Tree species classification from airborne hyperspectral and LiDAR data using 3D convolutional neural networks DOI Creative Commons
Janne Mäyrä, Sarita Keski‐Saari, Sonja Kivinen

et al.

Remote Sensing of Environment, Journal Year: 2021, Volume and Issue: 256, P. 112322 - 112322

Published: Feb. 15, 2021

During the last two decades, forest monitoring and inventory systems have moved from field surveys to remote sensing-based methods. These methods tend focus on economically significant components of forests, thus leaving out many factors vital for biodiversity, such as occurrence species with low economical but high ecological values. Airborne hyperspectral imagery has shown potential tree classification, most common analysis methods, random support vector machines, require manual feature engineering in order utilize both spatial spectral features, whereas deep learning are able extract these features raw data. Our research focused classification major Scots pine, Norway spruce birch, together an ecologically valuable keystone species, European aspen, which a sparse scattered boreal forests. We compared performance three-dimensional convolutional neural networks (3D-CNNs) machine, forest, gradient boosting machine artificial network individual data resolution. collected LiDAR along extensive ground reference measurements 83 km2 study area located southern zone Finland. A LiDAR-derived canopy height model was used match aerial imagery. The best performing 3D-CNN, utilizing 4 m image patches, achieve F1-score 0.91 overall 0.86 accuracy 87%, while lowest 3D-CNN 10 patches achieved 0.83 85%. In comparison, support-vector 0.82 82.4% 81.7%. Compared models, 3D-CNNs were more efficient distinguishing coniferous each other, concurrent aspen classification. Deep networks, being black box hide information about how they reach their decision. occlusion saliency maps interpret our models. Finally, we produce wall-to-wall map full that can later be prediction in, instance, mapping multispectral satellite images. improved demonstrated by benefit sustainable forestry biodiversity conservation.

Language: Английский

Machine learning approaches for crop yield prediction and nitrogen status estimation in precision agriculture: A review DOI
Anna Chlingaryan, Salah Sukkarieh,

Brett M. Whelan

et al.

Computers and Electronics in Agriculture, Journal Year: 2018, Volume and Issue: 151, P. 61 - 69

Published: June 5, 2018

Language: Английский

Citations

1137
Review of studies on tree species classification from remotely sensed data DOI
Fabian Ewald Fassnacht, Hooman Latifi, Krzysztof Stereńczak

et al.

Remote Sensing of Environment, Journal Year: 2016, Volume and Issue: 186, P. 64 - 87

Published: Aug. 20, 2016

Language: Английский

Citations

827
Tree Species Classification with Random Forest Using Very High Spatial Resolution 8-Band WorldView-2 Satellite Data DOI Creative Commons
Markus Immitzer, Clement Atzberger, Tatjana Koukal

et al.

Remote Sensing, Journal Year: 2012, Volume and Issue: 4(9), P. 2661 - 2693

Published: Sept. 14, 2012

Tree species diversity is a key parameter to describe forest ecosystems. It is, for example, important issues such as wildlife habitat modeling and close-to-nature management. We examined the suitability of 8-band WorldView-2 satellite data identification 10 tree in temperate Austria. performed Random Forest (RF) classification (object-based pixel-based) using spectra manually delineated sunlit regions crowns. The overall accuracy classifying was around 82% (8 bands, object-based). class-specific producer’s accuracies ranged between 33% (European hornbeam) 94% beech) user’s 57% 92% (Lawson’s cypress). object-based approach outperformed pixel-based approach. could show that 4 new bands (Coastal, Yellow, Red Edge, Near Infrared 2) have only limited impact on if main (Norway spruce, Scots pine, European beech, English oak) are be separated. However, increased significantly full spectral resolution further were included. Beside accuracy, importance evaluated with two measures provided by RF. An in-depth analysis RF output carried out evaluate reference quality resulting reliability final class assignments. Finally, an extensive literature review comprising about 20 studies presented.

Language: Английский

Citations

694
Domain Adaptation for the Classification of Remote Sensing Data: An Overview of Recent Advances DOI
Devis Tuia, Claudio Persello, Lorenzo Bruzzone

et al.

IEEE Geoscience and Remote Sensing Magazine, Journal Year: 2016, Volume and Issue: 4(2), P. 41 - 57

Published: June 1, 2016

The success of the supervised classification remotely sensed images acquired over large geographical areas or at short time intervals strongly depends on representativity samples used to train algorithm and define model. When training are collected from an image a spatial region that is different one for mapping, spectral shifts between two distributions likely make model fail. Such generally due differences in acquisition atmospheric conditions changes nature object observed. To design methods robust data set shifts, recent remote sensing literature has considered solutions based domain adaptation (DA) approaches. Inspired by machine-learning literature, several DA have been proposed solve specific problems classification. This article provides critical review advances approaches presents overview divided into four categories: 1) invariant feature selection, 2) representation matching, 3) classifiers, 4) selective sampling. We provide methodologies, examples applications techniques real characterized very high resolution as well possible guidelines selection method use application scenarios.

