Automatic segmentation of tea fields by using deep learning algorithms

Automatic classification of forest areas and tree species has vital importance for environmental management issue. Urban and forest map production is realized by using photogrammetric techniques in our country since 1940's. The traditional forest map generation by stereoscopic evaluation is time consuming. In this thesis, a new method is proposed using deep learning methods for automatic extraction of tea fields. Deep learning algorithms are machine learning algorithms that automatically extract features from data. It gained more popularity in recent times and have been successfully used in various computer vision tasks such as voice recognition, object recognition, handwriting recognition and image classification. The objective of this thesis was to extract tea fields from high resolution imagery using deep learning-based approach. To achieve this objective, SegNet architecture with VGG19 network structure had been adopted to classify images into tea fields and non-tea fields. The deep learning based algorithms were trained using various combination of various spectral channels. These combinations include RGB (red-green-blue), NirRG (near infrared-red-green), NDVI (normalized difference vegetation index) and NDVIRG (NDVI-red-green). The multispectral images were acquired using an aerial digital camera at 30x30 cm ground sample distance. The algorithms used in the thesis have been implemented using 2017b version of Matlab® which is licensed by Yildiz Technical University. In this thesis, 1:5000 scaled 25 very high resolution orthophoto images of Hayrat region of Trabzon have been used. The aerial images for orthophoto creation were taken on April 2013. Each orthophoto image consists of four bands which are red, green, blue and near-infrared (RGBNir). The radiometric resolution of the orthophoto images was 8 bits. Each orthophoto map covers 453.6 ha area on the earth. For evaluation of the proposed method, 15 images (60%) have been used for training and 10 (40%) images for testing. SegNet architecture with VGG19 network structure was trained using different amount of images with different band combinations to test its efficiency under varying amount of data. The training data has been prepared from 15 orthophoto images by randomly selecting data in size of 200x200 pixels image patches. 250,400,600 and 1000 training image-set have been used respectively and system has been trained and the results have been analyzed. The photogrammetrically evaluated tree field segments have been taken as reference data for accuracy assessment. Sørensen-Dice Similarity Coefficient algorithm has been utilized for accuracy assessment. As a result, three fields have been segmented in 90.91%, 91.16%, 91.51% and 53.39% accuracy for RGB, NirRG, NDVIRG and NDVI band combinations respectively. Compare to traditional photogrammetric evaluation, the proposed method is very efficient in terms of both accuracy and processing time. Normally, stereoscopic evaluation required approximately 8 hours to process a single image. On the other side, the average processing time of the proposed method to classify the same image was 5 minutes.