EWIS1: Maize, Sorghum & Weed Dataset

Description

Precision agriculture relies on Site-Specific Weed Management (SSWM) to replace traditional, uniform spraying with data-driven herbicide application and mechanical weeding. As weeds compete with crops for space and nutrients, they remain a primary threat to harvest quality. As a first step, the implementation of SSWM requires high precision in automated detection and classification of weed species. Therefore we publish a manually annotated and expert curated drone image dataset for weed segmentation (crop vs weed) in sorghum and maize fields with different crop growth stages and weather conditions.

Steps to reproduce

We used a consumer-grade drone (DJI Mavic 2 Pro) with a 20 MP Hasselblad camera (L1D-20c) to capture images with a resolution of 5464x3640 pixels². The drone conducted automated flight missions with the camera pointing nadir at a flight altitude of five meters above ground level. At this altitude, the corresponding GSD was approx. one millimeter. At the border of each field, we placed a MicaSense Calibrated Reflectance Panel (RP04-1847249-SC) to act as white balance. Afterwards, we used this panel to adjust the white balance in all RAW image files (DNG format) of a specific drone flight. The resulting images were exported as png, as it is a lossless format and uploaded in this dataset under "images". Further, we upload the JPG files of 4 additional drone flights of maize and sorghum. We selected a wide variety of images (n = 88) from different drone flights for annotation. Here, we used GIMP 2.8 (GNU Image Manipulation Program) to manually label weed and crop plants in the drone images. The exported semantic segmentation masks are uploaded in this dataset under "masks". These masks are also in png format with crop plants labeled with a blue color (RGB=[31, 119,180]) and weeds colored in orange (RGB=[255,127,14]). The background is transparent. Bounding Boxes are derived from the segmentation masks, are filtered by plant area (> 50px²) and uploaded into "bounding_boxes".

Categories

Funders

• Bavarian State Ministry for Food, Agriculture and Forests

  Grant ID: G2/N/19/13

Licence