Skip to content
SDGs Center UDINUS

Evidence record · UDINUS · SDGs

Land Cover Classification from UAV Imagery Using CNN Features and Machine Learning Techniques

Rektorat UDINUS 2025 Published 2026-10-30

Evidence record · UDINUS · SDGs

Unit
Rektorat UDINUS
Period
2025
Goals
Record
f633d9eb-5331-4760-9e98-36e372cad255
Published
2026-10-30

Universitas Dian Nuswantoro (UDINUS), with Sari, Yuslena; Mustamin, Nurul Fathanah; Wijaya, Eka Setya; Noor, Muhammad Fajrian; Alkaff, Muhammad; Tommy, Muhammad; Mambang; Pramunendar, Ricardus Anggi; and Maulidiya, Erika as activity participants, conducted a study on land-cover classification using unmanned aerial vehicle (UAV) imagery from the Liang Anggang Protected Forest. The study addressed the role of land cover in geographic analysis, including sustainable environmental planning. Its objective was to classify UAV imagery according to vegetation density.

The UAV data were processed with Convolutional Neural Network (CNN) architectures, namely ResNet 50 and Inception-V3, for feature extraction. The extracted features were subsequently classified with machine learning algorithms, including Support Vector Machine (SVM), Random Forest, and eXtreme Gradient Boosting (XGBoost). In CNN-based classification, ResNet 50 achieved the best reported performance, with an accuracy of 91.04% at the 100th epoch. The combination of ResNet 50 and SVM produced an accuracy of 85% when applied as a feature extraction and classification method. XGBoost produced an accuracy of 82%, while Inception-V3 showed a lower maximum accuracy of 57.22%.

The study highlights ResNet 50 for its time efficiency and accuracy in vegetation-density classification. The results report that ResNet 50 was more reliable and faster than Inception-V3 for feature extraction and classification tasks. The research also states that the method improves vegetation-based land-cover classification in high-resolution UAV imagery. The ResNet 50 and SVM combination achieved the highest reported accuracy for the combined method and outperformed other models in processing time, making it suitable for time-sensitive applications. The proposed model can be integrated with edge computing frameworks for real-time environmental monitoring, including early wildfire detection and forest health surveillance. The study also identifies the integration of learning approaches as a possible area for future work. In relation to SDG 15, the activity concerns land ecosystems through research on land cover, vegetation density, and environmental monitoring.

Reference https://www.scopus.com/pages/publications/105025967704?origin=resultslist