PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning for conifer/deciduous classification of airborne LiDAR 3D point clouds representing individual trees

Hamid Hamraz, Nathan B. Jacobs, Marco A. Contreras, Chase H. Clark

arXiv:1802.08872v1cs.LGcs.CV

TL;DR

Expert-designed features for LiDAR point clouds may be suboptimal because the data are difficult for people to interpret visually. The paper compares CNN-compatible representations of individual airborne-LiDAR tree crowns, finding that 2D views can yield similar or better accuracy with less processing cost, while performance differs between overstory and understory trees.

  • Problem

    Expert-selected features may be suboptimal for LiDAR point clouds because the data are difficult for people to interpret visually.

    Source: §1 · p. 3

  • Method

    The study compares CNN-based crown representations and uses ensembles trained on balanced data subsamples to assess classification and correct likely mislabels.

    Source: §4 · p. 22

  • Results

    2D views/profiles can yield similar or higher accuracy than bulkier representations with lower computational load; overstory accuracy was ~90%, versus ~90% for deciduous and ~65% for coniferous understory trees.

    Source: §4 · pp. 22–23

  • Takeaways & Limitations

    The results indicate deep learning can classify individual airborne-LiDAR crowns effectively and efficiently, a step toward operational tree-level remote quantification of large-scale forests.

    Source: §4 · p. 23

    Rotational augmentation is unlikely to match the classification quality that a higher number of real training instances could provide.

    Source: §3.4 · p. 19

Abstract

from arXiv · show

The purpose of this study was to investigate the use of deep learning for coniferous/deciduous classification of individual trees from airborne LiDAR data. To enable efficient processing by a deep convolutional neural network (CNN), we designed two discrete representations using leaf-off and leaf-on LiDAR data: a digital surface model with four channels (DSMx4) and a set of four 2D views (4x2D). A training dataset of labeled tree crowns was generated via segmentation of tree crowns, followed by co-registration with field data. Potential mislabels due to GPS error or tree leaning were corrected using a statistical ensemble filtering procedure. Because the training data was heavily unbalanced (~8% conifers), we trained an ensemble of CNNs on random balanced sub-samples of augmented data (180 rotational variations per instance). The 4x2D representation yielded similar classification accuracies to the DSMx4 representation (~82% coniferous and ~90% deciduous) while converging faster. The data augmentation improved the classification accuracies, but more real training instances (especially coniferous) likely results in much stronger improvements. Leaf-off LiDAR data were the primary source of useful information, which is likely due to the perennial nature of coniferous foliage. LiDAR intensity values also proved to be useful, but normalization yielded no significant improvements. Lastly, the classification accuracies of overstory trees (~90%) were more balanced than those of understory trees (~90% deciduous and ~65% coniferous), which is likely due to the incomplete capture of understory tree crowns via airborne LiDAR. Automatic derivation of optimal features via deep learning provide the opportunity for remarkable improvements in prediction tasks where captured data are not friendly to human visual system - likely yielding sub-optimal human-designed features.

1 Introduction

The introduction motivates deep learning for tree classification from airborne LiDAR because expert-designed features may miss useful information in point clouds that are difficult to interpret visually. The study addresses the limited prior work on airborne-LiDAR tree classification by engineering CNN-ready crown representations and investigating training and network design choices.

1 Introduction

  • Traditional machine-learning approaches depend on expert-selected features that may be suboptimal because LiDAR point clouds are difficult for people to interpret visually.The introduction contrasts this with deep learning’s end-to-end automatic feature extraction and mapping to predictions.Source: §1 · pp. 3–4
  • Deep learning on 3D data requires preprocessing to make point clouds usable, while voxel representations can be computationally expensive and prone to overfitting.The introduction also identifies DSMs and multiple 2D views as alternative discrete representations of 3D data.Source: §1 · p. 4
  • Earlier deep-learning tree-classification studies described here used ground-based LiDAR, leaving airborne-LiDAR classification of individual trees unaddressed.Airborne LiDAR provides information at a larger scale and from a different perspective than ground-based LiDAR.Source: §1 · p. 5
  • The study addresses the identified lack of deep-learning studies on individual-tree classification from airborne LiDAR by targeting crowns in a natural, dense forest.Its stated contributions are two CNN-ready representations and experiments on training-data preparation, network structure, data composition, and domain-specific data.Source: §1 · p. 5

2.1 Study site, LiDAR campaign, and field survey

The study uses Robinson Forest in eastern Kentucky, combining leaf-off and leaf-on airborne LiDAR collections with field measurements from plots distributed across the forest. The field survey supplies tree attributes and stem locations for linking segmented crowns to measured trees.

