ABSTRACT: During a disaster, emergency teams receive text messages, photographs and location data from citizens faster than staff can read them, and connectivity is often degraded exactly when reports matter most. This paper presents Jalad Pratisad (DisasterAI Response), a decision-support system that ranks incoming citizen reports by urgency and shows them to responders on a live map and priority list. A fine-tuned DistilBERT classifier assigns each report to one of seven humanitarian categories; a ResNet-50 classifier labels an attached photograph as fire, flooded area, collapsed building, traffic incident or normal; and a reported-location density term captures how many other reports lie within 500 m. The three signals are combined by a transparent weighted score on a 0–10 scale. Reports can be written in English or any of twelve Indian languages (via a keyword vocabulary), are accepted only from inside India, and can be collected offline by a store-and-forward mesh relay and a queueing web app. Responders can correct the model, and each correction is stored as a labelled example. The text model, trained on 41,671 de-duplicated tweets from the CrisisNLP and HumAID corpora, reaches 90.2% test accuracy and 0.80 macro-F1 on 5,211 held-out tweets; the image model, trained on the AIDER aerial dataset plus real ground-level photographs, reaches 97.6% accuracy and 0.97 macro-F1 on 1,552 held-out images (96.7% on aerial and 99.2% on ground-level images). We state the limits plainly: the priority score has no ground-truth labels, the fusion weights are heuristic, the ground-level photographs come from existing collections rather than citizens’ phones, and no end-to-end field evaluation has been carried out.
KEYWORDS: disaster management; crisis informatics; DistilBERT; ResNet-50; incident prioritisation; multilingual; offline-first; decision support
JALAD PRATISAD (DISASTERAI RESPONSE): AN AI SYSTEM FOR PRIORITISING CITIZEN DISASTER REPORTS
MRS. ANJALI DANDEKAR, MR. VEDANT MHAPANKAR