Recognition and extraction of palm trees from satellite images with high spatial resolution of Google Earth based on residual deep learning networks

Subject Areas : Spatial data infrastructures and standardisation

Mostafa Kabolizadeh ¹ , Kazem Rangzan ² , Mohammas Abbasi ³

1 - RS and GIS department, Earth science faculty, shahiud Chamran university of Ahvaz
2 - RS and GIS department, earth science faculty, Shahid Chamran university of Ahvaz
3 - RS and GIS Department

Received: 2022-09-05 Accepted : 2022-11-27

Keywords: Features extraction, convolutional neural network, Machine Learning, Automatic recognition,

Abstract :

Identifying the location of trees is the first step to manage green spaces, gardens and forests. The preparation of the location map of the trees can be done by ground mapping operations, which require a lot of money and time, or by using aerial or satellite images. In this research, satellite images with the high spatial resolution of Google Earth have been used to detect and extract palm trees, considering the role and importance of palm trees in the southern regions of Iran, but automatic recognition of trees from satellite images is a challenge. In this regard, deep learning methods are considered as an important solution for extracting objects from images. In this research, residual deep learning methods with the number of layers 18, 34 and 50 have been used. First, more than 3000 image samples were cut in two classes containing palm trees and without palm trees with dimensions of 64 x 64 pixels, then the models were trained with 80% samples for learning and 20% for validation with 30 epochs. The training accuracy of the models has been above 99%. The trained model was implemented on 500 test samples and the evaluation results of all three models show that the precision is more than 0.96, the recall is equal to 1, and the F1Score is more than 0.98. Running the models on Google Earth satellite images by moving the 64 x 64 pixel window with a step of 16 pixels and applying the non maximum suppression method shows that the satellite images of the Google Earth system can be used to prepare a map of palm trees. Considering the processing time and the possibility of better estimating the number and extracting the position of palm trees, the residual deep learning model with 34 layers is suggested.