Neural Network Prediction of Cropland Dynamics in the Vitebsk Region Using Google Earth Engine Land-Cover Data
Year: 2026
Type: article
Abstract: Accurate prediction of cropland redistribution is important for regional environmental assessment, agricultural planning, and biodiversity-oriented land management, especially in territories where agricultural activity is spatially interwoven with forests, natural vegetation, transport corridors, and protected areas. This study develops and validates a neural-network framework for forecasting cropland dynamics in the Vitebsk Region of Belarus using harmonised land-cover maps prepared in Google Earth Engine from the MODIS MCD12Q1 product for 2001, 2005, 2010, 2015, and 2020. The source maps were harmonised into five classes: forest, natural vegetation, cropland, urban land, and water. A compact spatiotemporal neural model was designed to estimate cropland probability from historical land-cover states, transition indicators, and neighbourhood descriptors. To strengthen methodological reproducibility, a retrospective validation experiment was performed: the model was trained using the 2001–2015 sequence and then used to forecast the known 2020 cropland distribution. The predicted 2020 map was compared with the observed 2020 land-cover map using spatially separated test blocks. The proposed neural model achieved an overall accuracy of 0.946, cropland F1-score of 0.876, intersection over union of 0.779, and Cohen’s kappa of 0.842 on the test subset, outperforming simple persistence and history-rule baselines. After validation, the model was applied to the 2005–2020 sequence to produce a 2030 cropland scenario. The results indicate that cropland remains concentrated in historically stable agricultural clusters, with limited redistribution at margins and no abrupt expansion into predominantly forested landscapes.
Language: English
Subfield: Ecology
Field: Environmental Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235544577
