Remote sensing‐based landscape indicators for the evaluation of threatened‐bird habitats in a tropical forest

Abstract Avian species persistence in a forest patch is strongly related to the degree of isolation and size of a forest patch and the vegetation structure within a patch and its matrix are important predictors of bird habitat suitability. A combination of space‐borne optical (Landsat), ALOS‐PALSAR (radar), and airborne Light Detection and Ranging (LiDAR) data was used for assessing variation in forest structure across forest patches that had undergone different levels of forest degradation in a logged forest—agricultural landscape in Southern Laos. The efficacy of different remote sensing (RS) data sources in distinguishing forest patches that had different seizes, configurations, and vegetation structure was examined. These data were found to be sensitive to the varying levels of degradation of the different patch categories. Additionally, the role of local scale forest structure variables (characterized using the different RS data and patch area) and landscape variables (characterized by distance from different forest patches) in influencing habitat preferences of International Union for Conservation of Nature (IUCN) Red listed birds found in the study area was examined. A machine learning algorithm, MaxEnt, was used in conjunction with these data and field collected geographical locations of the avian species to identify the factors influencing habitat preference of the different bird species and their suitable habitats. Results show that distance from different forest patches played a more important role in influencing habitat suitability for the different avian species than local scale factors related to vegetation structure and health. In addition to distance from forest patches, LiDAR‐derived forest structure and Landsat‐derived spectral variables were important determinants of avian habitat preference. The models derived using MaxEnt were used to create an overall habitat suitability map (HSM) which mapped the most suitable habitat patches for sustaining all the avian species. This work also provides insight that retention of forest patches, including degraded and isolated forest patches in addition to large contiguous forest patches, can facilitate bird species retention within tropical agricultural landscapes. It also demonstrates the effective use of RS data in distinguishing between forests that have undergone varying levels of degradation and identifying the habitat preferences of different bird species. Practical conservation management planning endeavors can use such data for both landscape scale monitoring and habitat mapping.

landscape in Southern Laos. The efficacy of different remote sensing (RS) data sources in distinguishing forest patches that had different seizes, configurations, and vegetation structure was examined. These data were found to be sensitive to the varying levels of degradation of the different patch categories. Additionally, the role of local scale forest structure variables (characterized using the different RS data and patch area) and landscape variables (characterized by distance from different forest patches) in influencing habitat preferences of International Union for Conservation of Nature (IUCN) Red listed birds found in the study area was examined. A machine learning algorithm, MaxEnt, was used in conjunction with these data and field collected geographical locations of the avian species to identify the factors influencing habitat preference of the different bird species and their suitable habitats. Results show that distance from different forest patches played a more important role in influencing habitat suitability for the different avian species than local scale factors related to vegetation structure and health. In addition to distance from forest patches, LiDARderived forest structure and Landsat-derived spectral variables were important determinants of avian habitat preference. The models derived using MaxEnt were used to create an overall habitat suitability map (HSM) which mapped the most suitable habitat patches for sustaining all the avian species. This work also provides insight that retention of forest patches, including degraded and isolated forest patches in addition to large contiguous forest patches, can facilitate bird species retention within tropical agricultural landscapes. It also demonstrates the effective use of RS data in distinguishing between forests that have undergone varying levels of degradation and identifying the habitat preferences of different bird species. Practical conservation management planning endeavors can use such data for both landscape scale monitoring and habitat mapping.

