The Use of Multi-Modal Satellite Imagery for Detection of Palmyra Trees A Machine Learning-Based Approach

Abstract

A manual census of trees over a large geographic area can be very costly. Remote sensing is a powerful tool for the task. In this regard, this research focuses on the Google Earth image-based detection and counting of Palmyra trees in the northern part of Sri Lanka. Freely accessible Google Earth images are for the first time used here for the detection of specific tree type. Color information is used to identify their foliage. As the color information itself can be ambiguous at times, a complimentary analysis in the form of the identification of shadows is also carried out. Here, the fact that these tall trees throw a considerable shadow on the ground or other lower lying features is exploited. Phase Stretch Transform is used to identify the shadows. Furthermore, object-based image analysis is used on high resolution QuickBird images for detection of Palmyra on larger area. Multi-resolution segmentation and supervised nearest neighbor classification is used for that purpose. The detection results in successfully extracted Palmyra among other vegetation in Google Earth images and shows with a precision of 92.6% and recall of 88%. On QuickBird images, precision and recall values are found to be 92% and 95.5% respectively