Environmental dynamics in arid and semi-arid regions have become increasingly challenging in recent years owing to climate change. Traditional Cellular Automata (CA) models used for Land Use Land Cover (LULC) change prediction rely on static transition rules that cannot adequately capture the complex, nonlinear spatial dependencies and temporal dynamics present in real-world land use changes. In this study, we developed a Hybrid Spectral Network (HybridSN) for LULC classification using Sentinel-2 A satellite images. Additionally, the study utilized a Markov-chain integrated Random Forest (RF) was used to predict future LULC patterns from 2024 to 2030. The dataset comprised 409 Sentinel-2 A images, with varying numbers collected each year (ranging from 22 to 194) in the Adhaim River Basin in Northern Iraq. The HybridSN exhibited exceptional performance in classifying LULC throughout the study period, as evidenced by the consistent improvement in the Overall Accuracy (OA), which increased from 0.796 in 2017 to 0.921 in 2023. Comparative analysis revealed that the HybridSN model consistently outperformed both 2D and 3D CNNs across all years, which can be attributed to its ability to effectively capture the spatial and spectral features of LULC. The LULC forecasting for 2022 and 2023 was conducted using a Markov-chain integrated RF model, which demonstrated high accuracy in short-term predictions. The predicted LULC maps for the period of 2023 to 2030 highlighted significant trends, including a 22.2% decline in water coverage, a 55.0% reduction in tree cover, and an 11.6% increase in built-up areas. The results also highlighted the potential of hybrid deep-learning architectures for improving the accuracy of LULC mapping and forecasting using remote sensing time series. By improving the capacity of classification and prediction models, this study provides enhanced solutions for understanding the dynamics of the environment in arid and semi-arid regions.