Visión por satélite y deep learning para el monitoreo de variables agrícolas en agricultura de precisión

Abstract

Abstract: The current study uses satellite remote sensing and deep learning methods to explore precision agriculture relevant variables. This workflow was built upon a multimodal 64x64 mosaic (including inputs from Sentinel-1, Sentinel-2, ERA5, and SRTM), and 22 spectral bands, including 11 tabular variables. After cleaning of the data for 2018 2023, a total of 9167 records were obtained from the crops class subset to model productivity variables. Mosaic extraction has been performed through Google Earth Engine following integrity control, as well as error treatment in order to maintain operational traceability. The model reached GBWP and NBWP (kgm⁻³) estimation for each cell in a multitasking fashion based on the spatial branch of a pre-trained ResNet34 and the tabular branch of an MLP, as well as NPP (net biomass). The results demarcated that the 2v model chosen during epoch 13 from validation produced R² values of 0.62, 0.57 and 0.77 during testing for GBWP, NBWP and NPP respectively. An additional system, SmartAgroEc, built on Streamlit and interlinked with Google Earth Engine was developed to map Ranching Areas of Interest (AOIs) using Random Forest classification, obtain annual NDVI for each AOI mapped by year and calculate P−PET water stress and export monthly data tables and series. This process creates a platform suitable for deep inference analysis. The workflow proposed here is put forward as a viable alternative for making decisions and sustainable monitoring in environments with high climate variability, technological limitations, and limited access to sophisticated methodologies.