| INTRODUCTION
Remote sensing (RS) data have been widely applied for studying forest structure and its impact on bird species persistence (Eldegard et al., 2014). Optical imagery from Landsat has often been used to estimate vegetation properties such as greenness and its influence on avian persistence (Ranganathan, Krishnaswamy, & Anand, 2010). Landsatderived variables, such as greenness, allow us to quantify how vegetation properties vary with different levels of degradation. This makes them useful for monitoring forest transformation and degradation, including varying stages of forest degradation in human modified tropical ecosystems (Cattau, 2010;Wiens et al., 2009).
In addition to optical imagery, airborne Light Detection and Ranging (LiDAR) is another method used for monitoring forest structure dynamics and their impact on species. Although the use of LiDAR has been known to be costly (Hummel, Hudak, Uebler, Falkowski, & Megown, 2011), LiDAR provides more detailed information on forest structure such as tree heights (Shimizu et al., 2014).
The three-dimensional vegetation structure has been mapped using LiDAR data and was found to be associated with the presence and persistence of animal species in a given ecosystem (Davies & Asner, 2014). Moreover, LiDAR-based measures of canopy structure and complexity were shown to be key predictors of the breeding habitat of a migrant Neotropical bird species in the United States (Goetz et al., 2010). Space-borne radar provides another source of forest structural information at a global scale. A Radar Forest Degradation Index (RFDI) derived from these data was successful at detecting degraded forests (Mitchard et al., 2011). However, RFDI as yet has not been used to quantify avian habitat preferences in tropical forests. This study has used LiDAR, Landsat, and ALOS-PALSAR RFDI to quantify their ability to distinguish between forest patches of different sizes that have undergone varying levels of degradation in a tropical forest-agricultural ecosystem in southern Laos. These RS variables were used in conjunction with a landscape level variable (distance from forest patches) to identify suitable habitat for endangered avian species. The study aimed to demonstrate the efficacy of different RS data and tools to monitor the inert-related problems of forest degradation and habitat identification for endangered species is currently facing with the view of contributing to practical conservation management.
Southeast Asia's biodiversity faces an impending crisis. Over the past several decades, the region has the highest rates of deforestation in the world and it could lose three quarters of its forests by 2100 (Sodhi, Koh, Brook, & Ng, 2004). An analysis using a 500-m resolution MODIS dataset from 2000 to 2011 in the Lower Mekong region revealed that Vietnam lost 1.1% of forests per annum, followed by Cambodia at 0.7% and Laos at 0.4% (Leinenkugel, Wolters, Oppelt, & Kuenzer, 2015). Such a rapid rate of deforestation could result in a large-scale extinction of different species, including avian species.
Given the high level of endemic species in the region, regional losses could translate into global extinctions (Sodhi et al., 2004).
The causes of forest loss vary across SE Asia. One of the leading causes of tropical forest degradation and biodiversity loss is selective logging; the practice of removing high value timber species. It has been considered as a more sustainable alternative to clear felling wherein all trees in the area are cut uniformly. Owing to high habitat specificity, many avian guilds need large patches of contiguous and near intact forests, and so decline in selectively logged forests (Tobias, Şekercioğlu, & Vargas, 2013). A meta-analysis of the response of tropical and subtropical bird species to anthropogenic disturbances revealed that these species were generally less likely to occur in logged/disturbed forests (Newbold et al., 2014). Frugivorous birds are particularly affected by selective logging and higher logging intensities (Tobias, 2015). However, species richness of birds in intact primary forest, once-logged and twice-logged forest in Borneo were found to be similar and the International Union for Conservation of Nature (IUCN) threatened birds (listed as Near-Threatened, Vulnerable or Endangered species) did not decline after the second round of logging. Hence, it is important to examine the ability of degraded and human-modified tropical ecosystems to provide habitat to endangered avian species.
This study explores the use of space-borne optical, radar, and airborne LiDAR datasets to quantify forest structural properties and to model the distribution of bird species in a human-modified tropical landscape in Laos, the two conservation issues that concern the wider tropical Asia as well. We envisage that a combination of RS variables is needed to characterize the different properties of the humanmodified landscapes and their effect on habitat suitability of different bird species.
In addition to local scale factors (characterized by RS derived vegetation parameters and patch area), landscape scale factors in form of distance from forest patches have been included as this may influence species persistence in an agricultural landscape (Angelieri, Adams-Hosking, de Barros, de Souza, & McAlpine, 2016). Using these inputs, we developed a combined habitat suitability map (HSM) for the bird species under consideration using a MaxEnt (maximum entropy) modeling approach (Elith et al., 2011). It has been suggested that landscape scale factors play an important role in influencing avian habitat preferences in temperate forests (Zhao, Azeria, Le Blanc, Lemaître, & Fortin, 2013). Landscape scale factors have been identified as being significant drivers of avian species habitat preferences in anthropogenically disturbed landscapes (De Barros, de Siqueira, Alexandrino, Da Luz, & Do Couto, 2012) and are known to be more important than local scale forest structure variables (Zhao et al., 2013). These findings make it important to quantify the role of both local and landscape scale factors in influencing avian habitat preferences in human-modified tropical ecosystems such as those found in Laos and elsewhere in SE Asia.

