Scientific literature on spatio-temporal data primarily focuses on predicting trajectories of moving objects or observing events over fixed geographical areas. Few studies address the predictive analysis of dynamic variations of spatial areas themselves over time. This research proposes a dual methodology to model the evolution of maritime piracy (?350 incidents/year globally): (1) quantitative prediction at macroscopic resolution (?4000 km/cell) for strategic resource allocation, and (2) qualitative classification at mesoscopic resolution (?1000 km/cell) to detect regional spatial transitions. Evaluation on 15,947 incidents (1978-2024) reveals that standard validation systematically overestimates complex models, particularly in regression where LSTM is significantly less effective in walk-forward validation. In classification, all models deteriorate in walk-forward validation, but Logistic Regression demonstrates robustness with degradation lower than complex architectures, becoming the best model under strict temporal validation. The proposed hybrid architecture (Ridge for regression and LSTM+Logistic for classification) offers temporal robustness for operational monitoring.