Toward Predictive Remote Sensing for Precision Agriculture: Deep Learning Forecasting of Vegetation Spectral Dynamics

This PhD thesis investigates the potential of satellite time-series forecasting to support precision agriculture through anticipatory monitoring of crop and grassland dynamics. The research is grounded in the operational limitation of optical remote sensing. Although Sentinel- 2 provides frequent, multispectral observations suitable for vegetation, soil, and water monitoring, its practical use is still constrained by cloud cover and by delays between acquisition and the availability of analysis-ready products. These limitations become even more critical under increasingly frequent drought and water-stress conditions, when timely information is essential for irrigation planning, stress detection, and adaptive land management. The thesis addresses this gap by exploring whether machine learning and deep learning models can reliably forecast future spectral responses and vegetation indices from past Sentinel-2 and meteorological observations. The thesis develops a progressive research path across three studies. The first study demonstrates that field-level corn NDVI can be forecast with high accuracy through a Bi-LSTM model using Sentinel-2 indices, meteorological variables, day-of-year information, and an adaptive minibatch strategy designed to handle variable data availability. The second study extends the problem beyond NDVI by forecasting EVI and NDMI over corn fields in northern Italy under recurrent drought and heat-stress conditions, while also comparing conventional machine learning with sequential deep learning models and explicitly testing out-of-year and cross-region generalization. The third study further broadens the methodological scope through DeepGrass, a Transformer-based framework designed to forecast the Sentinel-2 spectral signature and twenty-eight vegetation indices over managed grasslands in Sweden, with additional inference in Finland, thus moving from a crop-specific and index-specific setting toward a more scalable spectral forecasting framework. Across the three studies, the results consistently show that sequence-aware deep learning models provide strong predictive skill and greater operational value than more conventional approaches, especially when the objective is to support near-term agronomic decisions under uncertainty. Overall, this thesis shows that forecasting satellite-derived vegetation signals is a credible way to complement Earth observation in agricultural monitoring. It demonstrates a clear methodological evolution from single-index forecasting in corn to multi-index forecasting under drought stress, and finally to multi-spectral, multi-index forecasting in grasslands. In doing so, it contributes both scientifically and operationally to the development of proactive decisionsupport tools for precision agriculture, offering scalable methods based mainly on open satellite and meteorological data. At the same time, the thesis highlights the remaining challenges linked 2 to transferability, field-level heterogeneity, cloud-related data scarcity, and the limited availability of detailed management information