Predicting woodland bird species habitat with multi-temporal and multisensor remote sensing data

Abstract Remote sensing data capture ecologically important information that can be used to characterize, model and predict bird habitat. This study implements fusion techniques using Random Forests (RF) with spectral Landsat data and structural airborne laser scanning (ALS) data to scale habitat attributes through time and to characterize habitat for four bird species in dynamic young forest environments in the United Kingdom. We use multi-temporal (2000, 2005, 2012/13, 2015) multi-sensor (Landsat and ALS) data to (i) predict structural attributes via pixel-level fusion at 30 metre spatial resolution, (ii) model bird habitat via object-level fusion and compare with models based on ALS, Landsat and predicted structural attributes, and (iii) predict bird habitat through time (i.e., predict 2015 habitat based on 2000-2012 data). First, we found that models predicting mean height from spectral information had the highest accuracy, whilst maximum height, standard deviation of heights, foliage height diversity, canopy cover and canopy relief ratio had good accuracy, and entropy had low accuracy. The green band and the normalized burn ratio (NBR) were consistently important for prediction, with the red and shortwave infrared (SWIR) 1 bands also important. For all structural variables, high values were underpredicted and low values were overpredicted. Second, for Blue Tit ( Cyanistes caeruleus ) and Chaffinch ( Fringilla coelebs ), the most accurate model employed Landsat data, while object-level fusion performed best for Chiffchaff ( Phylloscopus collybita ) and Willow Warbler ( Phylloscopus trochilus ). ALS mean, maximum and standard deviation of heights and Landsat tasseled cap transformations (TCT) (i.e., wetness, greenness and brightness) were ranked as important to all species across various models. Third, we used our models to predict presence in 2015 and implemented a spatial intersection approach to assess the predictive accuracy for each species. Blue Tit and Willow Warbler presences were well predicted with the Landsat, ALS, and objectlevel fusion models. Chaffinch and Chiffchaff presences were best predicted with the ALS model. Predictions based on pixel-level predicted structure surfaces had low accuracy but were acceptable for Chaffinch and Willow Warbler. This study is significant as it provides guidance for Landsat and ALS data application and fusion in habitat modelling. Our results highlight the need to use appropriate remote sensing data for each study species based on their ecology. Object-level data fusion improved habitat characterization for all species relative to ALS, but not to Landsat for Blue Tit and Chaffinch. Pixel-level fusion for predicting structural attributes in years where ALS data are note available is increasingly being used in modelling but may not adequately represent within-patch wildlife habitat. Finally, incorporating predicted surfaces generated through pixel-level fusion in our habitat models yielded low accuracy. Highlights We used object- and pixel-level fusion with ALS and Landsat to examine bird habitat Pixel-level fusion predicted surfaces yielded low accuracy in habitat models Best models: Landsat (Blue Tit, Chaffinch); fusion (Chiffchaff, Willow Warbler) Best prediction: ALS (Chaffinch, Chiffchaff) Best prediction: ALS, Landsat, object-level fusion (Blue Tit, Willow Warbler) Graphical abstract

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.