Predicting butterfly species presence from satellite imagery using soft contrastive regularisation

TL;DR

This work addresses the challenge of predicting multi-species presence from satellite imagery by introducing the S2-BMS dataset, which links Sentinel-2 4-band imagery to UK Butterfly Monitoring Scheme observations across the UK. It proposes a ResNet-18-based predictor for 62 butterfly species with a 256-d embedding, trained under a binary cross-entropy objective, and augments learning with a novel soft supervised contrastive regularisation, Paired Embeddings Contrastive Loss (PECL), that uses squared cosine similarities between label vectors to form soft positive pairs. The key contributions are (i) the public release of S2-BMS as a benchmark for satellite-to-biodiversity prediction in a new region and taxon, and (ii) the PECL loss that improves top-$k$ predictive accuracy by incorporating semantic label relationships into contrastive learning. Empirically, the best model achieves around 70% top-10 accuracy and 63% top-5 accuracy, with PECL yielding modest yet robust gains, demonstrating the viability of remote-sensing-based biodiversity monitoring and suggesting directions for future improvements via transfer learning and richer geospatial context.”

Abstract

The growing demand for scalable biodiversity monitoring methods has fuelled interest in remote sensing data, due to its widespread availability and extensive coverage. Traditionally, the application of remote sensing to biodiversity research has focused on mapping and monitoring habitats, but with increasing availability of large-scale citizen-science wildlife observation data, recent methods have started to explore predicting multi-species presence directly from satellite images. This paper presents a new data set for predicting butterfly species presence from satellite data in the United Kingdom. We experimentally optimise a Resnet-based model to predict multi-species presence from 4-band satellite images, and find that this model especially outperforms the mean rate baseline for locations with high species biodiversity. To improve performance, we develop a soft, supervised contrastive regularisation loss that is tailored to probabilistic labels (such as species-presence data), and demonstrate that this improves prediction accuracy. In summary, our new data set and contrastive regularisation method contribute to the open challenge of accurately predicting species biodiversity from remote sensing data, which is key for efficient biodiversity monitoring.

Figures (3)

• Figure 1: Properties of S2-BMS data set.a) Locations (1329 in total) of S2-BMS data set in the UK. b) Number of days that locations were visited over a 2-year period. c) (Log) distribution of probability presence per species. d) Distribution of cosine similarity (blue) and cosine similarity squared (orange, used in \ref{['eq:loss_softk_pecl']}) of species vectors $(\mathbf{y}_i, \mathbf{y}_{j \neq i})$. Total number of pairs is $N (N - 1) / 2 = 882$k. e) Example locations of extracted 256 $\times$ 256 pixels S2 images (only RGB shown) and their species-presence vector (right side of each satellite image). f) Cosine similarity matrix between the example locations of e).
• Figure 2: Model schematic. Top: the prediction task (\ref{['eq:loss_bce_pred']}). Bottom: PECL loss function (\ref{['eq:loss_softk_pecl']}).
• Figure 3: Performance across locations and species, quantified using the MSE improvement factor ($f_{\text{MSE}}$ across locations and species. a) Map of $f_{\text{MSE}}$ values for each test location (values greater than 2 shown as 2), across the UK. b)$f_{\text{MSE}}$ as a function of mean encounter rate per species. c)$f_{\text{MSE}}$ as a function of number of species present per location (i.e., number of species ever observed per location).