Ensemble machine learning prediction of anaerobic co-digestion of manure and thermally pretreated harvest residues

This study aimed to clarify the statistical accuracy assessment approaches used in recent biogas prediction studies using state-of-the-art ensemble machine learning approach according to 10-fold cross-validation in 100 repetitions. Three thermally pretreated harvest residue types (maize stover, sunflower stalk and soybean straw) and manure were anaerobically co-digested, measuring biogas and methane yield alongside eight thermal preprocessing and biomass covariates. These were the inputs to an ensemble machine learning approach for biogas and methane yield prediction, employing three feature selection approaches. The Support Vector Machine prediction with the Recursive Feature Elimination resulted in the highest prediction accuracy, achieving the coefficient of determination of 0.820 and 0.823 for biogas and methane yield prediction, respectively. This study demonstrated an extreme dependency of prediction accuracy to input dataset properties, which could only be mitigated with ensemble machine learning and strongly suggested that the split-sample approach, often used in previous studies, should be avoided.

Geospatial prediction of total soil carbon in European agricultural land based on deep learning

Accurate geospatial prediction of soil parameters provides a basis for large-scale digital soil mapping, making efficient use of the expensive and time-consuming process of field soil sampling. To date, few studies have used deep learning for geospatial prediction of soil parameters, but there is evidence that it may provide higher accuracy compared to machine learning methods. To address this research gap, this study proposed a deep neural network (DNN) for geospatial prediction of total soil carbon (TC) in European agricultural land and compared it with the eight most commonly used machine learning methods based on studies indexed in the Web of Science Core Collection. A total of 6209 preprocessed soil samples from the Geochemical mapping of agricultural and grazing land soil (GEMAS) dataset in heterogeneous agricultural areas covering 4,899,602 km2 in Europe were used. Prediction was performed based on 96 environmental covariates from climate and remote sensing sources, with extensive comprehensive hyperparameter tuning for all evaluated methods. DNN outperformed all evaluated machine learning methods (R2 = 0.663, RMSE = 9.595, MAE = 5.565), followed by Quantile Random Forest (QRF) (R2 = 0.635, RMSE = 25.993, MAE = 22.081). The ability of DNN to accurately predict small TC values and thus produce relatively low absolute residuals was a major reason for the higher prediction accuracy compared to machine learning methods. Climate parameters were the main factors in the achieved prediction accuracy, with 23 of the 25 environmental covariates with the highest variable importance being climate or land surface temperature parameters. These results demonstrate the superiority of DNN over machine learning methods for TC prediction, while highlighting the need for more recent soil sampling to assess the impact of climate change on TC content in European agricultural land.

A wildfire growth prediction and evaluation approach using Landsat and MODIS data

The increasing wildfire occurrence due to global climate changes urged the improvement of present wildfire growth prediction and evaluation methods. This study aimed to propose novel solutions to their two primary limitations, including the lack of robust fuel classification method and the low spatial resolution of wildfire growth accuracy assessment while ensuring wide applicability using open data satellite missions and software. The first objective was to create a robust two-step fuel model classification method consisted of the supervised machine learning classification of generalized land cover classes in the 1st level and their individual unsupervised classification to vegetation subtypes in the 2nd level. The second objective was creating a wildfire prediction accuracy assessment method using MODIS 250 m images, which overcome the limitations of low spatial resolution while preserving sub-daily temporal resolution. The wildfire on the Korčula island in Croatia was analyzed in the study, being specific for its long duration from 18 to 24 July 2015. The wildfire ignition occurred in the isolated area, which prolonged the response time from emergency agencies. Random Forest (RF) with input Landsat 8 spectral bands and indices resulted in the highest classification accuracy in the 1st classification level with an overall agreement of 83.6%. The vegetation subclasses from the 2nd classification level were matched to the 13 standard fuel models for the input in FARSITE software. The predicted wildfire evaluation showed the highest mean accuracy of 0.906 for the first two days, which decreased to 0.722 in the latter stages of the active wildfire caused by overprediction. The proposed two-step fuel model classification presented a cost-efficient solution to the fuel map creation in any part of the world, with a disadvantage of no in-situ ground truth identification and accuracy assessment for 2nd classification level. The evaluation of wildfire growth prediction with 250 m images enabled high spatial and temporal resolution of the assessment, while its limitations of wildfire overprediction and the negative effects of wildfire smoke in MODIS images should be addressed in future research.