K E Y W O R D S
forest patches, habitat suitability, IUCN Red listed birds, Laos, production forest, remote sensing Evaluation of factors that influence habitat preference and identifying suitable habitat patches is important for practical conservation management. This research has the following objectives: (a) Evaluate how RS metrics pertaining to forest structure and degradation vary across forest patches that have undergone varying levels of degradation; (b) identify the most important local scale factor that contribute to habitat preferences of endangered birds; and (c) examine the role of a landscape level factor: distance from different patch types in influencing habitat suitability. It is expected that the results of (a) will help identify the extent to which different RS data can distinguish forest patches that have undergone varying levels of degradation and if they concur with the qualitative categorizations made on the ground. Further, (b) and (c) will help quantify the role of both local and landscape scale factors in influencing the habitat preferences of different avian species and help identify suitable habitat patches.

| Description of study site and bird surveys
Our study site is a 25,000-ha estate called Dongsithouane (Figure 1 (B) small forest patches with large canopy gaps, thick understory, and presence of many small trees. These were mostly found in the vicinity of villages (mean size: 0.38 km 2 ); (C) small patches of closed canopy forest. The presence of some understory and several small trees was detected in these. These were located within woodlands or scrublands F I G U R E 1 Dongsithouane production forest area located in Savannakhet Province of Laos (With Different Patch categories). Image © 2015 DigitalGlobe that had been heavily exploited for timber (mean size: 0.11 km 2 ); and (D) highly degraded forest patches with an open canopy and scrub development (mean size: 0.31 km 2 ). These different forest patches were then manually digitized on Google Earth.
Bird surveys were carried out in December 2015 with the assistance of a local forestry official in at least three patches from each of the categories, spanning as much of the production forest as possible, using a point count method. This method is useful for getting a spatially representative sample from each patch (Whitworth, Newman, Mundkur, & Harris, 2007). The surveys were carried out from 6 to 10 a.m. every morning with 200-m interval between each of the points, to avoid double counting (dos Anjos et al., 2015;Mollon, 2010). This was performed to ensure we obtain spatially well distributed geo-location data. About 15 min were spent at each point observing bird species, and birds seen or heard within a 20-m vicinity of the observation point were recorded. Laos has a problem of unexploded landmines from the wars of 1960s and 1970s (Lao Rehabilitation Foundation, 2010). The impact on the study area is not known but the suspected presence of landmines has been given high attention so as not to impede data collection activities. The bird species presence data was recorded for IUCN Red listed threatened bird species observed in the area (Table 1) (Table 1) inhabit woodlands and forests but also use the surrounding agricultural matrix for feeding and transverse the landscape. For instance, parakeets are known to inhabit a variety of forested habitats but also forage in agricultural areas for food (Thewlis, Timmins, Evans, & Duckworth, 1998). Same for the endangered green peafowl which inhabits a variety of forest habitats and can also use agricultural edges Hence, in a human-modified tropical forest ecosystem, it is vital we identify the local and landscape scale factors that may influence avian species persistence.

| Airborne LiDAR survey
The airborne LiDAR data were obtained in February 2009, using a Piper PA-31 Najavo aircraft at an altitude of 2,000 m and a flying speed on 19 knots. Leica Airborne Laser Scanner (ALS) was used with a 30-m field of view (FOV), delivering a point density of 1 point/m 2 and two returns (first and last) only. TerraScan was used for data preprocessing and initial point classification (Hou, Xu, & Tokola, 2011).
Area-based LiDAR metrics were derived using LAS tools (Isenburg, M., 2014) for the vertical and horizontal vegetation structure using a 20 m by 20 m (see Table 2). Vertical distribution ratio (VDR) also helps quantify the subcanopy variation in the tree canopies (Hill & Hinsley, 2015).
The avian geo-location data were collected in 2015. The forest change is assumed to be low as no forest clearance activity has occurred in the study area; hence, it is appropriate to compare these data. Earlier works suggest that a time lag of 6 years or more between LiDAR and field data collection has no impact on patterns of avian distribution in relatively undisturbed forests (Hill & Hinsley, 2015;Vierling, Swift, Hudak, Vogeler, & Vierling, 2014).

