Image Dataset of Local Cocoa, Maize and Soybean Seed Varieties from Ghana

Description

This dataset contains 46,843 raw, expert-confirmed seed/bean images covering 13 locally recognised varieties across three economically important Ghanaian crops: cocoa (4 varieties — Agric, Gardener, Hybrid, Tetteh Quarshie; 3,323 images), maize (5 varieties — Bhihilifa, Obaatampa, Sanzal Sima, Wang Basik, Wang Dataa; 25,696 images), and soybean (4 varieties — Afayak, FarmerSaved, Favor, Tenguma; 17,824 images). Maize and soybean seeds were pre-sorted by variety at Heritage Seeds Ghana; cocoa seeds were extracted from pods pre-sorted by variety by farm experts in Berekum, Ghana, and sun-dried for seven days before imaging. All seeds were photographed in groups on a plain background with a Google Pixel 9a smartphone, then cropped into individual seed images, with variety labels confirmed throughout by the respective experts. Images are resized to 224 × 224 px and organised into folders by crop and variety. The dataset supports development of low-cost, mobile deep-learning models for seed variety identification and certification, addressing the current lack of openly available, cultivar-labelled image data for African seed crops. Class distributions are close to balanced, though a few varieties (e.g., Agric cocoa, Obaatampa maize) are comparatively under-represented

Steps to reproduce

Maize and soybean seeds were sourced and already sorted by variety from Heritage Seeds Ghana, Tamale. Cocoa pods were sourced pre-sorted by variety from farm experts in Berekum, Ghana; seeds were removed from the pods, labelled by variety, and sun-dried for 7 days under the farmers' guidance. For each variety, the seeds/beans were arranged in groups on a plain background and photographed during daylight hours using a Google Pixel 9a smartphone (48 MP wide, ƒ/1.7 rear camera). Each grouped photograph was tagged with its known crop and variety, then individual seed images were cropped out, carrying the label through. Variety labels were verified throughout—by Heritage Seeds Ghana specialists (maize, soybean) and the Berekum farm experts (cocoa)—from initial sorting to the final cropped images. Confirmed images were organized into class folders by crop and variety and resized to 224 × 224 px. This acquisition setup (grouped capture, plain background, natural daylight) follows our previously published maize dataset (Asante et al., 2023, doi:10.17632/r6vvm5jkh6.2), adapted here for multi-crop variety classification rather than quality grading.

Institutions

• University of Mines and Technology

  Western, Tarkwa
• University of Energy and Natural Resources

  Bono, Sunyani

Categories

Licence