Abstract

Several Free and Open Source (FOSS) tools use Neural Networks for detection of objects from images and videos captured from hand-held imaging devices. Satellite based Remote Sensing images offer wide area coverage and hold potential for detecting, counting and mapping manmade objects, trees, etc., but, have embedded geospatial information and often have more than three bands. Hence, the existing FOSS tools are not able to directly process Remote Sensing images for Computer Vision (CV) applications. This research aims to devise a methodology to adapt a FOSS CV tool, namely the TensorFlow Object Detection (TFOD) API, for detection, counting and segmentation of coconut palms from satellite images and ascertain if the technique can facilitate automated census of coconut palms. Dataset of coconut palm crowns was custom-created using multi-band images from World View-3 satellite. The images were pan-sharpened, cropped and labelled. SSDLite MobileNet V2 CNN, which was pre-trained on COCO dataset, was subjected to transfer learning using coconut data on Tesla K80x GPU. This re-trained CNN could successfully detect and count coconut palms with F-1 score more than 96 %. Histogram thresholds were used to segment and delineate each detected coconut palm crown with 87 % accuracy. Assessment of relative health status of coconut palms was mapped using the Normalised Difference Red-Edge Index derived from satellite images. This study demonstrated that TFOD API can indeed be adapted for object detection and segmentation from Remote Sensing images, albeit with some limitations.