| Radar imagery (ALOS-PALSAR)
ALOS-PALSAR data collected in 2010 with a spatial resolution of 25 m were downloaded from the JAXA website (JAXA, 2010). These were provided in a color-composite form and had dual polarization HH (horizontal transmit, horizontal receive) and HV (horizontal transmit, vertical receive). These data are provided as digital numbers (DN).
In order to obtain the backscatter values, the DN values of both HH and HV were converted to the normalized radar cross section (σ 0 ) as follows (Avtar, Suzuki, Takeuchi, & Sawada, 2013): Here, CF is a calibration factor (−83 in case of ALOS-PALSAR data).
An enhanced Lee filter was applied to reduce speckle, while retaining the image texture using ENVI. HH and HV values were used to calculate the Radar Forest Degradation Index (RFDI), a radar-derived measure of forest degradation (Mitchard et al., 2011): The backscatter values are strongly associated with the forest structural components, orientation, and factors such as canopy cover (Mitchard et al., 2011). Hence, when the canopy opens up (as a consequence of logging), this causes RFDI to go up. Completely cleared areas have RFDI value of 1, while undisturbed forests have RFDI values ranging from 0.3 to 0.4 (Saatchi, 2007).
(1) σ 0 = 10 ⋅ log 10 (DN 2 ) + CF (2) RFDI = HH − HV HH + HV ClasLite implements sensor-specific offset and gains to convert the raw DN values to radiance (expressed in terms of watts/m 2 /solid angle). These were in turn put through the 6S atmospheric correction module provided by ClasLite to produce top of the atmosphere surface reflectance (Bryan et al., 2013). Tasselled cap transformation was implemented on these atmospherically corrected surface reflectance data. A tasselled-cap transformation is a technique that converts the original bands of Landsat data (blue, green, red, NIR, SWIR1, SWIR2) into a new set of variables (Kauth & Thomas, 1976). Tasselled cap transformations of the reflectance data (using ENVI) reduced data from seven to three ecologically significant bands (i.e., brightness, greenness, and wetness) (Ranganathan et al., 2010). Brightness is a measure of soil brightness, greenness is associated with chlorophyll content and canopy density, while wetness is related to water content of the vegetation (Foster, Townsend, & Zganjar, 2008).
Kruskal-Wallis tests were carried out to examine how the vegetation properties, derived using LiDAR, Landsat, and ALOS PALSAR, varies among the four patch categories. Using the "pgirmess" package in R, post hoc analysis was conducted to identify which patches were significantly different.

| Species and habitat suitability maps using MaxEnt
MaxEnt is a machine learning-based technique which uses species presence data in conjunction with environmental layers for modeling species distribution (Elith et al., 2011). The  The previously described RS variables were used as environmental layers (Huang et al. 2014). These variables represent the local scale forest structure-related variables that influence habitat suitability. Additionally, Euclidean distances to the four different patch categories (A, B, C, and D) were included as landscape level variables. In its rasterized format, the value of each pixel is its Euclidean distance from its nearest forest patch edge. Two separate models were developed using MaxEnt: one with a combination of local and landscape scale variables, and another one with local scale variables only. It was expected that the former will help us understand the role distance from forest patches can play in influencing habitat preference in a mixed ecosystem like ours, and the latter is expected to help identify the local scale factors that influence habitat suitability.
The modeling was performed using MaxEnt software (version 3.3.k) (Princeton, 2015). Regularization was implemented to penalizing the model in proportion to the coefficient magnitude, thus preventing overfitting (Merow, Smith, & Silander, 2013). Further, 1,000 iterations were carried out to allow for convergence, thereby reducing the chance of over or underprediction (Frake, 2014).
Variable importance estimation was carried out using both the heuristic test and the jackknife procedure; the latter helps to evaluate the independent contribution of variables to the habitat suitability model (Padalia, Srivastava, & Kushwaha, 2014). It should be noted that 75% of the training data were used for model training and 25% T A B L E 2 Area-based LiDAR metrics and their ecological significance

Name of the LiDAR metric Ecological significance References
Max. height 2D representation of the tallest trees in the landscape.
Average height 2D representation of the average tree height in the landscape.
Height percentiles (p10, p20, p50, p90) Captures the percentile distribution of laser returns within the canopy. Height percentiles were calculated from 5% to 95% with the view of representing the multilayered structure of the forest canopy.
Ediriweera (2013) Standard deviation Represents variation in canopy structure Isenburg (2014) Canopy cover Number of first returns above a given height cutoff is divided by the total number of first returns. It measures the extent of ground covered with vegetation.
Davies and Asner (2014), Isenburg (2014) Vertical distribution ratio (VDR) Vertical distribution ratio (VDR) quantifies the vertical distribution of foliage in the canopy. Forest patches with a closed canopy and reduced understory have a low VDR Goetz, Steinberg, Dubayah, and Blair (2007) data were employed for model testing (Lahoz-Monfort et al., 2010).
The area under the curve (AUC) was also computed for evaluating model performance using the test data. The AUC of a model is a threshold independent measure equivalent to the probability that the model will rank a randomly chosen species presence site higher than a randomly chosen absence site (Liu, White, & Newell, 2011

| Variation in vegetation structure across forest patches
Four patch types of varying sizes, spatial configuration, and degradation level categories (A-D) were evaluated as part of this research.
Vertical vegetation parameters (derived from LiDAR) varied significantly among the four patch categories (Figure 2a (Table 4).

