This dataset contains bird call detection data, rainfal, and satellite-derived vegetation metrics collected to examine arid zone bird community responses to resource pulses. Data were collected at Fowlers Gap Arid Zone Research Station, Australia, over two years using passive acoustic monitoring coupled with satellite remote sensing. The dataset demonstrates the integration of high temporal resolution passive acoustic monitoring with satellite remote sensing at matching temporal scales - an approach with significant potential for scaling ecological monitoring in remote, temporally variable ecosystems. Data are provided in analysis-ready format suitable for examining community-level responses to environmental variability, developing predictive models of species activity patterns, or adapting similar methodological approaches to other arid systems. Files include processed bird detection data, monthly rainfall, and vegetation metrics.

Acoustic Data Collection Continuous audio recordings were collected using six SM4 song meters deployed at Fowlers Gap Arid Zone Research Station, Australia, from April 2022 to August 2024. All devices recorded for 60 minutes before and after sunrise and sunset daily using manufacturer default settings, with recording windows adjusted seasonally based on changing sunrise/sunset times.

Satellite Vegetation Data Monthly fractional vegetation cover was derived from MODIS imagery (500m × 500m resolution) covering 195 pixels across the study area downloaded from https://eo-data.csiro.au/remotesensing/v310/australia/monthly/cover/. Data processing used pre-packaged MODIS products from CSIRO, specifically the Nadir BRDF-Adjusted Reflectance product (MCD43A4) collection 6. Monthly composites were generated from 8-day composites using the multi-dimensional median (medoid) for each pixel. Fractional cover of photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV) was calculated using linear spectral unmixing methods validated for Australian conditions.

Bird Call Classification A custom bird classifier was developed using BirdNET v2.4 as the foundational model through iterative transfer learning. The model underwent 21 successive rounds of fine-tuning using locally sourced call libraries, progressively adapting to site-specific acoustic characteristics and species-location combinations. Each audio clip was transformed into numerical embeddings capturing acoustic features (pitch, rhythm, harmonics, temporal dynamics).

Model Validation and Threshold Selection Classifier performance was validated using 8,253 one-hour recordings, with each 3-second call segment assigned a confidence score (0-1). For each species, 200 randomly selected segments were manually validated as true or false positives. Precision-confidence curves were generated at 0.01 intervals across the full confidence range. Species-specific confidence thresholds were set to achieve ≥0.7 precision rates. Species not achieving this threshold at any confidence level were excluded from analysis. Only calls exceeding species-specific confidence thresholds were retained in the final dataset.

Data Processing Monthly call rates were calculated as the proportion of days per month with species detections. 

NA Values This data set includes PV and NPV measurements taken up to 12 months prior to the commencement of acoustic monitoring to allow for calculating lagged coeffients between bird call rates and environmnetal covariates. Becasue there was no bird call data during this period, all cells excluding the year_month, rainfall.mm, meanGreen (PV), and meanDead(NPV) have NA values.