Geospatial Customizing Convolutional Neural Network (CNN) Model Application for Stunting in Tasikmalaya, Indonesia

Abstract

Stunting remains a persistent public health challenge, particularly in socio-economically diverse regions like Tasikmalaya, Indonesia. This study presents an innovative approach to predicting stunting hotspots by integrating geospatial and socio-economic data through a customizable Convolutional Neural Network (CNN) model. The model is designed with a dual-input architecture, processing geospatial raster data and structured variables simultaneously. Its flexibility allows users to modify key parameters, such as the number of layers, optimizer, batch size, learning rate, and geospatial focus areas, ensuring adaptability across different datasets and objectives. Geospatial data were acquired from the Tasikmalaya Open Data platform, while variables were collected through household surveys. Rigorous preprocessing, including data cleaning, augmentation, and georeferencing, ensured the integration of diverse data types. The model was trained and evaluated on 22 configurations to identify the optimal architecture, achieving the highest accuracy of 86.94% using the Adam optimizer, four layers, a learning rate of 0.0001, and a dropout rate of 0.2. Across configurations, accuracy ranged from 70.29% to 85.67%, demonstrating the model's robustness and adaptability in classifying stunting risk with stable accuracy.