| Species and habitat suitability mapping for avian species
The species habitats were also found to moderately overlap with Schoener's D index ranging from 0.1 to <0.7 (Table 5). A D index of 0.8 and above would indicate a significant overlap of species habitats.
Jackknife test of variable importance also revealed that distance of different patches had more information by themselves as compared to local forest structure variables (see Figures S1-S10). The individual SDMs produced by this model (also presented in the Figures S1-S10) with test AUC values >0.7 were retained for producing a cumulative HSM, as shown in Figure 4.  (Table 6).

| Factors affecting bird species habitat suitability
Several factors have been identified by MaxEnt modeling as influencing IUCN Red listed threatened bird species, including LiDAR-derived variables of vertical and horizontal canopy structure. Research by Tattoni, Rizzolli, and Pedrini (2012) indicates that inclusion of LiDAR variables improves the predictive ability of MaxEnt-based habitat suitability models (Tattoni et al., 2012). Variables related to forest T A B L E 3 Percentage contributions of local scale factors to habitat suitability of different species  in human-modified forest ecosystems. In case of certain species such as the ashy headed green pigeon, alexandrine parakeet, and blossomheaded parakeet, the tasselled cap metrics were found to contain some useful information. Certain tasselled cap metrics such as wetness are strongly linked with the landscape scale canopy cover and the persistence of bird species in forested landscapes (Ranganathan et al., 2010) and arboreal mammal species (Turner et al., 2012). These metrics are important determinants of richness of forest dependent bird species and allow for finer distinction between different habitats as compared to measures such as canopy cover (Ranganathan, Chan, & Daily, 2007). These findings are further corroborated by Eldegard et al. (2014) who discovered that inclusion of spectral variables along with LiDAR improved the performance of models predicting avian species persistence in a boreal forest. Further, the size of patches was discovered to be an important variable for influencing habitat suitability.
It has been discovered that in addition to forest structure within the patch, forest patch size is an important determinant of habitat preference, especially for species such as the ashy headed green pigeon and white rumped pygmy falcon. This finding is consistent with a previous study (Tobias et al., 2013) which indicates that patch area can influence the persistence of certain species (Davies & Asner, 2014). To the best of our knowledge, RFDI has not been previously employed to study avian habitat preferences. While RFDI is sensitive to variations in forest degradation across different forest patches, its contribution to explaining avian habitat preference was minimal.
The research has identified that different bird species have different local habitat structure preferences and limited habitat overlap (Whitmore, 1975). This information indicates different habitat preferences of the different avian species and can help customize conservation programs to suit the needs of the bird species in question. When landscape and local scale variables were modeled together, the former were discovered to be the more important variables for habitat suitability in almost all the cases. This is consistent with the existing body of research. Distance from forest patches was an important determinant avian persistence in the fragmented woodlands of England (Hinsley, Bellamy, Newton, & Sparks, 1995). It has further been established that mobile taxa such as mammals and birds can use agricultural landscapes for feeding and dispersal, but need forest patches for their continued persistence (Anand, Krishnaswamy, & Das, 2008;Angelieri et al., 2016). Indeed, in the study area, different avian species were observed foraging in the agricultural fields. However, this research has established the importance of both forest patch and landscape variables in facilitating avian species persistence and identified which of these are the most important for different species persistence in a tropical forest-agricultural matrix.
The presence of a contiguous forest cover is an important determinant of bird assemblages (Graham & Blake, 2001). In this research, distance from the large patches was found to be very important for the ashy headed green pigeon. Additionally, varying levels of canopy closure are also known to support bird species diversity (Gil-Tena, Saura, & Brotons, 2007). This arguably explains why the distance from moderately disturbed, noncontinuous forest patches are an important variable of habitat suitability for many bird species such as the alexandrine parakeet and the blossom headed parakeet. These find-  (Edwards et al., 2011).
Our research has also demonstrated that MaxEnt is a suitable tool for modeling distribution of multiple species (and identifying factors that affect habitat suitability) using a combination of field collected data, multiple RS data for characterizing forest structure, and landscape variables. Indeed, MaxEnt has been used for modeling species distribution and identifying factors influencing habitat suitability in a number of tropical forest ecosystems (Lahoz-Monfort et al., 2010;Nazeri et al., 2012;Wilting et al., 2010). Owing to its ability to produce useful results with a very small presence data, MaxEnt has proven itself to be useful for modeling the distribution of rare and endangered species (Rebelo & Jones, 2010;Tinoco et al., 2009

