Alfalfa is a valuable and intensively produced forage crop in the United States, and the timely estimation of its yield can inform precision management decisions. However, traditional yield assessment approaches are laborious and time-consuming, and thus hinder the acquisition of timely information at the field scale. Recently, unmanned aerial vehicles (UAVs) have gained significant attention in precision agriculture due to their efficiency in data acquisition. In addition, compared with other imaging modalities, hyperspectral data can offer higher spectral fidelity for constructing narrow-band vegetation indices which are of great importance in yield modeling. In this study, we performed an in-season alfalfa yield prediction using UAV-based hyperspectral images. Specifically, we firstly extracted a large number of hyperspectral indices from the original data and performed a feature selection to reduce the data dimensionality. Then, an ensemble machine learning model was developed by combining three widely used base learners including random forest (RF), support vector regression (SVR) and K-nearest neighbors (KNN). The model performance was evaluated on experimental fields in Wisconsin. Our results showed that the ensemble model outperformed all the base learners and a coefficient of determination (R²) of 0.874 was achieved when using the selected features. In addition, we also evaluated the model adaptability on different machinery compaction treatments, and the results further demonstrate the efficacy of the proposed ensemble model.

Potato is the largest non-cereal food crop in the world. Timely estimation of end-of-season tuber production using in-season information can inform sustainable agricultural management decisions that increase productivity while reducing impacts on the environment. Recently, unmanned aerial vehicles (UAVs) have become increasingly popular in precision agriculture due to their flexibility in data acquisition and improved spatial and spectral resolutions. In addition, compared with natural color and multispectral imagery, hyperspectral data can provide higher spectral fidelity which is important for modelling crop traits. In this study, we conducted end-of-season potato tuber yield and tuber set predictions using in-season UAV-based hyperspectral images and machine learning. Specifically, six mainstream machine learning models, i.e., ordinary least square (OLS), ridge regression, partial least square regression (PLSR), support vector regression (SVR), random forest (RF), and adaptive boosting (AdaBoost), were developed and compared across potato research plots with different irrigation rates at the University of Wisconsin Hancock Agricultural Research Station. Our results showed that the tuber set could be better predicted than the tuber yield, and using the multi-temporal hyperspectral data improved the model performance. Ridge achieved the best performance for predicting tuber yield (R² = 0.63) while Ridge and PLSR had similar performance for predicting tuber set (R² = 0.69). Our study demonstrated that hyperspectral imagery and machine learning have good potential to help potato growers efficiently manage their irrigation practices.

Over the past few years, there has been extensive exploration of machine learning (ML), especially deep learning (DL), for crop yield prediction, resulting in impressive levels of accuracy. However, such models are highly dependent on training samples with ground truth labels (i.e., crop yield records), which are not available in some regions. Additionally, due to the existence of domain shifts between different spatial regions, DL models trained within one region (i.e., source domain) tend to have poor performance when directly applied to other regions (i.e., target domain). Unsupervised domain adaptation (UDA) has become a promising strategy to improve the transferability of DL models by aligning the feature distributions in the source domain and the target domain. Despite the success, existing UDA models generally assume an identical label space across different domains. This assumption can be invalid in crop yield prediction scenarios, as crop yields can vary significantly in heterogeneous regions. Due to the mismatch between label spaces, negative transfer may occur if the entire source and target domains are forced to align. To address this issue, we proposed a novel partial domain adversarial neural network (PDANN), which relaxes the assumption of fully, equally shared label spaces across domains by downweighing the outlier source samples. Specifically, during model training, the PDANN weighs each labeled source sample based on the likelihood of its yield value given the expected target yield distribution. Instead of aligning the target domain to the entire source domain, the PDANN model downweighs the outlier source samples and performs partial weighted alignment of the target domain to the source domain. As a result, the negative transfer caused by source samples in the outlier label space would be alleviated. In this study, we assessed the model’s performance on predicting yields for two main commodities in the U.S., including corn and soybean, using the U.S. corn belt as the study region. The counties under study were divided into two distinct ecological zones and alternatively used as the source and target domains. Feature variables, including time-series vegetation indices (VIs) and sequential meteorological variables, were collected and aggregated at the county level. Next, the PDANN model was trained with the extracted features and corresponding crop yield records from the U.S. Department of Agriculture (USDA). Finally, the trained model was evaluated for three testing years from 2019 to 2021. The experimental results showed that the developed PDANN model had achieved a mean coefficient of determination (R2) of 0.70 and 0.67, respectively, in predicting corn and soybean yields, outperforming three other ML and UDA models by a large margin from 6% to 46%. As the first study performing partial domain adaptation for crop yield prediction, this research demonstrates a novel solution for addressing negative transfer and improving DL models’ transferability on crop yield prediction.

