A Data-Driven Latent Digital Twin Framework for Sustainable Agroecosystem Management

In recent years, digital twins have been increasingly used in agroecosystem management. However, integrating heterogeneous satellite observations into a unified framework remains challenging. This study developed a latent-state Digital Twin framework using Sentinel-2, MODIS, and SMAP observations from 23 agricultural fields at the North Kazakhstan Agricultural Experimental Station (NKAES) LLP during 2020–2025. These observations were combined with NASA POWER meteorological data. A three-dimensional multilayer perceptron autoencoder (MLP-AE-3D) was used to represent vegetation, thermal, and soil moisture conditions in a unified latent space. The model achieved a test R2 of 0.9654. To model temporal dynamics, MLP, GRU, and TSN architectures with different hidden dimensions were evaluated. GRU-512 was selected as the temporal prediction model, achieving a test R2 of 0.7962 and an RMSE of 0.0749. A Kalman-based Latent Assimilation Unit (LAU) was used to update predicted latent states while accounting for uncertainty. The adaptive J-RA-LAU configuration further reduced the RMSE to 0.0701. Controlled climate-stress scenarios were used to assess latent-state and vegetation responses to heat, drought, and compound environmental stressors. The proposed framework provides a unified dynamic representation to support field-level monitoring, climate-stress assessment, and sustainable agroecosystem management.