Leveraging Remote Sensing and Machine Learning to Assess the Impact of Home Proximity to Agriculture on Pesticide Exposure, Depression, and Anxiety in an Ecuadorian Agricultural Community.
Abstract
Introduction: Living near agriculture increases exposure risks through movement of volatilized pesticides, known as pesticide drift. Evidence from the Study of Secondary Exposures to Pesticides among Children and Adolescents (ESPINA) found suspected floricultural pesticide drift and subsequent worse mood for individuals living in Pedro Moncayo, Ecuador, underscoring the need to assess contributions from broader agriculture. Yet this region lacks contemporary agriculture maps.
Methods: To map agriculture (Aim 1), Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) and the Segment Anything Model (SAM) helped identify temporal endmembers for time series of Sentinel-2 spectrally unmixed vegetation fractions and PlanetScope enhanced vegetation index 2 imagery. Aim 2 and 3 used the agriculture map from Aim 1 and 2022 ESPINA data. These Aims assessed the relationship between home distance to the nearest agricultural field boundary (home-ag distance) and surface areas of agriculture within buffers around the home (home-ag SA) with urinary organophosphate, pyrethroid, and herbicide metabolite concentrations (Aim 2, n=500), and with anxiety and depression symptoms (Aim 3, n=484). Associations were estimated using generalized estimating equations, logistic regression, Getis-ord Gi* hotspot analysis, and geospatially weighted regression (GWR), adjusting for confounders.
Results: SAM identified more agricultural plots for PlanetScope compared to Sentinel-2 and UMAP pre-processing enhanced performance. Compared to PlanetScope, Sentinel-2 agricultural maps provided greater accuracy and sensitivity. The final agriculture map was 72.6% accurate and 75.3% precise (Aim 1). Greater home-ag distance was negatively associated with 2,4-Dichlorophenoxyacetic acid (2,4-D) and malathion dicarboxylic acid (MDA). GWR found the strongest associations between home-ag distance and 2,4-D near an industrial agriculture plot. Conversely, having a greater agricultural presence was negatively associated with para-nitrophenol and 3-phenoxybenzoic acid (Aim 2). Living farther from agriculture was associated with lower anxiety and depression symptoms. Spatial variations of home-ag distance with anxiety and depression symptoms was observed (Aim 3).
Conclusion: This study demonstrates that integrating manifold learning and SAM for agricultural characterization of high- and low-resolution imagery can achieve moderately accurate agricultural maps. Our results found that pesticide drift of 2,4-D and MDA may be stemming from agricultural production, and pesticide drift may increase risk of worse depression and anxiety symptoms.