tCrop: Thermal Imaging Based Plant Stress Identification Using On-Edge Deep Learning

Abstract

Plant stress identification is one of the critical tasks to secure food availability for the future. Stress in plants increases considerably due to changing environmental conditions. This paper proposes an innovative solution for the automatic detection of plant stress using computer vision and deep learning for edge computing platforms. The novel model classifies the disease and sends an alert to the cloud if the disease is detected for an early detection of the disease. The idea can be implemented on any edge platform, where the thermal camera captures the image of the crop, the deep learning model predicts the disease in the plant, gives the inference on edge and sends alerts if the disease is detected. The proposed novel CNN model had 90% validation accuracy and 93% test accuracy. The tCrop Lite model had 94% test accuracy and inference time of 0.001 seconds when evaluated on the Raspberry Pi 4 device.

Methodology

tCrop uses thermal images of paddy leaves to detect plant stress on edge devices. The pipeline begins with thermal image capture, followed by preprocessing and class-specific augmentation to address dataset imbalance. A custom CNN extracts local visual features from 224×224 thermal images and classifies disease categories using a softmax prediction head. The model is then optimized into tCrop Lite using TensorFlow Lite float-16 quantization, reducing the model size from 13.1 MB to 2.2 MB for efficient Raspberry Pi deployment.

Results

tCrop Lite reaches 94% test accuracy at 1.13M parameters and 2.18 MB on disk — roughly 50× smaller than Inception ResNetV2 and 120× smaller than ResNet 50, while outperforming both on this benchmark.

Citation

@INPROCEEDINGS{9864547,
  author    = {Bompilwar, Ritik and Rathor, Surya Pratap Singh and Das, Debanjan},
  booktitle = {2022 IEEE Region 10 Symposium (TENSYMP)},
  title     = {tCrop: Thermal Imaging Based Plant Stress Identification Using On-Edge Deep Learning},
  year      = {2022},
  doi       = {10.1109/TENSYMP54529.2022.9864547}
}