Soil respiration is the largest carbon efflux from the terrestrial ecosystem to the atmosphere, and selective logging influences soil respiration via changes in abiotic (temperature, moisture) and biotic (biomass, productivity, quantity and quality of necromass inputs) drivers. Logged forests are a predominant feature of the tropical forest landscape, their area exceeding that of intact forest. We quantified both total and component (root, mycorrhiza, litter, and soil organic matter, SOM) soil respiration in logged (n = 5) and old-growth (n = 6) forest plots in Malaysian Borneo, a region which is a global hotspot for emission from forest degradation. We constructed a detailed below-ground carbon budget including organic carbon inputs into the system via litterfall and root turnover. Total soil respiration was significantly higher in logged forests than in old-growth forests (14.3 ± 0.23 and 12.7 ± 0.60 Mg C ha-1 year-1 , respectively, p = 0.037). This was mainly due to the higher SOM respiration in logged forests (55 ± 3.1% of the total respiration in logged forests vs. 50 ± 3.0% in old-growth forests). In old-growth forests, annual SOM respiration was equal to the organic carbon inputs into the soil (difference between SOM respiration and inputs 0.18 Mg C ha-1 year-1 , with 90% confidence intervals of -0.41 and 0.74 Mg C ha-1 year-1 ), indicating that the system is in equilibrium, while in logged forests SOM respiration exceeded the inputs by 4.2 Mg C ha-1 year-1 (90% CI of 3.6 and 4.9 Mg C ha-1 year-1 ), indicating that the soil is losing carbon. These results contribute towards understanding the impact of logging on below-ground carbon dynamics, which is one of the key uncertainties in estimating emissions from forest degradation. This study demonstrates how significant perturbation of the below-ground carbon balance, and consequent net soil carbon emissions, can persist for decades after a logging event in tropical forests.
Along with propagating the input toward making a prediction, Bayesian neural networks also propagate uncertainty. This has the potential to guide the training process by rejecting predictions of low confidence, and recent variational Bayesian methods can do so without Monte Carlo sampling of weights. Here, we apply sample-free methods for wildlife call detection on recordings made via passive acoustic monitoring equipment in the animals' natural habitats. We further propose uncertainty-aware label smoothing, where the smoothing probability is dependent on sample-free predictive uncertainty, in order to downweigh data samples that should contribute less to the loss value. We introduce a bioacoustic dataset recorded in Malaysian Borneo, containing overlapping calls from 30 species. On that dataset, our proposed method achieves an absolute percentage improvement of around 1.5 points on area under the receiver operating characteristic (AU-ROC), 13 points in F1, and 19.5 points in expected calibration error (ECE) compared to the point-estimate network baseline averaged across all target classes.