An Adaptive and Modern Land Use / Land Cover Classification System for Indonesia Using Multi-Sensor Earth Observation Imagery and Data-Driven Techniques: Collecting Data and Training with Dashcam and Field Camera
Hadi F., Wahyuddin Y., Sabri L.M. , Suprayogi A., and Ramdani F

Universitas Diponegoro
University of Tsukuba

Abstract

The limited availability of publicly accessible land use and land cover (LULC) training datasets in Indonesia presents significant challenges for verifying historical image classification results. Traditional validation methods often rely on costly and logistically demanding field surveys, which are frequently prohibitive in scope and time. To address this critical gap, this study introduces a low-cost, scalable approach for LULC training data collection using consumer-grade dashcams equipped with Global Positioning System (GPS) functionality. These devices capture georeferenced video data embedded with spatial and temporal metadata. A comprehensive Python-based application was developed-assisted by Claude.ai and deployed in the Google Colab environment-to automate the extraction and processing of dashcam video frames. The system performs key tasks such as converting video to images, applying optical character recognition (OCR) for text detection, storing metadata in a structured database, and enabling public deployment via pyngrok for collaborative access. The application features four core modules: (1) camera metadata correction, (2) land cover labelling, (3) spatial coordinate adjustment, and (4) export functionality in both CSV and GeoJSON formats. From a test dataset of 21,945 images, 6,796 (30.96%) required manual verification of camera-derived information, highlighting the importance of integrated quality control in automated workflows. The platform^s web interface enables multi-user collaboration, significantly accelerating data validation and labelling compared to conventional field-based methods. Additionally, an interactive dashboard offers spatial filtering and region-specific download options, enhancing accessibility and usability across a range of research domains. This prototype lays the groundwork for a robust, accessible, and cost-efficient LULC classification framework in Indonesia.

Keywords: land cover classification, dashcam imagery, collaborative mapping, Indonesia datasets, automated processing

Topic: Topic D: Geospatial Data Integration