Location: Northern Great Plains Research Laboratory
Title: Raspberry Pi Lightweight Deep Learning Models Deployed for Corn Growth Stage Classification Using PhenoCam ImagesAuthor
![]() |
JOICE, ASTINA - North Dakota State University |
![]() |
TAZEEN, HUMEERA - North Dakota State University |
![]() |
TUFAIQUE, TALHA - North Dakota State University |
![]() |
GUPTA, MATHALA J. - North Dakota State University |
![]() |
IGATHINATHANE, C. - North Dakota State University |
![]() |
RAI, NITIN - University Of Florida |
![]() |
Whippo, Craig |
![]() |
ARCHER, DAVID - Retired ARS Employee |
|
Submitted to: Remote Sensing
Publication Type: Peer Reviewed Journal Publication Acceptance Date: 9/17/2026 Publication Date: 9/29/2026 Citation: Joice, A., Tazeen, H., Tufaique, T., Gupta, M., Igathinathane, C., Rai, N., Whippo, C.W., Archer, D.W. 2026. Raspberry Pi Lightweight Deep Learning Models Deployed for Corn Growth Stage Classification Using PhenoCam Images. Remote Sensing. 18(19), 3337. https://doi.org/10.3390/rs18193337. DOI: https://doi.org/10.3390/rs18193337 Interpretive Summary: Combining precision agriculture (PA), which uses data to manage farming, with deep learning (DL) models can really help in predicting crop yields, monitoring crop health, automating field tasks, and making smart decisions. Usually, these data-driven methods need desktops, supercomputers, and cloud clusters to crunch the numbers, but they can be tricky to use in the field. However, a simple computer like a Raspberry Pi (a credit card-sized single-board computer) can be a great, portable tool for making decisions in PA. For example, it can help you figure out the growth stage of corn in the field through image analysis so you producers can manage it better. In this study, we used PhenoCam cameras to take pictures of corn and classified eight different growth stages from ten different sites. We created four lightweight DL models—ELiteCrop0, ELiteCrop1, ELiteCrop4, and MobNetCropV2—and tested them on a supercomputer at five different levels of image clipping (from 0% to 40%). The best model was then put on a Raspberry Pi 5 to make classification in the field. Training the models took up over 97% of the CPU, but testing them was quick, taking just a few seconds. This makes it possible to use the models in real-time and applied in field scenarios. ELiteCrop0 did the best job overall, with a confusion-matrix diagonal ratio (CMDR) of 0.93, followed by ELiteCrop1 (CMDR = 0.92). Model performance was reduced with increased clipping, so it’s best to clip them only to a small amount (around 0%-10%). The supercomputer showed that the models were pretty good at the same sites they were trained on (0.90-0.93), but they weren’t as good at new sites (0.48-0.50). This suggests that the models might not be as good at working with new data and need more training with more images. The Raspberry Pi handled around 1000 images/min without any issues when operating safely at 68 °C. Moving forward, we should consider expanding the dataset across multiple locations to boost performance across different areas. So, this study offers a flexible and affordable way to keep an eye on corn growth stages in precision agriculture in real time. Technical Abstract: Integrating precision agriculture (PA, a data-driven agricultural management system) with deep learning (DL) models can effectively support various activities, including yield prediction, crop health monitoring, field task automation, and decision-making. Taking advantage of such data-driven methodologies typically requires desktops, high-performance computing systems, and cloud clusters for data analysis, but their importability limits in-field applications. However, a single board computer, such as Raspberry Pi, offers a compact, lightweight, cost-efficient, easy-to-use, and feature-rich portable computing device which is ideal for in-field decision-making in PA applications. One such in-field application is crop growth stage classification for better crop management. Therefore, in this study, eight corn growth stages were classified using PhenoCam (near-surface [proximal] remote sensing network camera) imagery collected from ten PhenoCam sites. Four lightweight DL models were developed, ELiteCrop0, ELiteCrop1, ELiteCrop4, and MobNetCropV2, and evaluated across five image vertical clipping levels (0%-40%) using a supercomputer. The optimized model was subsequently deployed on a Raspberry PI 5 for edge inference. Model training accounted for the majority of the total CPU time, exceeding 97%, while the testing times ranged from 0.01 min to 0.12 min, enabling real-time applications. Among the models, ELiteCrop0 achieved the most balanced performance with a confusion-matrix diagonal ratio (CMDR) of 0.93, followed by ELiteCrop1 (CMDR = 0.92). Overall, model performance decreased with increasing vertical clipping; therefore, a moderate image clipping (0%-10%) was recommended for improved computational efficiency. Analysis with a supercomputer produced an intrasite (same train sites) accuracy of 0.90-0.93 (Raspberry Pi: 0.78-0.81) and an intersite (new test sites) accuracy of 0.48-0.50 (Raspberry Pi: 0.41-0.43), indicating challenges with model generalization. Raspberry Pi successfully processed ~1000 images/min under safe operating conditions (68 °C). Future work should focus on extending the multi-site dataset to improve cross-site performance. Hence, this study presents a scalable and cost-effective solution for real-time corn growth stage monitoring in PA. |
