Huaccha Castillo,Annick Estefany2026-01-222026-01-222025-12-10http://hdl.handle.net/20.500.14689/1058Global warming, driven by rising greenhouse gas (GHG) concentrations, has agriculture as a major source of emissions. In coffee plantations, low sampling frequency and the absence of diurnal baselines introduce bias in emission estimates. The objective of this research was to estimate diurnal CO₂, N₂O, and CH₄ emissions from unfertilized coffee soils using recurrent neural networks (RNN). Gas fluxes were measured with a closed dynamic chamber (CDC) at 20-minute intervals between 8:00 and 18:00 over 22 days. For the estimation of GHG emissions, climatic data measured through a meteorological station were used, in addition to environmental parameters incorporated in the CDC. Five RNN models composed of two hidden layers of 20, 25, and 50 neurons were developed, trained, and validated for each GHG. Results indicate that N₂O contributed most to total emissions (734,689 ppm CO₂-eq), with CO₂ (237,579 ppm CO₂-eq) and CH₄ (215,426 ppm CO₂-eq) contributing less. Model performance was strong, with R² values of 0.98 (CO₂), 0.96 (N₂O), and 0.94 (CH₄). It is concluded that the RNNs proved to be reliable models for predicting GHG emissions in unfertilized coffee soils, with this study presenting a replicable framework with the potential to improve temporal estimation and reduce uncertainty in GHG inventories.application/pdfenginfo:eu-repo/semantics/openAccessclimate changeGasesSoilAir pollutionArtificial intelligencegasessoilair pollutionartificial intelligenceinfo:eu-repo/semantics/articlehttps://doi.org/10.18686/cest544https://purl.org/pe-repo/ocde/ford#2.07.00