Search

• Soil reflectance composites were generated with Sentinel-2 L2A data. • A multi-channel CNN learning framework for soil organic carbon prediction is proposed. • A dual input deep learning architecture improves the predictions. • The framework produced results that can be interpreted. The Copernicus Sentinel-2 multispectral imagery data may be aggregated to extract large-scale, bare soil, reflectance composites, which enable soil mapping applications. In this paper, this approach was tested in the German federal state of Bavaria, to provide estimations for soil organic carbon (SOC). Different temporal ranges were considered for the generation of the composites, including multi-annual and seasonal ranges. A novel multi-channel convolutional neural network (CNN) is proposed. By leveraging the advantages of deep learning techniques, it utilizes complementary information from different spectral pre-treatment techniques. The SOC predictions indicated little dissimilarity amongst the different composites, with the best performance attained for the six-year composite containing only spring months (RMSE = 12.03 g C · kg−1, R2 = 0.64, RPIQ = 0.89). It has been demonstrated that these outcomes outperform other well-known machine learning techniques. An ablation analysis was accordingly performed to evaluate the interplay of the CNN's different components to disentangle the advantages of each aspect of the proposed framework. Finally, a DUal inPut deep LearnIng architecture, named DUPLICITE, is proposed, which concatenates deep spectral features derived from the CNN mentioned earlier, as well as topographical and environmental covariates through an artificial neural network (ANN) to exploit their complementarity. The proposed approach was demonstrated to provide an improvement in the overall prediction performance (RMSE = 11.60 gC · kg−1, R2 = 0.67, RPIQ = 0.92). [ABSTRACT FROM AUTHOR]

Weather is a pivotal factor for crop production as it is highly volatile and can hardly be controlled by farm management practices. Since there is a tendency towards increased weather extremes in the future, understanding weather-related yield factors becomes increasingly important not only for yield prediction, but also for the design of insurance products. Although insurance products mitigate financial losses for farmers, they suffer from considerable basis risk, i.e., a discrepancy between losses and the indemnity payment. The objective of this paper was to explore the potential of machine learning for estimating the relationship between crop yield and weather conditions at the farm level and to use it as a tool for reducing basis risk in index insurance applications. An artificial neural network was set up and calibrated to a rich set of farm-level yield data in Germany, covering the period from 2003 to 2018. A nonlinear regression model, which uses rainfall, temperature, and soil moisture as explanatory variables for yield deviations, served as a benchmark. The empirical application revealed that compared with traditional estimation approaches, the gain in forecasting precision by using machine learning techniques was substantial. Moreover, the use of regionalized models and disaggregated high-resolution weather data improved the performance of artificial neural networks. A considerable part of yield variability at the farm level, however, could not be captured by statistical methods which solely use "big weather data". Our findings have important implications for the design of weather-index based insurance because they document that a rather high level of basis risk remains if insurance products are based on an estimation of the weather-yield relationship. [Display omitted] • Due to increasing weather extremes in the future, understanding weather-related yield factors becomes more important. • This paper explores the potential of machine learning for estimating farm-level yield losses and basis risk reduction. • Use of high-resolution weather data improves the performance of artificial neural networks for estimating yield losses. • More homogenous insurance zones can reduce the basis risk and improve estimation precision. • The findings have implications for the design of weather-index based insurances because a rather high level of basis risk remains. [ABSTRACT FROM AUTHOR]

• This paper explores probabilistic extensions to Deep Learning algorithms and their application on energy time series forecasting. • The methods tested are concrete Dropouts, deep Ensembles, bayesian neural Networks, deep gaussian Processes, and functional neural processes. • The methods are evaluated in terms of calibration, sharpness, and how well they indicate a lack of knowledge (Epistemic Uncertainty). • Good epistemic uncertainty estimates were given using functional neural Processes, deep Ensembles, and deep gaussian processes methods, while concrete dropouts and bayesian neural networks needed tunning for better performance. • The application of the methods was feasible for energy time series forecasting, and they produced good estimates of their confidence in a forecast. Deep Learning methods are widely successful and continue to be applied in new fields, including forecasting for energy time series. However, they lack an indication of their confidence in predictions, which helps evaluate, interpret, and improve the forecasting. This paper explores probabilistic extensions to Deep Learning algorithms and their application on energy time series forecasting. The methods tested are Concrete Dropouts, Deep Ensembles, Bayesian Neural Networks, Deep Gaussian Processes, and Functional Neural Processes. For evaluation, two load forecasting scenarios are considered: 1-hour-ahead and 24-hour-ahead load forecasting. The methods are evaluated in terms of calibration, sharpness, and how well they indicate a lack of knowledge (Epistemic Uncertainty). As a reference, simple Neural Network and Quantile Regression models are used. Also, the ARIMA model and Persistence model are implemented to compare the results. Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Continuous Ranked Probability Score (CRPS) are computed to measure the error performance. Based on endogenous load time series in Germany, the results showed that Concrete Dropout, Deep Ensembles, and Bayesian Neural Networks performed similarly well and better than the reference methods. They also indicated that the simpler Concrete Dropout and Deep Ensembles have an advantage in both speed and convergence time. Good Epistemic Uncertainty estimates were given using Functional Neural Processes, Deep Ensembles, and Deep Gaussian Processes methods, while Concrete dropouts and Bayesian Neural Networks needed tunning for better performance. The application of the methods was feasible for energy time series forecasting, and they produced good estimates of their confidence in a forecast. [ABSTRACT FROM AUTHOR]