2.1 Study site, LiDAR campaign, and field survey

  • Robinson Forest is a diverse, contiguous mixed mesophytic forest in eastern Kentucky, with steep terrain, about 93% average canopy cover, and roughly 330 stems per hectare.The forest includes a range of deciduous overstory and understory species, with eastern hemlock and pine among the conifers.Source: §2.1 · pp. 5–6
  • Two airborne LiDAR datasets were collected in 2013: a low-density leaf-off survey at about 2 pt/m² and a high-density leaf-on survey at about 50 pt/m².The datasets used the same LiDAR system and were processed into ground and non-ground points; ground returns supported a 1-meter DEM.Source: §2.1 · pp. 6–7
  • Field crews surveyed 271 plots and recorded tree attributes and stem locations; among 3,987 trees taller than 4 m, 7.27% were conifers.The survey included DBH, height, species, crown class, status, and stem class for qualifying trees.Source: §2.1 · p. 7

2.2 Data preparation

The data-preparation pipeline normalizes selected leaf-on intensity values, segments and registers individual crowns against field measurements, and converts crowns into two CNN-compatible image formats. The resulting dataset is augmented through rotational variants of each crown.

2.2.1 LiDAR intensity normalization

  • Leaf-on intensity varied significantly with range and emission angle, so those effects were removed by residualization; corresponding correlations were not significant in leaf-off data.For leaf-on datasets, log range correlated negatively with intensity and the cosine of the angle correlated positively, both at P < .0001.Source: §2.2.1 · pp. 8–9

2.2.2 Individual tree segmentation and registration with field data

  • The pipeline segmented canopy layers with at least 3 pt/m² and matched crowns to field stems using height and leaning-angle scores, then selected the maximum-score one-to-one assignment.It produced 2,528 co-registered trees: 124 conifers and 2,404 deciduous trees.Source: §2.2.2 · pp. 9–10

2.2.3 Discretization of segmented point clouds

  • Each crown was represented as either a four-channel DSM or four 2D images, using leaf-on and leaf-off data, and each point cloud generated 180 rotational variants.The DSM channels encode leaf-on and leaf-off elevation and intensity; the 2D format combines aerial and side-profile views, with height and crown width supplied separately.Source: §2.2.3 · pp. 10–11

2.3 Convolutional Neural Network models

The CNN models process either the four-channel DSM representation or four single-channel 2D views, with additional tree-shape features. The architectures differ in when they fuse information from the inputs, trading model richness against computational cost.

2.3 Convolutional Neural Network models

  • The DSM×4 network processes a 4×128×128 input through six convolution-and-pooling pairs before combining image features with crown area.The resulting image representation is flattened to 16 units and combined with the crown-area features before the final dense layers and softmax.Source: §2.3 · pp. 11–12
  • The 4×2D model applies five convolution-and-pooling pairs to each of four 1×64×64 images, then combines their features with crown width and tree height.The image branches contribute 16 flattened units, and the two additional features contribute two units before the final dense layers and softmax.Source: §2.3 · p. 12
  • The DSM×4 model fuses leaf-on and leaf-off data, intensity, and height early, whereas 4×2D delays fusion until each view has produced features.The DSM approach supports a richer training model, while 4×2D incurs less computational cost.Source: §2.3 · p. 12

2.4 Mislabel correction via iterative resampling

Because GPS error and tree leaning can misregister field stems and LiDAR crowns, the study uses iterative ensemble filtering to identify and correct likely label errors. The procedure trains networks on balanced random samples and applies statistical tests to their predictions until no further mislabels are identified.

2.4 Mislabel correction via iterative resampling

  • GPS error and tree leaning can misregister field stems to LiDAR crowns, creating labels that assign a coniferous stem to a deciduous crown or vice versa.The paper notes that plot-center GPS error (~5 m) can exceed distances between individual trees.Source: §2.4 · pp. 12–13
  • The procedure trains 100 networks on balanced random samples of 80 deciduous and 80 conifer instances, distributing instances nearly equally across networks.Each network is trained for three epochs to raise accuracy above the 50% base without reaching overfitting.Source: §2.4 · p. 13
  • For each instance, the method compares its average held-out accuracy across augmented forms with the expected accuracy for a correctly labeled instance.A lower average than the symmetric training-accuracy threshold indicates a likely mislabel.Source: §2.4 · p. 14
  • T-tests on network-level accuracy differences identify likely mislabels; labels are flipped and the training-and-testing cycle repeats until none are flagged.The significance level is 10^-8 to control false positives across 2,528 tests per iteration.Source: §2.4 · p. 14

2.5 Classification and evaluation

Classification uses ensembles trained on balanced random samples, with held-out softmax probabilities averaged to classify each instance. The study evaluates cross-validated accuracy across training-data sizes, augmentation levels, domain-data variants, crown classes, and point densities.

