Soil water is essential for maintaining global food security and for understanding hydrological, meteorological, and ecosystem processes under climate change. Successful monitoring and forecasting of soil water dynamics at high spatio-temporal resolutions globally are hampered by the heterogeneity of soil hydraulic properties in space and complex interactions between water and the environmental variables that control it. Current soil water monitoring schemes via station networks are sparsely distributed while remote sensing satellite soil moisture maps have a very coarse spatial resolution. In this study, an empirical surface soil moisture (SSM) model was established via data fusion of remote sensing (Sentinel-1 and Soil Moisture Active and Passive Mission - SMAP) and land surface parameters (e.g. soil texture, terrain) using a quantile random forest (QRF) algorithm. The model had a spatial resolution of 100 m and performed moderately well across the globe under cropland, grassland, savanna, barren, and forest soils (R = 0.53, RMSE = 0.08 m m). SSM was retrieved and mapped at 100 m every 6-12 days in selected irrigated cropland and rainfed grassland in the OZNET network, Australia. It was concluded that the high-resolution SSM maps can be used to monitor soil water content at the field scale for irrigation management. The SSM model is an additive and adaptable model, which can be further improved by including soil moisture network measurements at the field scale. Further research is required to improve the temporal resolution of the model and map soil water content within the root zone.

Recently, statistical machine learning and deep learning methods have been widely explored for corn yield prediction. Though successful, machine learning models generated within a specific spatial domain often lose their validity when directly applied to new regions. To address this issue, we designed an unsupervised adaptive domain adversarial neural network (ADANN). Specifically, through domain adversarial training, the ADANN model reduced the impact of domain shift by projecting data from different domains into the same subspace. Also, the ADANN model was designed to be trained in an adaptive way, which guaranteed the model can learn the domain-invariant features and perform accurate yield prediction simultaneously. Informative variables including time-series vegetation indices and sequential weather observations were first collected from multiple data sources and aggregated to the county level. Then, we trained the ADANN model with the extracted features and corresponding reported county-level corn yield from the U.S. Department of Agriculture (USDA). Finally, the trained model was evaluated in four testing years 2016-2019. The U.S. corn belt was used as the study area and counties under study were grouped into two diverse ecological regions. The experimental results showed that the developed ADANN model had better performance than three other state-of-the-art machine learning models in both local experiments (train and test in the same region) and transfer experiments (train and test in different regions). As the first study using adversarial learning for crop yield prediction, this research demonstrates a novel solution for improving model transferability on crop yield prediction.

Maize (Zea mays L.) is one of the most consumed grains in the world and improving maize yield is of great importance for food security, especially under global climate change and more frequent severe droughts. However, traditional phenotyping methods relying on manual assessment are time-consuming and prone to human errors. Recently, the application of unmanned aerial vehicles (UAVs) has gained increasing attention in plant phenotyping due to their efficiency in data collection. Moreover, hyperspectral sensors integrated with UAVs can offer data streams with high spectral and spatial resolutions, which are essential for estimating plants’ physiological and biochemical traits. In this study, we developed machine learning models to estimate grain yield and flowering time of maize breeding lines using multi-temporal UAV-based hyperspectral imagery. The performance of multiple machine learning models and the efficacy of different hyperspectral features were evaluated on Genomes to Fields (G2F) experimental sites in Wisconsin. Results showed that ridge regression is the most robust model in estimating grain yield and flowering time, compared to random forest and support vector regression models. Furthermore, the ridge regression model achieved a correlation coefficient (r) of 0.551 for yield, 0.906 for days to silking, and 0.914 for days to anthesis when using the full-bands spectra features for estimation. In addition, we assessed the modeling performance using data acquired from different growing stages. The best time of applying the UAV survey was also identified in order to reduce the data collection efforts.