UAV LAND COVER CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK FEATURE MAP WITH A COMBINATION OF MACHINE LEARNING

Abstract

In geographic analysis, land cover plays an important role in everything from environmental analysis to sustainable planning methods and physical geography studies. The Indonesian National Standard (SNI 7645:2014) classifies vegetation analysis based on density. There are four categories of vegetation density index: non-vegetation, bare, medium, and high. Technically, vegetation data can be obtained through remote sensing. Satellite and UAV data are two types of data used in remote sensing to collect information. This research will analyze land cover based on vegetation density information that can be collected through remote sensing. Based on vegetation density information from remote sensing, the information can help in land processing, Land Cover Classification is carried out based on vegetation density. Convolutional neural networks (CNN) have been trained extensively to apply their properties to land cover classification. This research will evaluate features extracted from Convolutional Neural Networks (ResNet 50, Inception-V3, DenseNet 121) which have previously been trained and continued with Decision Tree algorithms, Random Forest, Support Vector Machine and eXtreme Gradient Boosting to perform classification. From the comparison results of classification tests between machine learning methods, Support Vector Machine is superior to other machine learning methods. This is proven by the accuracy results obtained at 85% with feature extraction using ResNet-50 where the processing time is 8 minutes. Followed by the second-best model, namely ResNet-50 with XGBoost which obtained accuracy results of 82% with a processing time of 55 minutes. Meanwhile, the use of feature extraction using the DenseNet-121 method was obtained using a combination of the Support Vector Machine method and the XGBoost method with the accuracy obtained being 81%.