Quantifying the Effects of Sensor Noise and Spatial Resolution on Small Target Detection in Multispectral Remote Sensing Imagery

In the field of high-resolution optical remote sensing image analysis， target detection and recognition technologies serve as foundational pillars for critical applications such as environmental monitoring， urban planning， agricultural yield estimation， and military reconnaissance. In recent years， deep learning algorithms have revolutionized this domain by leveraging their exceptional ability to extract intricate features： they efficiently capture both the geometric characteristics of ground objects （e.g.， shape contours， spatial distribution patterns， and structural details） and spectral attributes （e.g.， band reflectance variations， spectral curve uniqueness across land covers）. This capability has driven transformative breakthroughs in target detection for high-resolution optical remote sensing images， enabling more accurate and automated identification of diverse ground targets compared to traditional handcrafted-feature-based methods.However， the practical application of deep learning in multispectral remote sensing image analysis is hindered by inherent challenges stemming from data acquisition and scene complexity. First， multispectral images exhibit significant variability in both image quality and spatial resolution due to differences in sensor platforms—ranging from satellite-borne sensors （with meter to sub-meter resolution） to Unmanned Aerial Vehicle （UAV） cameras and airborne imaging systems. These hardware discrepancies lead to inconsistencies in Signal-to-Noise Ratio （SNR）， radiometric accuracy， and spatial granularity， directly impacting the reliability of feature extraction. Second， remote sensing scenes are inherently complex and variable： factors such as topographic relief， cloud cover， atmospheric scattering， and heterogeneous land cover （e.g.， intermingled vegetation， buildings， and water bodies） create a cluttered background that obscures small targets （e.g.， small vehicles， individual buildings， or vegetation patches）. These challenges are particularly acute for small target detection in heterogeneous scenes， where traditional deep learning models often suffer from high false detection rates， missed detections， or degraded localization accuracy due to insufficient discriminative features or severe background interference.To address these issues， this study focuses on the SuperYOLO model—a state-of-the-art object detection architecture optimized for remote sensing scenarios—and conducts a systematic investigation to evaluate its performance in small target detection across diverse multispectral datasets. The research specifically analyzes two core factors influencing detection capability： the characteristics of scene datasets and the quality/resolution of multispectral images. For dataset analysis， the study employs two types of experimental data： independent datasets （VEDAI and HRRSD） and a mixed dataset. The VEDAI dataset specializes in small ground targets （primarily vehicles） with consistent data distribution and high annotation precision， while the HRRSD dataset covers multiple small target categories （e.g.， buildings， roads， and small vegetation clusters） in complex urban and natural scenes. In contrast， the mixed dataset is constructed by fusing multispectral data from different sensors， geographic regions， and acquisition times， simulating the high data heterogeneity encountered in real-world remote sensing applications.The research findings yield three key insights. First， dataset homogeneity significantly impacts SuperYOLO's detection accuracy： the model achieves notably higher performance on independent datasets （VEDAI and HRRSD） compared to the mixed dataset. This is attributed to the consistent data distribution and reduced background variability in independent datasets， which minimize the model's sensitivity to irrelevant background features； in contrast， the mixed dataset's diverse land cover types and spectral variations increase the difficulty of distinguishing small targets from cluttered backgrounds， leading to reduced detection stability. Second， instrument noise emerges as a critical limiting factor for SuperYOLO's precision. When 5% and 10% instrument noise （simulating real-world sensor electronic noise， radiometric distortion， and atmospheric interference） are introduced into multispectral images， the model's mean Average Precision （AP） at an Intersection-over-Union （IoU） threshold of 0.5 （mAP@0.5） plummets sharply from the original 0.813 to 0.420 and 0.058， respectively. This degradation occurs because noise disrupts the integrity of spectral features in multispectral bands， interfering with the multi-branch network architecture of SuperYOLO—specifically impairing its ability to extract and fuse discriminative spectral and geometric features for small targets. Notably， integrating a Denoising Convolutional Neural Network （DCNN） into the data preprocessing module effectively mitigates this issue： the DCNN's adaptive noise-filtering capability restores spectral feature integrity， leading to a significant recovery of SuperYOLO's detection accuracy. Third， the spatial resolution of multispectral images exerts a relatively minor impact on SuperYOLO's performance. After downsampling the original images to simulate lower-resolution scenarios， the model's mAP@0.5 decreases from 0.813 to 0.741 （for mild downsampling） and 0.464 （for severe downsampling）. This resilience is attributed to SuperYOLO's feature pyramid structure， which enables multi-scale feature fusion to compensate for partial resolution loss—though severe downsampling still causes accuracy degradation due to the loss of fine-grained details in small targets.In conclusion， instrument noise significantly impairs the extraction and fusion of spectral features in SuperYOLO's multi-branch architecture， ultimately limiting its detection accuracy and practical utility for small targets in multispectral remote sensing images. Enhancing optical image quality （e.g.， optimizing sensor hardware to reduce inherent noise） and incorporating denoising algorithms （such as DCNN） into the SuperYOLO framework can substantially improve the detection capability， accuracy， and applicability of small target detection algorithms in multispectral remote sensing imagery. This study emphasizes that maximizing the potential of deep learning models in remote sensing applications requires prioritizing solutions to sensor-induced noise and optimizing image preprocessing strategies. The proposed integration of denoising techniques with advanced detection architectures like SuperYOLO not only provides a more practical technical solution for small target detection in multispectral remote sensing but also points to a promising direction for future research （e.g.， multi-modal data fusion detection， real-time remote sensing image processing） and engineering applications （e.g.， UAV inspection systems， satellite remote sensing monitoring platforms） in this field.