# Plant stress detection using multimodal imaging and machine learning: from leaf spectra to smartphone applications

## TL;DR

This paper explores using smartphone-based imaging and machine learning to detect plant stress at low cost, comparing traditional and modern methods.

## Contribution

The paper introduces machine learning integration with multimodal imaging and smartphone tech for scalable, affordable plant stress detection.

## Key findings

- Multispectral and thermal imaging can detect plant stress earlier than traditional methods.
- Smartphone-based platforms offer a low-cost alternative for real-time plant stress monitoring.
- Machine learning improves automation and reduces reliance on expensive equipment.

## Abstract

Plant leaf spectrophotometry has been used successfully as a means to detect stress, and it has been complemented by fluorescence analysis. This identification can be achieved in the ultraviolet (UV), visible (red, green, blue; RGB), near-infrared (NIR), and infrared (IR) spectral regions. Hyperspectral (measuring continuous wavelength bands) and multispectral (measuring discrete wavelength bands) imaging modalities can provide detailed information concerning the physiological well-being of plants, often diagnosing them at an earlier stage than visual or other more traditional biochemical assays. Because hyperspectral methods are highly sensitive and accurate, they cost a lot and produce vast quantities of data, which demand sophisticated computing software, and compared to multimedia, multispectral, and RGB cameras, they are less expensive and easier to carry but have reduced spectral resolution. Such methods are justified by thermal and fluorescence images revealing variations in the temperature and efficiency of photosynthesis of the leaves in response to stress. New digital imaging, thermal imaging, and optical filter technologies, and advancements in smartphone cameras have rendered low-cost, field-deployable platforms to monitor plant stress in real time feasible. Machine learning also supports these techniques by automating feature extraction, classification, and prediction to reduce the use of expensive instrumentation and human skill. But also problems like sensor calibration in a changing field, low model generalization across species and environments, and large, annotated datasets are needed. Beyond highlighting the relative strengths of the conventional and contemporary sensing approaches, the paper also examines the possibility of applying machine learning to multimodal images, as well as the growing impact of smartphone- based solutions in supplying inexpensive agricultural diagnostics. It concludes by overviewing the current limitations and limits to future research into scalable, cost-effective, and generalizable plant stress models.