Abstract: Plankton are fundamental components of marine ecosystems, underpinning biodiversity and forming the base of the marine food web. Fluctuations in plankton populations not only regulate biodiversity and trophic dynamics but also serve as sensitive indicators of marine environmental conditions. Conventional survey methods, which rely on manual sampling and microscopic examination, are time-consuming, labor-intensive, and inadequate for continuous, high-resolution in-situ monitoring. To address these limitations, we developed and deployed a compact in-situ plankton observation system based on backlight imaging. The system integrates automated image acquisition, onboard preprocessing, and a convolutional neural network (CNN) classifier based on ResNet50, enabling real-time plankton identification directly in the marine environment. The system was deployed during an ecological survey conducted along the southeastern coast of Hainan Island between August and September 2023. It successfully captured and classified in-situ plankton images, achieving a classification accuracy of 98.2% on the collected dataset. Statistical analysis revealed pronounced spatial variations: zooplankton abundance ranged from 94.04 to 4325.46 ind/m³ (mean: 1 472.91 ind/m³), whereas phytoplankton abundance ranged from 82.71 to 18 704.70 ind/m³ (mean: 2 398.86 ind/m³). These findings demonstrate that the developed system provides accurate plankton classification and reliable in-situ abundance estimates. By enabling automated, continuous, and high-resolution monitoring, the system constitutes a valuable tool for marine ecological research and environmental assessment. It also offers significant potential for supporting biodiversity assessments, tracking plankton community dynamics, and providing early warning of ecosystem change.