Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tile Compression and Embeddings for Multi-Label Classification in GeoLifeCLEF 2024

Anthony Miyaguchi, Patcharapong Aphiwetsa, Mark McDuffie

TL;DR

This work tackles multi-label plant species classification in GeoLifeCLEF 2024 by combining tile-based remote-sensing inputs compressed with 2D Discrete Cosine Transform and a neighborhood-aware, geolocation-driven approach. It evaluates Nearest Neighbor, Tile CNN, and Tile2Vec pipelines, leveraging ASL, BCE, Hill, and sigmoidF1 losses, with a self-supervised Tile2Vec objective built on triplet loss $L(t_a, t_n, t_d)=\left[||f_{\theta}(t_a)-f_{\theta}(t_n)||_2-||f_{\theta}(t_a)-f_{\theta}(t_d)||_2+m\right]_+$ and an ASL component to balance positives and negatives. The pipeline benefits from LSH-based approximate nearest neighbors (50 km cutoff), 128×128 tiles, and DCT-based feature reduction, achieving a public-leaderboard score of 0.161 with geolocation-fronted predictions, while post-competition top submissions reached around 0.409, illustrating substantial potential in geospatial cues and learned tile embeddings. Source code and models are publicly available at https://github.com/dsgt-kaggle-clef/geolifeclef-2024, underscoring the approach’s reproducibility and practical relevance for scalable biodiversity monitoring.

Abstract

We explore methods to solve the multi-label classification task posed by the GeoLifeCLEF 2024 competition with the DS@GT team, which aims to predict the presence and absence of plant species at specific locations using spatial and temporal remote sensing data. Our approach uses frequency-domain coefficients via the Discrete Cosine Transform (DCT) to compress and pre-compute the raw input data for convolutional neural networks. We also investigate nearest neighborhood models via locality-sensitive hashing (LSH) for prediction and to aid in the self-supervised contrastive learning of embeddings through tile2vec. Our best competition model utilized geolocation features with a leaderboard score of 0.152 and a best post-competition score of 0.161. Source code and models are available at https://github.com/dsgt-kaggle-clef/geolifeclef-2024.

Tile Compression and Embeddings for Multi-Label Classification in GeoLifeCLEF 2024

TL;DR

This work tackles multi-label plant species classification in GeoLifeCLEF 2024 by combining tile-based remote-sensing inputs compressed with 2D Discrete Cosine Transform and a neighborhood-aware, geolocation-driven approach. It evaluates Nearest Neighbor, Tile CNN, and Tile2Vec pipelines, leveraging ASL, BCE, Hill, and sigmoidF1 losses, with a self-supervised Tile2Vec objective built on triplet loss and an ASL component to balance positives and negatives. The pipeline benefits from LSH-based approximate nearest neighbors (50 km cutoff), 128×128 tiles, and DCT-based feature reduction, achieving a public-leaderboard score of 0.161 with geolocation-fronted predictions, while post-competition top submissions reached around 0.409, illustrating substantial potential in geospatial cues and learned tile embeddings. Source code and models are publicly available at https://github.com/dsgt-kaggle-clef/geolifeclef-2024, underscoring the approach’s reproducibility and practical relevance for scalable biodiversity monitoring.

Abstract

We explore methods to solve the multi-label classification task posed by the GeoLifeCLEF 2024 competition with the DS@GT team, which aims to predict the presence and absence of plant species at specific locations using spatial and temporal remote sensing data. Our approach uses frequency-domain coefficients via the Discrete Cosine Transform (DCT) to compress and pre-compute the raw input data for convolutional neural networks. We also investigate nearest neighborhood models via locality-sensitive hashing (LSH) for prediction and to aid in the self-supervised contrastive learning of embeddings through tile2vec. Our best competition model utilized geolocation features with a leaderboard score of 0.152 and a best post-competition score of 0.161. Source code and models are available at https://github.com/dsgt-kaggle-clef/geolifeclef-2024.
Paper Structure (32 sections, 8 equations, 7 figures, 4 tables)

This paper contains 32 sections, 8 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Overview
  4. Processing Pipeline
  5. Satellite and Raster Data
  6. Tile Compression via Discrete Cosine Transform (DCT)
  7. Time-Series Data
  8. Data Augmentation
  9. Locality-Sensitive Hashing for Nearest Neighbor Queries
  10. Models
  11. Nearest Neighbor Model
  12. Geolocation Model
  13. Tile CNN Model
  14. Tile2Vec Model
  15. Model Evaluation and Loss Functions
  16. ...and 17 more sections

Figures (7)

  • Figure 1: GeoJSON polygon definition.
  • Figure 2: Polygon region overlayed on a map.
  • Figure 3: Overview of the data and modeling pipeline. Raw data is pre-processed to maintain survey site per row semantics, with 2D DCT coefficients as features. Data is cached as columnar parquet files in cloud storage for efficient access.
  • Figure 4: Example of a tiled raster image. The image is a 128x128 tile of the RGB-NIR satellite imagery. The image is associated with a survey site and is used as input to the model.
  • Figure 5: Example of low-pass filtering using the DCT. (a) Original bio1 raster image. (b) Low-pass filter using the first 50 coefficients of the 1D-DCT, reshaping on the first axis (row-major order). (c) Low-pass filter using the 2D-DCT using the top-left 8x8 coefficients.
  • ...and 2 more figures