Accurate wildfire prediction in sub-Saharan Africa has long been limited by coarse burned area (BA) products that fail to detect the small, low-intensity fires that dominate savanna fire regimes. Recent advances in Sentinel-2 based BA mapping have revealed far higher BA totals, motivating the need to reassess predictive models. This study develops a ConvLSTM- and Vision Transformer-based spatiotemporal wildfire forecasting models trained with the high resolution (20 meter) FireCCISFD11 dataset for western Burkina Faso. Incorporating key climatic, ecological, and anthropogenic predictors, the model learns fine scale fire dynamics and achieves substantial performance gains compared to traditional machine learning baselines and a ConvLSTM trained only on coarse BA labels. Significantly, while pixel-level precision remains challenging, the model represents regional fire patterns more accurately than legacy BA estimates which have historically been treated as ground truth. Results highlight both the promise of deep spatiotemporal wildfire modeling and the critical importance of continued high resolution BA dataset development.