| Avian persistence in human-modified landscapes and proposed conservation action
As human pressure on natural ecosystems grows, it is important to evaluate the ability of the matrix surrounding nature reserves, agroecosystems, human-modified landscapes, agricultural fields, and plantations to provide habitat and feeding/foraging space to avian species (Petit, Petit, Christian, & Powell, 1999). The most suitable habitats are isolated patches located within a sea of marginal to unsuitable habitat matrix (see Figure 4). This is to be expected, given the magnitude of human modification of the landscape and conversion to paddy fields (Azman et al., 2011). Conversion to monoculture systems, such as rice paddies, brings about a decline in bird species abundance (Azman et al., 2011). Further, increased fragmentation and isolation of forest patches also leads to a decline in avian species abundances (Haddad et al., 2015). Rice paddy agricultural systems have been a part of tropical Asia's ecosystems for more than a millennium.  (2013) and Flaspohler et al. (2010) indicates that smaller forest patches can play an important role in avian species conservation (see Table 4). The research has also identified that forest structure and spectral variables (derived from the LiDAR and Landsat data, respectively) are important for influencing individual avian species habitat preferences. The forest patches are isolated in a sea of unsuitable habitat (see Figure 4). This could increase the risk of "extinction debt," the possibility of species going extinct after a time lag (Cowlishaw, 1999). However, ecological restoration of forest ecosystems is a stated goal of the Sustainable Forestry and Rural Development (SUFORD) program. Conservation action could focus on retaining and enhancing the forest structure and health variables that were identified in order to inform future ecological restoration programs, along with preserving the suitable habitats mapped. Restoration of marginal habitats can yield significant conservation benefits for avian species (Reid, Mendenhall, Rosales, Zahawi, & Holl, 2014).
Our research was conducted in a production forest in Laos. The tropical forest-agricultural matrix found here is a microcosm of tropical landscapes in the majority of the Greater Mekong region. Hence, it is important to devise strategies that can help with biodiversity retention in human-modified landscapes and allow for landscape scale monitoring of forest degradation. The role large contiguous forest patches can play in bird species conservation has been well established, and it has been recommended that retention of large forest patches can help maintain a diverse species assemblage in human-modified ecosystems (Fischer, Lindenmayer, & Manning, 2006). In addition to this, the research also makes a case for retaining small, degraded, isolated forest patches within agricultural landscapes. Previous research indicates that the presence of different kinds of forest patches helped improve the diversity of butterfly assemblages and small, isolated forest patches can contribute to landscape scale diversity (Benedick et al., 2006

| CONCLUSION
This research has demonstrated how a combination of optical, radar, and LiDAR data may be used for assessing forest degradation and variation in forest structure across varying disturbance regimes/patch configurations. This is of potential importance in postconflict countries such as Laos which are rich in tropical forests. These countries face an ongoing problem of the presence of unexploded landmines and shells, which makes the collection of field data extremely hazardous. As this work demonstrates, RS data from different sources can be employed for rapid conservation monitoring by facilitating landscape scale identification of forest structure patterns and suitable habitats.
Data from multiple sensors are useful for identifying the factors that influence avian habitat preference in a tropical forest-agricultural matrix ecosystem. Our findings make a case for promoting the retention of forest patches that includes degraded forest patches within an agricultural ecosystem for bird species conservation. Coomes for his support and guidance in improving the manuscript.

CONFLICTS OF INTEREST
The authors have no conflicts of interest to declare.

AUTHOR CONTRIBUTIONS
MS collected the field data, analyzed the data, and wrote the manuscript. TT, ZH, and CN provided suggestions on manuscript improvement. TT and ZH provided information about the study site and CN advised on ALOS PALSAR data processing