Implementation of deep learning based semantic segmentation method to determine vegetation density

Abstract

The dryness of peatlands is influenced by the density of vegetation. If peatlands are dry, they become vulnerable to a fire risk. To calculate the drought index, professionals must conduct a vegetation density analysis. However, field analysis requires vast amounts of resources. Moreover, the accuracy of the analysis based on satellite data is not adequate. Therefore, this research presents drone-captured two-dimensional image data. The object of this research is The Liang Anggang Protection Forest Block I in Banjarbaru, South Kalimantan, Indonesia. It is surveyed for information on its vegetation cover. Afterwards, There are 300 images of vegetation cover collected and utilized in total. The method of deep learning with semantic segmentation will be used to compare the results of determining methods with expert results as ground truth. The contribution of this study is to determine the optimal performance of deep learning model used for classifying vegetation density into three categories: bare/ungrazed, lightly grazed, and heavily grazed. Performance is evaluated based on correctness and intersection over union (IoU). Obtaining the proper parameters for the classification model using deep learning techniques and comparing the results of the best segmentation model are the objectives of the following contribution. From experimental studies conducted, the optimal momentum parameter value for MobileNetV2, Xception, and Inception-ResNet-v2 is 0.9, and the optimal accuracy performance is 82.69 percent on average. The most appropriate momentum for ResNet 18 architecture is 0.1. The result of semantic segmentation using the DeepLabV3 model with Inception-ResNet-v2 architecture is the optimal model for estimating vegetation density compared to U-Net model.