Integrating Multi-Sensors Data for Species Distribution Mapping Using Deep Learning and Envelope Models

Abstract

The integration of ecological and atmospheric characteristics for biodiversity management is fundamental for long-term ecosystem conservation and drafting forest management strategies, especially in the current era of climate change. The explicit modelling of regional ecological responses and their impact on individual species is a significant prerequisite for any adaptation strategy. The present study focuses on predicting the regional distribution of Rhododendron arboreum, a medicinal plant species found in the Himalayan region. Advanced Species Distribution Models (SDM) based on the principle of predefined hypothesis, namely BIOCLIM, was used to model the potential distribution of Rhododendron arboreum. This hypothesis tends to vary with the change in locations, and thus, robust models are required to establish nonlinear complex relations between the input parameters. To address this nonlinear relation, a class of deep neural networks, Convolutional Neural Network (CNN) architecture is proposed, designed, and tested, which eventually gave much better accuracy than the BIOCLIM model. Both of the models were given 16 input parameters, including ecological and atmospheric variables, which were statistically resampled and were then utilized in establishing the linear and nonlinear relationship to better fit the occurrence scenarios of the species. The input parameters were mostly acquired from the recent satellite missions, including MODIS, Sentinel-2, Sentinel-5p, the Shuttle Radar Topography Mission (SRTM), and ECOSTRESS. The performance across all the thresholds was evaluated using the value of the Area Under Curve (AUC) evaluation metrics. The AUC value was found to be 0.917 with CNN, whereas it was 0.68 with BIOCLIM, respectively. The performance evaluation metrics indicate the superiority of CNN for species distribution over BIOCLIM.

Forest tree species distribution for Europe 2000–2020: mapping potential and realized distributions using spatiotemporal machine learning

“…As any ML-derived product, our predictions would benefit from having more and better quality data on tree species, in particular those that come from NFI plots: it is now crucial to have such data freely available to monitor processes such as species compositional changes, niche shifts, forest regrowth and degradation, as recently stated by Nabuurs et al (2022) . Exploring more sophisticated and different ML algorithms such as Deep Learning (DL) techniques ( Lakshminarayanan, Pritzel & Blundell, 2016 ) for our ensemble framework is also another area of improvement given the wide variety of applications these methods possess and the results obtained in comparison with other conventional ML algorithms ( Choe, Chi & Thorne, 2021 ; Deneu et al, 2021 ; Anand et al, 2021 ).…”

Forest tree species distribution for Europe 2000-2020: mapping potential and realized distributions using spatiotemporal Machine Learning

“…As any ML-derived product, our predictions would benefit from having more and better quality data on tree species, in particular those that come from NFI plots: it is now crucial to have such data freely available to monitor processes such as species compositional changes, niche shifts, forest regrowth and degradation, as recently stated by Nabuurs et al (2022). Exploring more sophisticated and different ML algorithms such as Deep Learning (DL) techniques (Lakshminarayanan et al, 2016) to our ensemble framework is also another area of improvement given the wide variety of applications these methods possess and the results obtained in comparison with other conventional ML algorithms (Choe et al, 2021;Deneu et al, 2021;Anand et al, 2021).…”

Incorporating environmental time series into species distribution models

“…Our results extend the findings from previous studies (Botella et al . 2018; Benkendorf & Hawkins 2020; Anand et al . 2021; Capinha et al .…”

“…Recently, Capinha et al (2021) demonstrated that some deep learning architectures allow using spatial time series data directly as predictors of ecological phenomena, hence overcoming the need of using temporally unvarying, pre-assembled, predictors sets. In addition, SDM practitioners have been also turning to deep learning (Anand et al . 2021), but as an alternative algorithms to use in conventional workflows with static predictors (e.g., Botella et al .…”