UCLA

The primary mission of NASA's Earth Surface Mineral Dust Source Investigation (EMIT) is to improve our understanding of mineral dust in atmospheric radiative forcing using a novel set of spectroscopically-based mineral observations. In order to estimate mineral mass abundance, information on median surface grainsize is required. Knowledge of clay and silt fractions as well as the sub-components of the sand fractions is required for calculation of median grainsize. Therefore, this study aimed to produce a map of surface soil texture suitable for the intended application. Spectral reflectance data were incorporated into a multi-output Random Forest model to predict mass fraction for seven particle size classes: silt (TSI), clay (Clay), and five sand size classes: very coarse sand (S1, 1–2 mm diameter), coarse sand (S2, 1/2–1 mm), medium sand (S3, 1/4–1/2 mm), fine sand (S4, 1/8–1/4 mm), and very fine sand (S5, 1/16–1/8 mm diameter). To account for the presence of vegetation in EMIT spectral data, spectra of soils were mixed linearly with spectra of green and nonphotosynthetic vegetation in various combinations and amounts and the mixed spectra were used as training data. A five-fold cross-validation approach allowed us to estimate independent errors for each of the texture classes as well as three parameters estimated post hoc: total sand fraction (TSA, calculated as the sum of the five sand classes), mean grain size (GSmean), and median grain size (GSmedian). Sand fractions have estimated mean absolute errors (MAEs) <6% (R2 = 0.54–0.81), while the error for TSA is <12% (R2 = 0.78). MAE for TSI and Clay is <4% (R2 = 0.66) and < 7% (R2 = 0.81), respectively. Estimated MAE for GSmedian is 30 μm (R2 = 0.78). For GSmean estimated MAE is 7.3 μm (R2 = 0.73). Above 70% soil cover (below 30% vegetation cover), vegetation did not appear to influence estimates. A final model trained on all available data was used to produce estimates of soil texture class mass fractions for dust-producing areas worldwide at 0.1° resolution suitable for supporting estimates of mineral mass fraction from EMIT data. Despite drastically different estimation approaches, our results are broadly consistent with other global mineral datasets, although differences are apparent. Overall, our approach indicates the value of imaging spectrometer data for estimation of surface grain size for application to EMIT's primary mission of better constraining dust radiative forcing, though additional paired soil spectra-texture measurements would improve results. The products may also be useful for other applications in the Earth sciences, especially those related to global modeling of dust.