Stardust

Forestfire Detection using Meta and XAI Techniques The problem: Forest fires are cheapest to stop early, but smoke is easy to miss, and a plain classifier only gives a label with no context. In our baseline, about a third of smoke images were missed. What it does: The system watches a video feed and screens every sampled frame with a fast local classifier. When it flags a possible fire or smoke event, Muse Spark 1.3 reviews selected keyframes and returns a structured verdict, a severity level, and a plain-language summary. SAM 3.1 then segments the smoke and flame regions from short text prompts, so each alert comes with visual evidence of where the system sees the event. The alert goes to a Telegram chat with the summary, model confidence, a before/after video clip, and Confirm / False-alarm buttons, so a person always makes the final call. How Meta’s models are used: Spark 1.3 and SAM 3.1 run live through the Meta Model API, and only on flagged events. That keeps cost and latency low Prior work: The local classifier is a MobileNetV2 fine-tuned in my earlier mini-project work. Everything else, including the video pipeline, agent, SAM evidence, Telegram alerts, and evaluation, was built during the hackathon. Limitations: Evaluation used a small test subset and a synthetic test video, confidences are model-reported rather than calibrated, and SAM masks show where, not why. Real-world footage is the next step.