CalibCompress-YOLOv9: Sensor-Calibrated Rate-Adaptive Video Inference for Real-Time Tomato Disease Monitoring under Edge Bandwidth and Latency Constraints

Existing agricultural video disease monitoring systems fail on three fronts: uncalibrated sensor response under changing field illumination, bandwidth-induced loss of lesion texture during transmission, and unstable frame-level predictions caused by motion blur, partial occlusion, and background clutter in live crop videos. This paper introduces Sensor-Calibrated Rate- Adaptive YOLOv9 Video Inference (CalibCompress-YOLOv9), integrating radiometric and colour calibration, temporal exposure normalization, structural-similarity-preserving compression control, quantization-aware You Only Look Once version 9 detection, shortwindow temporal consensus, and uncertainty-aware recapture gating within a deployable edgeimaging pipeline. In the implementation process, the received video frames are processed with sensor calibration to compensate illumination and colour differences, adaptive compression to maintain lesion structure while reducing bandwidth, and temporal consensus to build stable frame sequences prior to YOLOv9 inference, and thus the output of this stage is the calibrated, bandwidth friendly, and temporally coherent frames that can be used to perform real-time tomato disease detection with high confidence. Evaluation is performed on a five-class tomato disease video dataset covering healthy leaves, bacterial spot, early blight, late blight, and powdery mildew under controlled and natural outdoor illumination. CalibCompress-YOLOv9 achieves throughput of 30 frames per second, median end-to-end latency below 65 ms, Structural Similarity Index Measure preservation of at least 0.96 after adaptive compression, and mean Average Precision at 0.5 intersection-over-union of 0.95. The outcomes demonstrate that the integrated calibrated acquisition process, bandwidth-aware compression, temporal decision smoothing and lightweight edge inference offers a realistic approach to reliable monitoring of tomato diseases at the field scale. This joint design narrows the gap between laboratory disease detection and at-scale agricultural video intelligence by considering image quality, network load, and inference reliability as interconnected system requirements, rather than separate model-training goals.