Chroma-Sense 2.0: A Memory-Efficient Two-Stage Pipeline for Lightweight On-Device Plant Disease Segmentation and Classification

On-device plant disease perception must reconcile two competing demands: enough spatial detail to localise diseased tissue within a field image and a memory and compute budget small enough for microcontroller-class hardware. Single-network solutions that jointly learn a pixel-wise mask and a fine-grained disease label tend to oversubscribe both the Flash and activation SRAM of such devices. This paper presents Chroma-Sense 2.0, a two-stage lightweight pipeline that decouples the two subproblems and sizes each stage for its own budget. The two stages run sequentially on the same frame: Stage 1 is a per-channel convolutional classifier, derived from Chroma-Sense, that names the disease, and Stage 2 is a compact ESPNet segmenter that produces a binary diseased-versus-healthy mask localising it. Because the stages run one after the other rather than concurrently, the peak working memory of the pipeline is the maximum of the two stages rather than their sum. We evaluate the pipeline on the in-the-wild PlantSeg dataset using a curated 10-species, 34-class subset and a leakage-controlled protocol in which all training crops are derived from PlantSeg’s official training images and all reported metrics are measured on a held-out test set of 5002 crops built from the official test images. The segmentation stage attains a mean foreground recall of 0.97 (mean foreground IoU of 0.49; 0.53 pooled over pixels), a deliberately recall-oriented operating point. Against Fast-SCNN, a small U-Net, LR-ASPP, and DeepLabV3+, ESPNet is the smallest-footprint model (140k parameters, 193 KB Int8 Flash) while retaining the highest foreground recall; the per-channel classifier reaches accuracies comparable to much larger ImageNet-pretrained backbones (MobileNetV3 and EfficientNet) using 10–13× fewer parameters. End to end, the coupled pipeline classifies the disease correctly on 87.8% of the test crops. An on-device profile on the OpenMV H7 and H7 Plus shows that the binding constraint at 256 × 256 is the segmenter’s ≈4 MB contiguous activation arena, rather than parameter Flash: Even on the 32 MB-SDRAM H7 Plus, the usable interpreter heap is only about 4 MB, and the arena cannot be allocated as a single contiguous block from it, so on the tested firmware, the classifier runs on microcontrollers while the segmenter does not; the full pipeline instead fits the gigabyte-scale single-board-computer tier (for example, Raspberry Pi or NVIDIA Jetson Nano), and enabling the segmenter to run on microcontrollers remains the open gap.