Wetland methane (CH₄) emissions have a significant impact on the global climate system. However, the current estimation of wetland CH₄ emissions at the global scale still has large uncertainties. Here we developed six distinct bottom‐up machine learning (ML) models using in situ CH₄ fluxes from both chamber measurements and the Fluxnet‐CH₄ network. To reduce uncertainties, we adopted a multi‐model ensemble (MME) approach to estimate CH₄ emissions. Precipitation, air temperature, soil properties, wetland types, and climate types are considered in developing the models. The MME is then extrapolated to the global scale to estimate CH₄ emissions from 1979 to 2099. We found that the annual wetland CH₄ emissions are 146.6 ± 12.2 Tg CH₄ yr⁻¹ (1 Tg = 10¹² g) from 1979 to 2022. Future emissions will reach 165.8 ± 11.6, 185.6 ± 15.0, and 193.6 ± 17.2 Tg CH₄ yr⁻¹ in the last two decades of the 21st century under SSP126, SSP370, and SSP585 scenarios, respectively. Northern Europe and near‐equatorial areas are the current emission hotspots. To further constrain the quantification uncertainty, research priorities should be directed to comprehensive CH₄ measurements and better characterization of spatial dynamics of wetland areas. Our data‐driven ML‐based global wetland CH₄ emission products for both the contemporary and the 21st century shall facilitate future global CH₄ cycle studies.

Spatially explicit crop yield datasets with continuous long-term series are essential for understanding the spatiotemporal variation of crop yield and the impact of climate change on it. There are several spatial disaggregation methods to generate gridded yield maps, but these either use an oversimplified approach with only a couple of ancillary data or an overly complex approach with limited flexibility and scalability. This study developed a spatial disaggregation method using improved spatial weights generated from machine learning. When applied to Chinese maize yield, extreme gradient boosting (XGB) derived the best prediction results, with a cross-validation coefficient of determination (R²) of 0.81 at the municipal level. The disaggregated yield at 1 km grids could explain 54% of the variance of the county-level statistical yield, which is superior to the existing gridded maize yield dataset in China. At the site level, the disaggregated yields also showed much better agreement with observations than the existing gridded maize yield dataset. This lightweight method is promising for generating spatially explicit crop yield datasets with finer resolution and higher accuracy, and for providing necessary information for maize production risk assessment in China under climate change.

Potato growth depends largely on nitrogen (N) availability in the soil. However, the shallow-root crop coupled with its common cultivation in coarse-textured soils leads to its poor N use efficiency. Fast and accurate estimations of potato tissue N concentrations are urgently needed to assist the decision making in precision fertilization management. Remote sensing has been utilized to evaluate the potato N status by correlating spectral information with lab tests on leaf N concentrations. In this study, a systematic comparison was conducted to quantitatively evaluate the performance of hyperspectral and multispectral images in estimating the potato N status, providing a reference for the trade-off between sensor costs and performance. In the experiment, two potato varieties were planted under four fertilization rates with replicates. UAV images were acquired multiple times during the season with a narrow-band hyperspectral imager. Multispectral reflectance was simulated by merging the relevant narrow bands into broad bands to mimic commonly used multispectral cameras. The whole leaf total N concentration and petiole nitrate-N concentration were obtained from 160 potato leaf samples. A partial least square regression model was developed to estimate the two N status indicators using different groups of image features. The best estimation accuracies were given by reflectance of the full spectra with 2.2 nm narrow, with the coefficient of determination (R²) being 0.78 and root mean square error (RMSE) being 0.41 for the whole leaf total N concentration; while, for the petiole nitrate-N concentration, the 10 nm bands had the best performance (R² = 0.87 and RMSE = 0.13). Generally, the model performance decreased with an increase of the spectral bandwidth. The hyperspectral full spectra largely outperformed all three multispectral cameras, but there was no significant difference among the three brands of multispectral cameras. The results also showed that spectral bands in the visible regions (400–700 nm) were the most highly correlated with potato N concentrations.