Learning vs. understanding: When does artificial intelligence outperform process-based modeling in soil organic carbon prediction?

In recent years, machine learning (ML) algorithms have gained substantial recognition for ecological modeling across various temporal and spatial scales. However, little evaluation has been conducted for the prediction of soil organic carbon (SOC) on small data sets commonly inherent to long-term soil ecological research. In this context, the performance of ML algorithms for SOC prediction has never been tested against traditional process-based modeling approaches. Here, we compare ML algorithms, calibrated and uncalibrated process-based models as well as multiple ensembles on their performance in predicting SOC using data from five long-term experimental sites (comprising 256 independent data points) in Austria. Using all available data, the ML-based approaches using Random forest and Support vector machines with a polynomial kernel were superior to all process-based models. However, the ML algorithms performed similar or worse when the number of training samples was reduced or when a leave-one-site-out cross validation was applied. This emphasizes that the performance of ML algorithms is strongly dependent on the data-size related quality of learning information following the well-known curse of dimensionality phenomenon, while the accuracy of process-based models significantly relies on proper calibration and combination of different modeling approaches. Our study thus suggests a superiority of ML-based SOC prediction at scales where larger datasets are available, while process-based models are superior tools when targeting the exploration of underlying biophysical and biochemical mechanisms of SOC dynamics in soils. Therefore, we recommend applying ensembles of ML algorithms with process-based models to combine advantages inherent to both approaches.

Understanding deep learning in land use classification based on Sentinel-2 time series.

The use of deep learning (DL) approaches for the analysis of remote sensing (RS) data is rapidly increasing. DL techniques have provided excellent results in applications ranging from parameter estimation to image classification and anomaly detection. Although the vast majority of studies report precision indicators, there is a lack of studies dealing with the interpretability of the predictions. This shortcoming hampers a wider adoption of DL approaches by a wider users community, as model's decisions are not accountable. In applications that involve the management of public budgets or policy compliance, a better interpretability of predictions is strictly required. This work aims to deepen the understanding of a recurrent neural network for land use classification based on Sentinel-2 time series in the context of the European Common Agricultural Policy (CAP). This permits to address the relevance of predictors in the classification process leading to an improved understanding of the behaviour of the network. The conducted analysis demonstrates that the red and near infrared Sentinel-2 bands convey the most useful information. With respect to the temporal information, the features derived from summer acquisitions were the most influential. These results contribute to the understanding of models used for decision making in the CAP to accomplish the European Green Deal (EGD) designed in order to counteract climate change, to protect biodiversity and ecosystems, and to ensure a fair economic return for farmers.

Multispectral high resolution sensor fusion for smoothing and gap-filling in the cloud.

Remote sensing optical sensors onboard operational satellites cannot have high spectral, spatial and temporal resolutions simultaneously. In addition, clouds and aerosols can adversely affect the signal contaminating the land surface observations. We present a HIghly Scalable Temporal Adaptive Reflectance Fusion Model (HISTARFM) algorithm to combine multispectral images of different sensors to reduce noise and produce monthly gap free high resolution (30 m) observations over land. Our approach uses images from the Landsat (30 m spatial resolution and 16 day revisit cycle) and the MODIS missions, both from Terra and Aqua platforms (500 m spatial resolution and daily revisit cycle). We implement a bias-aware Kalman filter method in the Google Earth Engine (GEE) platform to obtain fused images at the Landsat spatial-resolution. The added bias correction in the Kalman filter estimates accounts for the fact that both model and observation errors are temporally auto-correlated and may have a non-zero mean. This approach also enables reliable estimation of the uncertainty associated with the final reflectance estimates, allowing for error propagation analyses in higher level remote sensing products. Quantitative and qualitative evaluations of the generated products through comparison with other state-of-the-art methods confirm the validity of the approach, and open the door to operational applications at enhanced spatio-temporal resolutions at broad continental scales.

Exploring differences in spatial patterns and temporal trends of phenological models at continental scale using gridded temperature time-series.

Phenological models are widely used to estimate the influence of weather and climate on plant development. The goodness of fit of phenological models often is assessed by considering the root-mean-square error (RMSE) between observed and predicted dates. However, the spatial patterns and temporal trends derived from models with similar RMSE may vary considerably. In this paper, we analyse and compare patterns and trends from a suite of temperature-based phenological models, namely extended spring indices, thermal time and photothermal time models. These models were first calibrated using lilac leaf onset observations for the period 1961-1994. Next, volunteered phenological observations and daily gridded temperature data were used to validate the models. After that, the two most accurate models were used to evaluate the patterns and trends of leaf onset for the conterminous US over the period 2000-2014. Our results show that the RMSEs of extended spring indices and thermal time models are similar and about 2 days lower than those produced by the other models. Yet the dates of leaf out produced by each of the models differ by up to 11 days, and the trends differ by up to a week per decade. The results from the histograms and difference maps show that the statistical significance of these trends strongly depends on the type of model applied. Therefore, further work should focus on the development of metrics that can quantify the difference between patterns and trends derived from spatially explicit phenological models. Such metrics could subsequently be used to validate phenological models in both space and time. Also, such metrics could be used to validate phenological models in both space and time.

Optimized Kernel Entropy Components.

This brief addresses two main issues of the standard kernel entropy component analysis (KECA) algorithm: the optimization of the kernel decomposition and the optimization of the Gaussian kernel parameter. KECA roughly reduces to a sorting of the importance of kernel eigenvectors by entropy instead of variance, as in the kernel principal components analysis. In this brief, we propose an extension of the KECA method, named optimized KECA (OKECA), that directly extracts the optimal features retaining most of the data entropy by means of compacting the information in very few features (often in just one or two). The proposed method produces features which have higher expressive power. In particular, it is based on the independent component analysis framework, and introduces an extra rotation to the eigen decomposition, which is optimized via gradient-ascent search. This maximum entropy preservation suggests that OKECA features are more efficient than KECA features for density estimation. In addition, a critical issue in both the methods is the selection of the kernel parameter, since it critically affects the resulting performance. Here, we analyze the most common kernel length-scale selection criteria. The results of both the methods are illustrated in different synthetic and real problems. Results show that OKECA returns projections with more expressive power than KECA, the most successful rule for estimating the kernel parameter is based on maximum likelihood, and OKECA is more robust to the selection of the length-scale parameter in kernel density estimation.