Remote sensing studies of illicit crop detection have demonstrated strong retrospective performance, yet their operational value depends on how early and how confidently suspicious sites can be identified. Building directly on the Aceh Besar cannabis-detection framework of [1], this study re-analyzes the same cloud-screened Sentinel-2 time series, the same eradication-confirmed reference locations, and the same high-resolution visual validation source to convert a high-accuracy retrospective classifier into an early-warning surveillance system. No new imagery or field campaigns were introduced; the contribution lies in a controlled methodological extension of the existing dataset rather than in data expansion. We partition the original 17 usable Sentinel-2 acquisitions into progressively truncated temporal windows, retain the original class definitions, and extract a richer temporal-context feature space from the same red, green, blue, near-infrared, and NDVI channels. A calibrated stacked ensemble, combining the source-style backpropagation neural network with a random forest under site-blocked validation, is then used to quantify the earliest actionable detection point and to produce probability-ranked operational priorities. Using five-fold spatially blocked cross-validation across 254 labeled patch centers (87 cannabis, 90 forest, 77 shrub/scrub), the full-window model achieved 96.5% overall accuracy, a kappa coefficient of 0.947, cannabis precision of 0.976, and cannabis recall of 0.954, with limited fold-to-fold dispersion in the principal metrics. Critically, with only eight usable acquisitions, the model retained 93.7% overall accuracy and 0.897 cannabis recall, enabling actionable screening approximately four weeks earlier than full-window inference under the observed acquisition cadence. A targeted ablation confirmed that local context features account for most of the recall gain, while probability calibration reduced expected calibration error from 0.086 to 0.027 and materially improved field prioritization. The results therefore support earlier and more decision-ready screening within the observed Aceh Besar setting, while broader geographic transfer still requires separate validation.