Deep Learning Based Tree Detection and Counting for Remotely Sensed Images

Trees are an essential part of the environment and some trees are important economic crops whose count in a specific region is an important factor in the prediction of the yield of the product which they give. The number of trees in a particular region helps to monitor the growing situation of trees. In this proposed method, tree detection has been done using a deep learning based framework and the counting of these trees has been done using remote sensing high-resolution images for two regions in the state of Uttarakhand, India. The trees in our areas of study are congested, often leading to an overlap of crowns. Two multi-spectral images have been provided for the paper. The first image has four channels namely Red, Green, Blue (RGB) and Near-Infrared (NIR). For the first image provided, a variety of manually interpreted samples for the training as well as the optimization of the convolutional neural network (CNN) have been used. Thereafter, using the sliding window technique, the prediction of the labels of the samples in the image dataset has been carried out. The proposed model provides a weighted accuracy of over 98% during training and validation. Additionally, the text analyzes the results obtained in case the near-infrared band is removed from this image with four channels (i.e. in second image).