Language: Английский

Citations

477
Hyperspectral and LiDAR Data Fusion: Outcome of the 2013 GRSS Data Fusion Contest DOI Open Access

Christian Debes,

Andreas Merentitis,

Roel Heremans

et al.

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2014, Volume and Issue: 7(6), P. 2405 - 2418

Published: March 20, 2014

The 2013 Data Fusion Contest organized by the Technical Committee (DFTC) of IEEE Geoscience and Remote Sensing Society aimed at investigating synergistic use hyperspectral Light Detection And Ranging (LiDAR) data. data sets distributed to participants during Contest, a imagery corresponding LiDAR-derived digital surface model (DSM), were acquired NSF-funded Center for Airborne Laser Mapping over University Houston campus its neighboring area in summer 2012. This paper highlights two awarded research contributions, which investigated different approaches fusion LiDAR data, including combined unsupervised supervised classification scheme, graph-based method spectral, spatial, elevation information.

Language: Английский

Citations

474
Urban tree species mapping using hyperspectral and lidar data fusion DOI
Michael Alonzo, Bodo Bookhagen, Dar A. Roberts

et al.

Remote Sensing of Environment, Journal Year: 2014, Volume and Issue: 148, P. 70 - 83

Published: April 13, 2014

Language: Английский

Citations

462
Building Extraction in Very High Resolution Remote Sensing Imagery Using Deep Learning and Guided Filters DOI Creative Commons
Yongyang Xu, Liang Wu, Zhong Xie

et al.

Remote Sensing, Journal Year: 2018, Volume and Issue: 10(1), P. 144 - 144

Published: Jan. 19, 2018

Very high resolution (VHR) remote sensing imagery has been used for land cover classification, and it tends to a transition from land-use classification pixel-level semantic segmentation. Inspired by the recent success of deep learning filter method in computer vision, this work provides segmentation model, which designs an image neural network based on residual networks uses guided extract buildings imagery. Our includes following steps: first, VHR is preprocessed some hand-crafted features are calculated. Second, designed architecture trained with urban district at pixel level. Third, employed optimize map produced learning; same time, salt-and-pepper noise removed. Experimental results Vaihingen Potsdam datasets demonstrate that our method, benefits filtering, achieves higher overall accuracy when compared other machine methods. The proposed shows outstanding performance terms building extraction diversified objects district.

Language: Английский

Citations

440
Allometric equations for integrating remote sensing imagery into forest monitoring programmes DOI Creative Commons
Tommaso Jucker, John P. Caspersen, Jérôme Chave

et al.

Global Change Biology, Journal Year: 2016, Volume and Issue: 23(1), P. 177 - 190

Published: July 6, 2016

Remote sensing is revolutionizing the way we study forests, and recent technological advances mean are now able - for first time to identify measure crown dimensions of individual trees from airborne imagery. Yet make full use these data quantifying forest carbon stocks dynamics, a new generation allometric tools which have tree height size at their centre needed. Here, compile global database 108753 stem diameter, diameter all been measured, including 2395 harvested aboveground biomass. Using this database, develop general models estimating both biomass attributes can be remotely sensed specifically diameter. We show that jointly quantify find single equation predicts two variables across world's forests. These provide an intuitive integrating remote imagery into large-scale monitoring programmes will key importance parameterizing next dynamic vegetation models.

Language: Английский

Citations

344
Tree Species Classification in Boreal Forests With Hyperspectral Data DOI
Michele Dalponte, Hans Ole Ørka, Terje Gobakken

et al.

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2012, Volume and Issue: 51(5), P. 2632 - 2645

Published: Oct. 15, 2012

Tree species mapping in forest areas is an important topic inventory. In recent years, several studies have been carried out using different types of hyperspectral sensors under various conditions. The aim this work was to evaluate the potential two high spectral and spatial resolution (HySpex-VNIR 1600 HySpex-SWIR 320i), operating at wavelengths, for tree classification boreal forests. To address objective, many experiments were out, taking into consideration: 1) three classifiers (support vector machines (SVM), random (RF), Gaussian maximum likelihood); 2) resolutions (1.5 m 0.4 pixel sizes); 3) subsets bands (all a selection); 4) levels (pixel levels). study area characterized by presence four classes Norway spruce, Scots pine, together with scattered Birch other broadleaves. Our results showed that: HySpex VNIR sensor effective kappa accuracies over 0.8 (with Pine Spruce reaching producer's higher than 95%); role 320i limited, its alone are able properly separate only species; has strong effect on accuracy (an overall decrease more 20% between 1.5 resolution); there no significant difference SVM or RF classifiers.

Language: Английский

Citations

340
A framework for mapping tree species combining hyperspectral and LiDAR data: Role of selected classifiers and sensor across three spatial scales DOI Creative Commons
Aniruddha Ghosh, Fabian Ewald Fassnacht, P. K. Joshi

et al.

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2013, Volume and Issue: 26, P. 49 - 63

Published: June 28, 2013

Language: Английский

Citations

320