The first release of JWST-SPRING NIRCam science images will cover a total area of approximately 1200 square arcminutes in the F444W band, with MIRI science images covering an area of around 400 square arcminutes in the F770 band, across all five CANDELS fields (COSMOS, EGS, GOODS-N, GOODS-S & UDS). This figure presents the color images of the first-released NIRCam mosaics. 

The data processing procedure is briefly summarized as follows:

We initiate our data reduction by processing the uncalibrated raw images from individual exposures using the Stage 1 pipeline, Detector1Pipeline, with its default configuration. This stage performs essential detector-level corrections, including a series of initial adjustments such as group scale correction, data quality initialization, superbias subtraction, reference pixel correction, linearity correction, persistence correction, dark current subtraction, jump detection, and gain scale correction. In particular, we have adopted an improved method similar to that described by Bagley et al. (2023) for identifying "snowballs"—unusual cosmic ray events—and expanding their footprints. We then re-run the ramp-fitting step to generate count-rate maps that exclude the regions labeled as snowballs. 

We proceed to execute the Stage 2 pipeline, Image2Pipeline, on the count-rate images using default parameters. This critical stage includes WCS assignment, flat-fielding, and photometric calibration, resulting in fully calibrated individual exposures. Subsequently, we apply a series of bespoke corrections designed to mitigate specific features. These include addressing scattered-light effects, such as wisps and claws, managing 1/f noise, and implementing additional masking to refine the data quality. We implemented masks to address various features that left imprints on the mosaics, including persistence effects, dragon breath, ginkgo leaf patterns, uncorrected wisp features, and other artifacts. To ensure comprehensive coverage, we conducted a visual examination of all Stage 2 images to identify any residual artifacts. Additionally, we identified certain moving targets from the images and applied masking accordingly. 

We then execute the Stage 3 pipeline, Image3Pipeline, to generate a unified mosaic for each filter of every observation. This process integrates all calibrated images from various dither positions and detectors. During this stage, we perform astrometric alignment to ensure precise spatial correspondence, background matching for visual consistency, outlier detection to maintain data integrity, and resampling of images onto a unified output grid for seamless integration.

Finally, a customized background subtraction routine is re-applied to the individual mosaics before the co-added mosaics are created. 

To prevent over-subtraction of the background, we conducted a meticulous sky background estimation and subtraction process. Initially, we followed the method outlined by Bagley et al. (2023) to implement a four-tiered source detection approach. After identifying and masking the detected sources, we estimated the background using the biweight location estimator from the Photutils Background2D class. This estimation was performed within a grid of 20×20 pixel sigma-clipped boxes across the image. To construct a low-resolution gridded background model, we applied a median filter to 5×5 neighboring boxes. Finally, we used the BkgZoomInterpolator algorithm to interpolate the median-filtered array, resulting in a smooth background model. This method ensures precise and reliable background subtraction.