Final project exploring subdistrict/year poverty-risk estimation for East Java by combining satellite-derived features, POI/LULC signals, news-based NLP event scores, deep-learning models, SHAP explanations, and retrieval-augmented analysis.
This repository contains the curated research notebooks, reports, lightweight result tables, and selected documentation. The interactive frontend is maintained separately as east-java-poverty-map-web.
PCD/: satellite image-feature preprocessing, LULC/thresholding experiments, and model aggregation.NLP/: local-news scraping, pseudolabeling/event-score workflows, and RAG experiments.DL/: multimodal/temporal deep-learning experiments and SHAP analyses.Data-DL/: compact model-ready event, target, and metric tables retained for context.- Root Markdown/PDF/DOCX files: methodology, revision notes, and final project reporting.
Large raw news corpora, BPS/RPJMD documents, satellite patches, shapefiles, full prediction dumps, checkpoints, and archives are not committed. The included CSV/JSON files are compact research outputs or small feature tables. Obtain authoritative source data from BPS, government planning documents, satellite providers, or the course-provided sources before rerunning full pipelines.
Notebook outputs were removed during cleanup. Some notebooks retain Kaggle/Modal-oriented setup because they were executed there. API settings must be supplied through environment variables. Results should be interpreted according to the methodology and caveats in the included reports.