IMPLEMENTATION OF GEOSPATIAL INTELLIGENCE FOR SENTIMENT ANALYSIS ON STUNTING POLICY IN BATANG REGENCY USING INDOBERT

Abstract

This study develops a Geospatial Artificial Intelligence (GeoAI)-based WebGIS that integrates IndoBERT sentiment classification to evaluate public percetion of stunting-management policy in Batang Regency, Central Java, Indoneia. The Cross-Industri Standart Process for Data Mining (CRISP-DM) framework was applied to the sentiment-analysis pipeline, while Rapid Application Development (RAD) governed the system construction. A total of 478 public-opinion responses were collected through questionnaires from residents of fifteen sub-districts, preprocessed through data cleaning, case folding, tokenizing, stopword removal, and stemming, then labeled and classified into positive, neutral, and negative sentiment using a fine-tuned IndoBERT model. The system was built with Python, Flask, Leaflet.js, and QGIS to visualize sentiment spatially. On a held-out test set of 96 samples, the model achieved 82.29% accuracy, 83.49% weighted precision, 82.29% weighted recall, and an 82.58% weighted F1-score (macro F1-score of 0.77), with class-weighted loss applied during fine-tuning to counter a severe class imbalance in the labeled dataset (Imbalance Ratio = 5.82). Spatial analysis showed that Bandar Sub-district recorded both the highest number of positive (37) and negative (8) responses, indicating the highest level of public engagement, Batang Sub-district recorded the highest number of neutral responses (46). System functionality was further validated through User Acceptance Testing and Black Box Testing, each covering nine functional scenarios spanning authentication, dashoard acces, sentiment-analysis display, spatial map interaction, and page navigation; all eighteen test scenarios were completed successfully (100% valid), confirming that the system operates correctly and satisfies the intended user requirements. The resulting GeoAI-based WebGIS enable policymakers to identify sub-districts requiring closer attention and design more targeted, evidence-based interventions. These findings demonstrate that integrating sentiment classification with spatial visualization provides greater insight into public perception than statistical data.