Classification of Vegetation Land Cover Area Using Convolutional Neural Network

Abstract

The decrease and reduction of vegetation land or forest area over time has become a serious and significant problem to be considered. Increasing the Earth’s temperature is a consequence of deforestation, which can contribute to climate change. The other issues that researchers face concern diversity and various objects in satellite imagery that may be difficult for computers to identify using traditional methods. This research aims to develop a model that can classify vegetation land cover areas on high-resolution images. The data used is sourced from the ISPRS (International Society for Photogrammetry and Remote Sensing) Vaihingen. The model used is a Convolutional Neural Network (CNN) with a VGG16-Net Encoder architecture. Tests were conducted on eight scenarios with training and test data ratios of 80:20% and 70:30%. The classifier method that we employed in this research is argmax and threshold. We also compared the performance of Neural Networks with two hidden layers and three hidden layers to investigate the impact of adding another layer on the Neural Network's performance in classifying vegetation land cover areas. The results show that using the threshold classifier method can save training time compared to the argmax method. By increasing the number of hidden layers in the neural network, model performance improves, as shown by increases in recall, accuracy, and F1-score metrics. However, there is a slight decrease in the precision metric. The model achieved its best performance with a precision (Pre) of 99.5%, accuracy (Acc) of 83.3%, and F1-score (Fs) of 70.3%, requiring a training time (T-time) of 16 minutes and 41 seconds and an inference time (I-time) of 0.1535 seconds.