on a stack-ensemble machine learning model framework to explore and visualize the spatial distribution of the contributions of known risk factors to lung and bronchus cancer (LBC) mortality rates in the conterminous United States. We used five base-learners-generalized linear model (GLM), random forest (RF), Gradient boosting machine (GBM), extreme Gradient boosting machine (XGBoost), and Deep Neural Network (DNN) for developing stack-ensemble models. Then we applied several model-agnostic approaches to interpret and visualize the stack ensemble model's output in global and local scales (at the county level). The stack ensemble generally performs better than all the base learners and three spatial regression models. A permutation-based feature importance technique ranked smoking prevalence as the most important predictor, followed by poverty and elevation. However, the impact of these risk factors on LBC mortality rates varies spatially. This is the first study to use ensemble machine learning with explainable algorithms to explore and visualize the spatial heterogeneity of the relationships between LBC mortality and risk factors in the contiguous USA.

Geographically weighted machine learning model for untangling spatial heterogeneity of type 2 diabetes mellitus (T2D) prevalence in the USA.

Type 2 diabetes mellitus (T2D) prevalence in the United States varies substantially across spatial and temporal scales, attributable to variations of socioeconomic and lifestyle risk factors. Understanding these variations in risk factors contributions to T2D would be of great benefit to intervention and treatment approaches to reduce or prevent T2D. Geographically-weighted random forest (GW-RF), a tree-based non-parametric machine learning model, may help explore and visualize the relationships between T2D and risk factors at the county-level. GW-RF outputs are compared to global (RF and OLS) and local (GW-OLS) models between the years of 2013-2017 using low education, poverty, obesity, physical inactivity, access to exercise, and food environment as inputs. Our results indicate that a non-parametric GW-RF model shows a high potential for explaining spatial heterogeneity of, and predicting, T2D prevalence over traditional local and global models when inputting six major risk factors. Some of these predictions, however, are marginal. These findings of spatial heterogeneity using GW-RF demonstrate the need to consider local factors in prevention approaches. Spatial analysis of T2D and associated risk factor prevalence offers useful information for targeting the geographic area for prevention and disease interventions.

Detecting functional field units from satellite images in smallholder farming systems using a deep learning based computer vision approach: A case study from Bangladesh.

Improving agricultural productivity of smallholder farms (which are typically less than 2 ha) is key to food security for millions of people in developing nations. Knowledge of the size and location of crop fields forms the basis for crop statistics, yield forecasting, resource allocation, economic planning, and for monitoring the effectiveness of development interventions and investments. We evaluated three different full convolutional neural network (F-CNN) models (U-Net, SegNet, and DenseNet) with deep neural architecture to detect functional field boundaries from the very high resolution (VHR) WorldView-3 satellite imagery from Southern Bangladesh. The precision of the three F-CNN was up to 0.8, and among the three F-CNN models, the highest precision, recalls, and F-1 score was obtained using a DenseNet model. This architecture provided the highest area under the receiver operating characteristic (ROC) curve (AUC) when tested with independent images. We also found that 4-channel images (blue, green, red, and near-infrared) provided small gains in performance when compared to 3-channel images (blue, green, and red). Our results indicate the potential of using CNN based computer vision techniques to detect field boundaries of small, irregularly shaped agricultural fields.

Mapping U.S. Food System Localization Potential: The Impact of Diet on Foodsheds.

In the long term, food systems must heed natural resource limits. Localized production and dietary changes are often suggested as potential solutions. However, no U.S. analyses fully evaluate the feasibility to scale localization across a range of diets. We therefore modeled the biophysical capacity for regional food systems based on agricultural land area and productivity, population, and 7 diet scenarios ranging in meat-intensity, from current consumption to vegan. We estimated foodshed size, colloquially known as "food miles" for 378 U.S. metropolitan centers, in a hypothetical nationwide closed system that prioritizes localized food. We found that foodshed size (weighted average distance traveled) for three land types ranged from 351-428 km (cultivated cropland), 80-492 km (perennial forage cropland), and 117-799 km (grazing land). Localized potential varies regionally: foodsheds are generally larger in the populous Northeast, Southeast, and Southwest than in the Northwest and the center of the country. However, depending on consumption of animal-based foods, a sizable proportion of the population could meet its food needs within 250km: from 35%-53% (cultivated cropland), 39%-94% (perennial forage cropland, 100% for vegan), and 26%-88% (grazing land, 100% for ovolacto-vegetarian and vegan). All seven scenarios leave some land unused. This reserve capacity might be used to supply food to the global market, grow bioenergy crops, or for conservation.

Factors affecting paddy soil arsenic concentration in Bangladesh: prediction and uncertainty of geostatistical risk mapping.

Knowledge of the spatial correlation of soil arsenic (As) concentrations with environmental variables is needed to assess the nature and extent of the risk of As contamination from irrigation water in Bangladesh. We analyzed 263 paired groundwater and paddy soil samples covering highland (HL) and medium highland-1 (MHL-1) land types for geostatistical mapping of soil As and delineation of As contaminated areas in Tala Upazilla, Satkhira district. We also collected 74 non-rice soil samples to assess the baseline concentration of soil As for this area. The mean soil As concentrations (mg/kg) for different land types under rice and non-rice crops were: rice-MHL-1 (21.2)>rice-HL (14.1)>non-rice-MHL-1 (11.9)>non-rice-HL (7.2). Multiple regression analyses showed that irrigation water As, Fe, land elevation and years of tubewell operation are the important factors affecting the concentrations of As in HL paddy soils. Only years of tubewell operation affected As concentration in the MHL-1 paddy soils. Quantitatively similar increases in soil As above the estimated baseline-As concentration were observed for rice soils on HL and MHL-1 after 6-8 years of groundwater irrigation, implying strong retention of As added in irrigation water in both land types. Application of single geostatistical methods with secondary variables such as regression kriging (RK) and ordinary co-kriging (OCK) gave little improvement in prediction of soil As over ordinary kriging (OK). Comparing single prediction methods, kriging within strata (KWS), the combination of RK for HL and OCK for MHL-1, gave more accurate soil As predictions and showed the lowest misclassification of declaring a location "contaminated" with respect to 14.8 mg As/kg, the highest value obtained for the baseline soil As concentration. Prediction of soil As buildup over time indicated that 75% or the soils cropped to rice would contain at least 30 mg/L As by the year 2020.