Search for: All records

Masked autoencoders employ random masking to effectively reconstruct input images using self-supervised techniques, which allows for efficient training on large datasets. However, the random masking strategy does not adequately tap into information encapsulated within high-dimensional hyperspectral satellite imagery that is used in several domains. We propose a novel masking strategy, HOGMAE, based on the Histogram of Oriented Gradients that incorporates rich information inherent within satellite images during the mask creation step. Our experiments, over a hyperspectral satellite dataset, demonstrate the effectiveness of our methodology. 

Off-road vehicle mobility assessments rely on fine-resolution (~10 m) estimates of soil moisture and strength across the region of interest. Such estimates are often produced by downscaling soil moisture from a microwave satellite like SMAP, then using the soil moisture in a soil strength model. Soil moisture downscaling methods typically assume consistent relationships between the moisture and topographic, vegetation, and soil composition characteristics within the microwave satellite grid cells. The objective of this study is to examine whether soil moisture and strength exhibit heterogenous dependencies on topography, vegetation, and soil composition characteristics within a SMAP grid cell. Soil moisture and strength data were collected at four geographically separated regions within a 9 km SMAP grid cell in the Front Range foothills of northern Colorado. Laboratory methods and pedotransfer functions were used to characterize soil attributes, and remote sensing data were used to determine topographic and vegetation attributes. Pearson correlation analyses were used to quantify the direction, strength, and significance of the relationships of both soil moisture and strength with topography, vegetation, and soil composition. Contrary to the common assumption, spatial variations in the slope and correlation of the relationships are observed for both soil moisture and strength. The findings indicate that improved predictions of soil moisture and soil strength may be achievable by soil moisture downscaling procedures that use spatially variable parameters across the downscaling extent. 

This study explores machine learning for estimating soil moisture at multiple depths (0–5 cm, 0–10 cm, 0–20 cm, 0–50 cm, and 0–100 cm) across the coterminous United States. A framework is developed that integrates soil moisture from Soil Moisture Active Passive (SMAP), precipitation from the Global Precipitation Measurement (GPM), evapotranspiration from the Ecosystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS), vegetation data from the Moderate Resolution Imaging Spectroradiometer (MODIS), soil properties from gridded National Soil Survey Geographic (gNATSGO), and land cover information from the National Land Cover Database (NLCD). Five machine learning algorithms are evaluated including the feed-forward artificial neural network, random forest, extreme gradient boosting (XGBoost), Categorical Boosting, and Light Gradient Boosting Machine. The methods are tested by comparing to in situ soil moisture observations from several national and regional networks. XGBoost exhibits the best performance for estimating soil moisture, achieving higher correlation coefficients (ranging from 0.76 at 0–5 cm depth to 0.86 at 0–100 cm depth), lower root mean squared errors (from 0.024 cm3/cm3 at 0–100 cm depth to 0.039 cm3/cm3 at 0–5 cm depth), higher Nash–Sutcliffe Efficiencies (from 0.551 at 0–5 cm depth to 0.694 at 0–100 cm depth), and higher Kling–Gupta Efficiencies (0.511 at 0–5 cm depth to 0.696 at 0–100 cm depth). Additionally, XGBoost outperforms the SMAP Level 4 product in representing the time series of soil moisture for the networks. Key factors influencing the soil moisture estimation are elevation, clay content, aridity index, and antecedent soil moisture derived from SMAP.