Advancing Spatial Pattern Integration in Species Distribution Modelling With Computer Vision: How Grain Size and Extent Concepts Shape Predictions of Emerging Aquatic Insects
Abstract
1. Capturing spatial patterns in ecology necessitates jointly accounting for the concepts of grain size (i.e., the spatial resolution of the input data) and extent (i.e., the area covered by the input data). Species distribution models (SDMs) that rely on traditional machine learning models (e.g., Random Forest, MaxEnt) only integrate the grain size through environmental variables derived from raster data. Recent advances in deep learning have led to the development of Deep SDMs. Unlike traditional SDMs, Deep SDMs directly process image windows representing landscape context. Therefore, they can manipulate both the grain size and extent, through the spatial resolution and size of the windows, respectively. 2. This study evaluates how grain size and window extent affect Deep SDMs, and how grain size alters the importance of variables (i.e., pixels of the image windows). We used multispectral and LiDAR data from drone and satellite acquisitions, with spatial resolutions ranging from 31 cm to 10 m, and varied the window extent between 40 m and 100 m. Grad-CAM was used to assess the importance of variables for each data type and measure their spatial agreement.As a case study, the terrestrial distribution of Plecoptera and Trichoptera, two orders of emerging aquatic insects with contrasting dispersal capacities, was predicted. 3. Results demonstrated the capacity of Deep SDMs to integrate both the grain size and extent. They highlighted substantial variations in performance depending on the grain size (up to 0.26 AUC) and on the extent (up to 0.16 AUC). In particular, the best predictions were obtained using 2.5 m multispectral data at a 40 m extent for Plecoptera (limited dispersers), and the same multispectral resolution combined with LiDAR at a 100 m extent for Trichoptera (greater dispersers). Results also showed a moderate but notable change in variable importance with different grain sizes. 4. Given the growing availability of remote sensing products, our findings advocate for systematic analysis of grain size and extent with Deep SDMs. This is essential to efficiently capture spatio-ecological patterns and to better understand the studied species. Datasets and code are publicly available.
Related
- Advancing grapevine phenology predictions: a comparison of regression-based statistical learning approach and conventional models in mediterranean environment.
- Advancing Marine Mammal Conservation Through Omics Tools and Integration of Indigenous Ecological Knowledge
- Advancing Human Skin Equivalents: The Crucial Role of Neurovascular Integration
- Computer Vision for Cattle Health and Welfare Monitoring: A Comprehensive Review of Methods, Applications, and Interdisciplinary Integration in Smart Agriculture
- Advancing specific local wine identification through integration of ICP-MS mineral wine profiles and Machine Learning: A case study on discrimination among the ten Beaujolais crus
BibTeX
@misc{brosse2026advancing,
title = {{Advancing Spatial Pattern Integration in Species Distribution Modelling With Computer Vision: How Grain Size and Extent Concepts Shape Predictions of Emerging Aquatic Insects}},
author = {Augustin de la Brosse and Aimi Okabayashi and Thomas Corpetti and Thomas Houet},
howpublished = {HAL},
year = {2026},
month = oct,
doi = {10.64898/2026.10.29.762482},
url = {https://hal.science/hal-05844564v1},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Advancing Spatial Pattern Integration in Species Distribution Modelling With Computer Vision: How Grain Size and Extent Concepts Shape Predictions of Emerging Aquatic Insects
AU - Augustin de la Brosse
AU - Aimi Okabayashi
AU - Thomas Corpetti
AU - Thomas Houet
PY - 2026
DA - 2026/10/29
PB - HAL
DO - 10.64898/2026.10.29.762482
UR - https://hal.science/hal-05844564v1
AB - 1. Capturing spatial patterns in ecology necessitates jointly accounting for the concepts of grain size (i.e., the spatial resolution of the input data) and extent (i.e., the area covered by the input data). Species distribution models (SDMs) that rely on traditional machine learning models (e.g., Random Forest, MaxEnt) only integrate the grain size through environmental variables derived from raster data. Recent advances in deep learning have led to the development of Deep SDMs. Unlike traditional SDMs, Deep SDMs directly process image windows representing landscape context. Therefore, they can manipulate both the grain size and extent, through the spatial resolution and size of the windows, respectively. 2. This study evaluates how grain size and window extent affect Deep SDMs, and how grain size alters the importance of variables (i.e., pixels of the image windows). We used multispectral and LiDAR data from drone and satellite acquisitions, with spatial resolutions ranging from 31 cm to 10 m, and varied the window extent between 40 m and 100 m. Grad-CAM was used to assess the importance of variables for each data type and measure their spatial agreement.As a case study, the terrestrial distribution of Plecoptera and Trichoptera, two orders of emerging aquatic insects with contrasting dispersal capacities, was predicted. 3. Results demonstrated the capacity of Deep SDMs to integrate both the grain size and extent. They highlighted substantial variations in performance depending on the grain size (up to 0.26 AUC) and on the extent (up to 0.16 AUC). In particular, the best predictions were obtained using 2.5 m multispectral data at a 40 m extent for Plecoptera (limited dispersers), and the same multispectral resolution combined with LiDAR at a 100 m extent for Trichoptera (greater dispersers). Results also showed a moderate but notable change in variable importance with different grain sizes. 4. Given the growing availability of remote sensing products, our findings advocate for systematic analysis of grain size and extent with Deep SDMs. This is essential to efficiently capture spatio-ecological patterns and to better understand the studied species. Datasets and code are publicly available.
ER -