Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Master → Pays Bas → 2020 → Exploring the use of neural networks for object instance detection below the pixel resolution level : finding date palm trees using Sentinel-2 imagery and FCNN’s

Wageningen University (2020)

Exploring the use of neural networks for object instance detection below the pixel resolution level : finding date palm trees using Sentinel-2 imagery and FCNN’s

Buddingh, Laurens

Titre : Exploring the use of neural networks for object instance detection below the pixel resolution level : finding date palm trees using Sentinel-2 imagery and FCNN’s

Auteur : Buddingh, Laurens

Université de soutenance : Wageningen University

Grade : Master of Science (MS) Geo-information Science and Remote Sensing 2020

Résumé partiel
Fully convolutional neural networks (FCNN’s) are a promising technique to use spatial and spectral information within imagery for object detection and image segmentation. A well-built and functioning FCNN model that can successfully classify small objects using relatively coarse satellite imagery would open pathways for the creation of extensive datasets in both space and time. This study seeks to implement and assess the use of FCNN models to detect Phoenix Canariensis date palm trees in the Canary Islands. Due to the use of Sentinel-2 imagery as a source of reflectance data, the size of the objects (date palms) is between 0.6 and 0.8 times the size of a reflectance pixel. Since the palm trees are smaller than the pixel size of the reflectance data, it is difficult to correctly identify the palm trees. However, an evaluation over similar data was done by a previous study with promising results. The starting point of this study was to see if the same could be done for the data and problem of this study. This study first establishes the full pipeline of the data from its source to the finished model and then discusses the predictions. The ground truth data used for the study was a pointbased dataset gathered through field surveys, covering the whole of the Canary island region, making the whole of the Canary Islands the sampling area for this study. The sampling area was cut into chunks and the ones containing no positive ground truths were discarded (no palm trees), before being loaded into the model in a 8/2 training validation split. Building on previous works as a foundation, Deeplabv3 model architecture was implemented through python (Pytorch) code to create the finished model, and its fidelity was assessed using F1 and Au-ROC scores and its model convergence was assessed using the losses obtained during training and testing. With an optimal F1 of 2.763% the usability and fidelity of the model is at a point where it is rarely able to correctly assign the palm tree label on pixels correctly. The model is thus unsuitable for the accurate mapping of palm tree locations for further research. While the Au_ROC score scored significantly better at 0.74476 this was judged to be a side effect due to bias obvious within the ground truth data, greatly oversampling urban areas in comparison to other types of land use, likely as a result of the ease of access these locations provided to the field surveyors compared to the more dense forests and natural areas.

Mots Clés : Remote sensing, neural networks, deep learning, semantic segmentation

Présentation

Version intégrale

Page publiée le 15 avril 2021