Remote sensing tree classification with a multilayer perceptron

Abstract

To accelerate scientific progress on remote tree classification—as well as biodiversity and ecology sampling—The National Institute of Science and Technology created a community-based competition where scientists were invited to contribute informatics methods for classifying tree species and genus using crown-level images of trees. We predicted tree species and genus at the pixel level using hyperspectral and LIDAR observations. We compared three algorithms that have been implemented extensively across a broad range of research applications: support vector machines, random forests, and multilayer perceptron. At the pixel level, the multilayer perceptron algorithm predicted species or genus with high accuracy (92.7 and 95.9%, respectively) on the training data and performed better than the other algorithms (85.8-93.5%). This indicates promise for the use of the MLP algorithm for tree-species classification and coincides with a growing body of research in which neural network-based algorithms outperform other types of classification algorithms for machine vision. To aggregate patterns across the images, we used an ensemble approach that averages the pixel-level outputs of the MLP algorithm to predict species at the crown level. The accuracy of these predictions on the test set was 68.8% for species.

A data science challenge for converting airborne remote sensing data into ecological information

“…We had the most participants in this task (six): BRG (Sumsion et al, 2018), Conor (McMahon, 2018), FEM (Dalponte, Frizzera & Gianelle, 2018), GatorSense (Zou, Gader & Zare, 2018), StanforCCB (Anderson, 2018), and our baseline system (Fig. 6).…”

“…The class (species) of the tree in the test set was assigned by using the same label of the individual tree with highest likelihood. The BRG group used a neural network multi-layer perceptron on the hyperspectral images (Sumsion et al, 2018). Crown probabilities were aggregated by averaging the pixel scale predicted probabilities.…”

“…We present the details of the tasks and data, and summarize and synthesize the results from the participants. In a set of short accompanying papers and preprints the participants describe the methods used (Table 1), present detailed results for those methods, and discuss lessons learned and future directions for these methods (Anderson, 2018; Dalponte, Frizzera & Gianelle, 2018; Taylor, 2018; McMahon, 2018; Sumsion et al, 2018; Zou, Gader & Zare, 2018). Finally, we discuss the broad potential for competitions in ecology and the biological sciences more generally.…”

Hyperspectral tree crown classification using the multiple instance adaptive cosine estimator

“…One of these challenges was to perform species classification of individual trees given airborne hyperspectral data. The challenge provided competitors (Anderson, 2018; Sumsion et al, 2018; Dalponte, Frizzera & Gianelle, 2018) with training and testing hyperspectral signatures extracted from individual tree crowns in the National Ecological Observatory Network (NEON) hyperspectral data collected at the Ordway-Swisher Biological Station in north-central Florida. These signatures were extracted from the imagery and labeled by the competition organizers by generating individual tree crown polygons using a tablet computer, GIS software, and an external GPS device in the field as described by Marconi et al (2018).…”

Peer Review #1 of "Hyperspectral tree crown classification using the multiple instance adaptive cosine estimator (v0.1)"

“…One of these challenges was to perform species classification of individual trees given airborne hyperspectral data. The challenge provided competitors (Anderson, 2018;Sumsion et al, 2018;Dalponte et al, 2018) with training and testing hyperspectral signatures extracted from individual tree crowns in the National Ecological Observatory Network (NEON) hyperspectral data collected at the Ordway-Swisher Biological Station in north-central Florida. These signatures were extracted from the imagery and labeled by the competition organizers by generating individual tree crown polygons using a tablet computer, GIS software, and an external GPS device in the field as described by Marconi et al (2018).…”