SEARCHOTHERagriculture, plant_phenotyping

other·agriculture, plant_phenotyping, computer_vision, deep_learning, remote_sensing·PMC10076056

Rice Plant Counting, Locating, and Sizing Method Based on High-Throughput UAV RGB Images

Plant Phenomics · 9 authors, 5 centres

AI SUMMARY

FIDELITY 88%

POPULATIONRice plants in paddy fields (indica-type rice) imaged via UAV at 7 m altitude

INTERVENTIONRiceNet deep learning network with multiscale feature fusion, plant attention mechanism, and positive-negative loss function

COMPARISONMCNN, CSRNet, SANet, TasselNetV2, FIDTM (on URC dataset); MCNN, CSRNet, BCNet, SFC²Net, TasselNet, TasselNetV2 (on MTC and WED datasets)

This summary was generated by AI from a single paper. It has not been reviewed by a clinician and is not clinical advice. Verify against the source before acting on it.

AGENT 01 · STRUCTURED SUMMARY

TRACE TO SOURCE

“

TL;DR

This paper presents RiceNet, a deep learning network that uses UAV RGB images to automatically count, locate, and size rice plants in paddy fields. On a new dataset of 355 images with 257,793 manually labeled points, RiceNet achieved a mean absolute error of 8.6 and root mean square error of 11.2, significantly outperforming existing methods. The approach enables contactless, high-throughput rice plant phenotyping that could replace labor-intensive manual counting in agricultural research and production.

Full summary

3,383 CHARS

**Background:** Rice plant counting is essential for yield estimation, growth diagnosis, disaster loss assessment, and breeding. Traditional manual counting is labor-intensive, time-consuming, prone to human error, and limited to small sampling areas. Existing automated methods for other crops (corn, wheat, maize) cannot be directly applied to rice due to differences in plant morphology, imaging conditions, and dataset characteristics. The authors aimed to develop a deep learning approach using affordable UAV RGB imagery to simultaneously count, locate, and size rice plants in paddy fields.

**Methods:** The authors collected RGB images using a DJI Phantom 4 Advanced quadcopter UAV at 7 m altitude with vertical downward shooting, front overlap of 80%, and side overlap of 70%. Images were captured every three days during tillering to jointing stages from 2018–2020 in Nanchang, Jiangxi Province, China. The resulting URC dataset contains 355 images (original resolution 5,472 × 3,648, downsampled to 1,368 × 912 for training) with 257,793 manually labeled center points. RiceNet architecture uses the first 13 layers of VGG16_bn as a feature extractor, followed by three decoder modules: density map estimator (DME) with plant attention mechanism, plant location detector (PLD) using local non-maximum suppression, and plant size estimator (PSE). A novel positive-negative loss function was introduced to suppress background false positives. Training used Adam optimizer with learning rate 1×10⁻⁴, batch size 3, and hyperparameters λ=0.1 and γ=0.1. Performance was evaluated using MAE and RMSE.

**Key Results:** On the URC dataset, RiceNet achieved MAE of 8.6 and RMSE of 11.2, outperforming MCNN (MAE 25.5, RMSE 34.1), CSRNet (MAE 12.9, RMSE 17.5), SANet (MAE 10.1, RMSE 13.4), TasselNetV2 (MAE 29.5, RMSE 39.4), and FIDTM (MAE 26.3, RMSE 31.5). Compared to TasselNetV2 and FIDTM, MAE improved by 70.8% and 67.3%, respectively. On the MTC maize tassel dataset, RiceNet achieved MAE of 3.4 and RMSE of 5.3, with 81.0% MAE improvement over MCNN. On the WED wheat ear dataset, RiceNet achieved MAE of 3.7 and RMSE of 4.6, with 67.8% MAE improvement over MCNN. R² values were 0.997 (URC), 0.957 (MTC), and 0.975 (WED). Ablation experiments showed positive-negative loss reduced MAE from 10.6 to 8.6 for RiceNet, and from 25.5 to 11.7 for MCNN. Optimal Gaussian kernel size σ was 6. The network also demonstrated robustness to sun reflection and raindrops in field conditions.

**Clinical Implications:** While not a clinical study, this agricultural technology has significant implications for food security and crop production. Accurate, automated rice plant counting enables more effective planting density management, objective disaster assessment from typhoons and floods, and improved rice breeding through precise phenotyping. The ability to estimate plant location and size from point-level annotations (rather than expensive bounding box labels) makes the approach practical for large-scale deployment. The authors note limitations including sensitivity to weeds, sun reflection on water, and the need for further research on quantitative evaluation of location and size estimation. Future directions include developing networks robust to sun reflection, further suppressing false positives, and creating lightweight versions for field robots and embedded platforms.

PICO

PPOPULATION

Rice plants in paddy fields (indica-type rice) imaged via UAV at 7 m altitude

IINTERVENTION

RiceNet deep learning network with multiscale feature fusion, plant attention mechanism, and positive-negative loss function

OOUTCOME

Mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), plant location and size estimation accuracy

STUDY TYPE

other

SPECIALTY

agriculture

SUMMARISED BY

AI pipeline

FIDELITY CHECK

88% · B

ORIGINAL PAPER

PMC10076056

Rice Plant Counting, Locating, and Sizing Method Based on High-Throughput UAV RGB Images

Plant Phenomics

READ THE FULL PAPER

ABSTRACT

Rice plant counting is crucial for many applications in rice production, such as yield estimation, growth diagnosis, disaster loss assessment, etc. Currently, rice counting still heavily relies on tedious and time-consuming manual operation. To alleviate the workload of rice counting, we employed an UAV (unmanned aerial vehicle) to collect the RGB images of the paddy field. Then, we proposed a new rice plant counting, locating, and sizing method (RiceNet), which consists of one feature extractor frontend and 3 feature decoder modules, namely, density map estimator, plant location detector, and plant size estimator. In RiceNet, rice plant attention mechanism and positive–negative loss are designed to improve the ability to distinguish plants from background and the quality of the estimated density maps. To verify the validity of our method, we propose a new UAV-based rice counting dataset, which contains 355 images and 257,793 manual labeled points. Experiment results show that the mean absolute error and root mean square error of the proposed RiceNet are 8.6 and 11.2, respectively. Moreover, we validated the performance of our method with two other popular crop datasets. On these three datasets, our method significantly outperforms state-of-the-art methods. Results suggest that RiceNet can accurately and efficiently estimate the number of rice plants and replace the traditional manual method.

CITATION EXPORT

Xiaodong Bai, Pichao Liu, Zhiguo Cao et al.. Rice Plant Counting, Locating, and Sizing Method Based on High-Throughput UAV RGB Images. Plant Phenomics. (2023). https://doi.org/10.34133/plantphenomics.0020

AUTHORS

9 · 5 CENTRES

Xiaodong Bai, Pichao Liu, Zhiguo Cao, Hao Lu, Haipeng Xiong, Aiping Yang, Zhe Cai, Jianjun Wang, Jianguo Yao

1School of Computer Science and Technology, Hainan University, Haikou 570228, China.
2School of Telecommunication and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China.
3School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China.
4School of Computing, National University of Singapore, Singapore 119077, Singapore.
5Agricultural Meteorological Center, Jiangxi Meteorological Bureau, Nanchang 330045, China.