2.5 Classification and evaluation

  • For each classification ensemble, 50 networks are trained on random balanced samples of 100 deciduous and 100 conifer instances, and their held-out softmax probabilities are averaged for prediction.The same ensemble procedure is applied to DSM×4 and 4×2D representations; 4×2D uses five epochs rather than fifteen and is chosen for subsequent experiments because of its lower computational load.Source: §2.5 · pp. 14–15
  • Cross-validated classification is repeated on stratified samples containing 20% to 100% of the dataset, with resampling-instance sizes adjusted to each sample.The ensemble-network count remains constant across dataset sizes.Source: §2.5 · p. 15
  • The domain-data experiments exclude leaf-off data, leaf-on data, or intensity values, and compare normalized with non-normalized leaf-on intensities.When leaf-on or leaf-off data are excluded, the final two dense layers are reduced to 16 and 8 units.Source: §2.5 · p. 15
  • The study examines whether crown point density correlates with the correct-label softmax probability and compares classification performance for overstory and understory trees.Overstory includes dominant and co-dominant trees; understory includes intermediate and overtopped trees.Source: §2.5 · p. 15

3 Results and Discussion

The results show that 4×2D achieves accuracy similar to DSM×4 with lower computational load, while mislabel correction, more training data, and augmentation improve the classification pipeline. Performance varies by data source and crown class, and point density showed no significant correlation with accuracy.

3.1 Mislabel correction

  • Mislabel correction converged after 13 iterations, flipping 35 of 124 initial conifers and 125 of 2,404 initial deciduous labels; the conifer share rose from 4.90% to 8.46%.Average training accuracy across 100 networks rose from 67.1% to a plateau of 83.6%.Source: §3.1 · pp. 15–16, Figure 1 · p. 16

3.2 Classification accuracy

  • Cross-validated DSM×4 accuracies were 80.4±5.3% for conifers and 90.1±1.3% for deciduous trees; 4×2D achieved 82.7±5.1% and 90.2±1.3%, respectively, with similar accuracy at lower computational load.The lower-load result was reported for a late-fused network.Source: §3.2 · p. 17

3.3 Effect of training data size on the classification

  • Deciduous accuracy plateaued using 40% of the dataset (~925 trees), while conifer accuracy continued increasing beyond the original 214 conifer instances.The authors suggest that a balanced dataset near one thousand instances per class could bring conifer accuracy closer to deciduous accuracy.Source: §3.3 · p. 18, Figure 3 · p. 18

3.4 Effect of data augmentation on classification

  • 73.8% conifer and 87.7% deciduous accuracy with 20 rotations were lower than with 180 rotations; gains plateaued at ~150 rotations for conifers and ~60 for deciduous trees.The authors caution that augmentation is unlikely to match the classification quality of additional real training instances.Source: §3.4 · pp. 18–19, Figure 4 · p. 19

3.5 Effect of domain data on classification

  • Using binary presence values instead of intensity reduced conifer accuracy from 82.7% to 69.2%; intensity normalization yielded only minor changes when leaf-off data were excluded.The authors report that LiDAR intensity values were useful, while normalization yielded no significant improvement.Source: §3.5 · p. 20
  • Excluding leaf-off data reduced conifer accuracy from 82.7% to 61.2%, while excluding leaf-on data reduced it from 82.7% to 81.6%; deciduous accuracy changed only slightly.The authors conclude that leaf-off data provided the most useful features, despite its much lower point density.Source: §3.5 · p. 19, Figure 5 · p. 21

3.6 Effects of crown class and point density on classification

  • Overstory accuracy was 92.1±4.7% for conifers and 87.2±2.2% for deciduous trees; understory accuracy was 69.0±9.8% and 92.1±1.4%, respectively.The authors attribute the less balanced understory results likely to partial crown capture; they found no significant point-density correlation with accuracy.Source: §3.6 · p. 21, Figure 6 · p. 22

4 Conclusions

Four 2D views offer a more efficient alternative to bulkier representations, while training-data composition, LiDAR season, and crown visibility shape classification outcomes.

4 Conclusions

  • The study segmented individual airborne LiDAR trees, registered them with field-surveyed trees, and classified their crowns using CNNs and two discrete representations.Ensemble learning was used for mislabel correction and cross-validated accuracy measurement.Source: §4 · p. 22
  • Four 2D views of a crown point cloud are more efficient to process and can match or exceed bulkier 2.5D or 3D representations.Late feature fusion may also match early fusion with less computational load.Source: §4 · p. 22
  • More real training instances can yield much stronger improvements than data augmentation alone.Augmentation can improve classification accuracy, but the conclusions favor increasing real instances.Source: §4 · p. 23
  • Leaf-off LiDAR was the main source of useful information, and intensity was useful, but normalizing intensity produced no significant improvement.The authors associate leaf-off usefulness with the perennial nature of conifer foliage.Source: §4 · p. 23
  • Overstory accuracy was about 90% and more balanced than understory accuracy, which was about 90% for deciduous and 65% for coniferous trees.The authors associate the understory difference with airborne LiDAR capturing only partial crown shapes.Source: §4 · p. 23
  • Richer datasets and more complicated tasks, such as species classification, remain needed, while deep learning can automatically extract task-relevant features.The authors highlight potential relevance where data are not friendly to human perception and expert-designed features may be suboptimal.Source: §4 · p. 23

Related papers

Loading 1802.08872v1…