Contrastive Representation Learning for Natural World Imagery: Habitat prediction for 30, 000 species

Recent work in contrastive representation learning has pushed the boundaries of classification tasks in computer vision, achieving state of the art results on many established benchmarks. However, their performance on natural imagery tasks which fall into the category of fine-grained image classification can be further improved. In this paper, I present a methodology that explores this issue and achieves state of the art results on species distribution modelling from remote sensing imagery as part of the GeoLifeCLEF2021 challenge. My method is able to beat the current state of the art on this challenge (trained on 4 types of imagery) using only base RGB imagery. Initial experiments indicate that modifying the architecture to include additional image modalities leads to further improvements in performance on the task of location-based species recommendation. Additionally, I introduce a consistency function, which relies on the strategy of withholding data from the model and is useful in checking for model generality without relying on a validation split.

[1]  Sachith Hasaranga Seneviratne.  Automatic Code Generation for Statistical Models with Augmentation and Collapsing , 2021 .

[2]  Maximilien Servajean,et al.  How citizen scientists contribute to monitor protected areas thanks to automatic plant identification tools , 2020, Ecological Solutions and Evidence.

[3]  Sanka Rasnayaka,et al.  Multi-Dataset Benchmarks for Masked Identification using Contrastive Representation Learning , 2021, 2021 Digital Image Computing: Techniques and Applications (DICTA).

[4]  Moorthy Gokula Krishnan,et al.  Impact of Pretrained Networks For Snake Species Classification , 2020, CLEF.

[5]  Maximilien Servajean,et al.  Convolutional neural networks improve species distribution modelling by capturing the spatial structure of the environment , 2021, PLoS Comput. Biol..

[6]  Xavier Giro-i-Nieto,et al.  Seasonal Contrast: Unsupervised Pre-Training from Uncurated Remote Sensing Data , 2021, ArXiv.

[7]  Vladimir Risojevic,et al.  Self-Supervised Learning of Remote Sensing Scene Representations Using Contrastive Multiview Coding , 2021, 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).

[8]  Quoc V. Le,et al.  Randaugment: Practical automated data augmentation with a reduced search space , 2019, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).

[9]  J. Koenderink.  Q… , 2014, Les noms officiels des communes de Wallonie, de Bruxelles-Capitale et de la communaute germanophone.

[10]  이화영.  X , 1960, Chinese Plants Names Index 2000-2009.

[11]  Serge Belongie,et al.  Benchmarking Representation Learning for Natural World Image Collections , 2021, 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

[12]  Stefano Ermon,et al.  Geography-Aware Self-Supervised Learning , 2020, ArXiv.

[13]  Christophe Botella,et al.  Overview of LifeCLEF Location-based Species Prediction Task 2020 (GeoLifeCLEF) , 2020, Conference and Labs of the Evaluation Forum.

[14]  Xuan Yang,et al.  When Does Contrastive Visual Representation Learning Work? , 2021, ArXiv.

[15]  Geoffrey E. Hinton,et al.  ImageNet classification with deep convolutional neural networks , 2012, Commun. ACM.

[16]  Christophe Botella,et al.  The GeoLifeCLEF 2020 Dataset , 2020, ArXiv.

[17]  Stefan Kahl,et al.  Overview of LifeCLEF 2020: A System-Oriented Evaluation of Automated Species Identification and Species Distribution Prediction , 2020, CLEF.

[18]  Maximilien Servajean,et al.  Overview of GeoLifeCLEF 2021: Predicting species distribution from 2 million remote sensing images , 2021, CLEF.

[19]  G. Tiwari,et al.  Land use, transport, and population health: estimating the health benefits of compact cities , 2016, The Lancet.