Info: Zenodo’s user support line is staffed on regular business days between Dec 23 and Jan 5. Response times may be slightly longer than normal.

Published October 25, 2023 | Version 1.0
Dataset Open

Maize Phosphorus Leaf Deficiency (MPLD) Database 224x224

Creators

  • 1. ROR icon Universidad EAFIT

Contributors

  • 1. ROR icon Universidad EAFIT
  • 2. ROR icon Universidad de Las Palmas de Gran Canaria

Description

Monitoring the nutritional status of crops is crucial to assessing high productivity, optimizing cost, and minimizing environmental impact. Given that nutritional deficiencies primarily manifest through visual characteristics, artificial vision stands out as a competitive choice to assess the nutritional status of individual plants.

However, in order to train a supervised artificial vision system driven by convolutional neural networks (CNNs), a high amount of data, properly formatted and labeled is necessary. This work presents a curated image database to study single-nutrient deficiencies, specifically, phosphorus deficiency in maize leaves, named Maize Phosphorus Leaf Deficiency (MPLD) Database.

This database is composed of 20892 samples of maize leaves placed on a withe background. Images are 224x224 pixels size, representing three levels of phosphorus deficiency: complete absence of the nutrient (labeled -P), half dose of the required phosphorus for normal plant development (-P50), and complete supply (C). 

Methods

Data collection

Image acquisition was done by searching a diversity of conditions, although the plants were grown in laboratory conditions. In this way, photographs included the growth stages of seedling, jointing, and flowering. The images were acquired inside a plastic shed, involving natural illumination, and five consecutive images were taken by plant. 

Five acquisition devices were utilized, encompassing two types of regular smartphones (Xiaomi Redmi 8T and Moto G5 Plus), a digital camera (Samsung ES65), a single-lens reflex camera (Nikon D3100), and a compact scientific camera (ThorLabs DCC1645C-HQ).

Image preprocessing and augmentation

The original images underwent a preprocessing stage in terms of cropping and resizing by four concurrent methods, named: 

  1. Cut to square: Image cropped to a central square.
  2. Quadrant division: Square image divided into four equal parts.
  3. Split into two parts: Rectangular image divided by two.
  4. Padding to square: A withe pad is added on the sides of rectangular image to complete it square.

All cropped images were resized to 224x224 pixels size for fast deep learning model adaptation.

Photos taken by each camera experienced at least two of four possible preprocessing methods, depending on the characteristics of the image. 

 

Table of contents

ZIP folder is structured as follows: 

  1. Five main folders represent images taken by each camera. 
  2. Inside every folder, there are three more folders containing each nutritional treatment: -P, -P50, and C. 
  3. Inside each treatment folder, the preprocessed images are placed with no division between preprocessing methods. 

Files

MPLD_databse.zip

Files (289.6 MB)

Name Size Download all
md5:3a8e25429ed2b8e9399e66d0d8df0d0d
289.6 MB Preview Download

Additional details

Dates

Collected
2022-05-13
Image collection
Submitted
2023-10-25
Dataset V 1.0 submission