Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FrogDeepSDM: Improving Frog Counting and Occurrence Prediction Using Multimodal Data and Pseudo-Absence Imputation

Chirag Padubidri, Pranesh Velmurugan, Andreas Lanitis, Andreas Kamilaris

TL;DR

This study tackles frog species distribution modeling under sparse, imbalanced citizen-science data by introducing a late-fusion multimodal architecture that combines high-resolution imagery with TerraClimate covariates. It couples adaptive oversampling and a custom loss with a pseudo-absence imputation strategy to generate reliable non-detections, achieving strong results in both presence/absence classification and frog counting. The approach dramatically reduces counting error from about $MAE$ = $189$ to $29$ and achieves an AUC of $0.90$ with accuracy of $84.9\%$ on validation, while demonstrating cross-regional generalization. These findings advance scalable biodiversity monitoring and provide actionable insights for amphibian conservation and habitat prioritization.

Abstract

Monitoring species distribution is vital for conservation efforts, enabling the assessment of environmental impacts and the development of effective preservation strategies. Traditional data collection methods, including citizen science, offer valuable insights but remain limited in coverage and completeness. Species Distribution Modelling (SDM) helps address these gaps by using occurrence data and environmental variables to predict species presence across large regions. In this study, we enhance SDM accuracy for frogs (Anura) by applying deep learning and data imputation techniques using data from the "EY - 2022 Biodiversity Challenge." Our experiments show that data balancing significantly improved model performance, reducing the Mean Absolute Error (MAE) from 189 to 29 in frog counting tasks. Feature selection identified key environmental factors influencing occurrence, optimizing inputs while maintaining predictive accuracy. The multimodal ensemble model, integrating land cover, NDVI, and other environmental inputs, outperformed individual models and showed robust generalization across unseen regions. The fusion of image and tabular data improved both frog counting and habitat classification, achieving 84.9% accuracy with an AUC of 0.90. This study highlights the potential of multimodal learning and data preprocessing techniques such as balancing and imputation to improve predictive ecological modeling when data are sparse or incomplete, contributing to more precise and scalable biodiversity monitoring.

FrogDeepSDM: Improving Frog Counting and Occurrence Prediction Using Multimodal Data and Pseudo-Absence Imputation

TL;DR

This study tackles frog species distribution modeling under sparse, imbalanced citizen-science data by introducing a late-fusion multimodal architecture that combines high-resolution imagery with TerraClimate covariates. It couples adaptive oversampling and a custom loss with a pseudo-absence imputation strategy to generate reliable non-detections, achieving strong results in both presence/absence classification and frog counting. The approach dramatically reduces counting error from about = to and achieves an AUC of with accuracy of on validation, while demonstrating cross-regional generalization. These findings advance scalable biodiversity monitoring and provide actionable insights for amphibian conservation and habitat prioritization.

Abstract

Monitoring species distribution is vital for conservation efforts, enabling the assessment of environmental impacts and the development of effective preservation strategies. Traditional data collection methods, including citizen science, offer valuable insights but remain limited in coverage and completeness. Species Distribution Modelling (SDM) helps address these gaps by using occurrence data and environmental variables to predict species presence across large regions. In this study, we enhance SDM accuracy for frogs (Anura) by applying deep learning and data imputation techniques using data from the "EY - 2022 Biodiversity Challenge." Our experiments show that data balancing significantly improved model performance, reducing the Mean Absolute Error (MAE) from 189 to 29 in frog counting tasks. Feature selection identified key environmental factors influencing occurrence, optimizing inputs while maintaining predictive accuracy. The multimodal ensemble model, integrating land cover, NDVI, and other environmental inputs, outperformed individual models and showed robust generalization across unseen regions. The fusion of image and tabular data improved both frog counting and habitat classification, achieving 84.9% accuracy with an AUC of 0.90. This study highlights the potential of multimodal learning and data preprocessing techniques such as balancing and imputation to improve predictive ecological modeling when data are sparse or incomplete, contributing to more precise and scalable biodiversity monitoring.
Paper Structure (37 sections, 7 equations, 5 figures, 8 tables, 2 algorithms)

This paper contains 37 sections, 7 equations, 5 figures, 8 tables, 2 algorithms.

Table of Contents

  1. Background
  2. Related Work
  3. Statistical Methods
  4. Machine Learning Methods
  5. Multimodal Deep Learning Methods
  6. Our Contribution
  7. Methodology
  8. Dataset
  9. Environmental Covariates (Predictor Variables/Covariates)
  10. Frog Occurrence Dataset (Target Variable)
  11. Data Preprocessing
  12. Grid Creation
  13. Data balancing using Adaptive Oversampling
  14. Data Balancing using Custom Loss Function
  15. Data Balancing using Log Transformation
  16. ...and 22 more sections

Figures (5)

  • Figure 1: (a) Grid creation for calculating frog density across three countries using 30 km² cells. (b,c,d) Frog presence points in Costa Rica, Australia, and South Africa.
  • Figure 2: Histogram of Unbalanced, Balanced and Log transformed frequency of frog occurrence.
  • Figure 3: Architectures of the multimodal learning models for frog presence/absence classification and frog counting regression, presented together for clarity. Both models combine highresolution satellite images and TerraClimate data as inputs. The classification model uses a sigmoid activation function to predict frog presence/absence, while the regression model employs a ReLU activation function to output frog count predictions.
  • Figure 4: (a) Train and Test MAE curves for Imbalanced data, (b) Train and Test MAE curves for balanced data
  • Figure 5: Comparison of MAE for model trained on different input data source combinations and the ensemble model