SPAGRI-AI: Smart precision agriculture dataset of aerial images at different heights for crop and weed detection using super-resolution

Recently, smart agriculture has become an essential part of modern agriculture approaches from tillage, via plant seeding and grow support to their collection. With modern technologies, farmers can use substances like pesticides, herbicides, or fertilizers at precise dosages or to identify places on a field with specific production rates. The main objective of this study is to introduce a novel and a unique aerial image dataset of various fields acquired by UAV containing crops/weeds in the early phenophases captured in two different resolutions (2 mm and 7 mm per pixel). Secondly, the best super-resolution technique for high-resolution images, substitution with lower resolution is explored. For data acquisition, we employed DJI Matrice 600 equipped with a full-frame Sony Alpha A7R IV285 image sensor. Data were captured at flight heights of 26 and 95 m from 4 different fields in Central Europe. In addition, we proposed a methodology focused on the selection of an appropriate super-resolution method to enhance low-resolution aerial images to obtain better accuracy of crop/weed detection. As a baseline crop/weed detector for super-resolution effect evaluation, YOLOv5 architecture was used. Next, we explored the performance of several super-resolution models (U-Net++, ESRGAN, SwinIR), and fine-tuned the best-performed one. We present the new dataset named SPAGRI-AI: a novel unique dataset of aerial images for super-resolution experiments in smart precision agriculture. The dataset contains 27,638 aerial images (1024 × 1024 px) and additionally, it contains a subset of 2014 labeled images with 45,548 bounding boxes of 12 classes. The main purpose of the SPAGRI-AI is to provide the scientific community with real-world data to test new methods for super-resolution (SR) and crop/weed detection. During the evaluation of selected super-resolution models, the YOLOv5 model trained on high-resolution images resulted in corn [email protected] of 94.48%. The YOLOv5 model trained on low-resolution images resulted in corn [email protected] of only 51.43%. Nevertheless, if the low-resolution images were pre-processed using the SwinIR super-resolution method, corn [email protected] of 62.36% was achieved. To the best of our knowledge, it is one of the largest datasets available to the paper's publication date. Overall, the SPAGRI-AI dataset and the findings from our experiments contribute to the advancement of super-resolution techniques and crop/weed detection methods in the field of smart agriculture. By utilizing real-world data and optimizing image enhancement approaches, we paved the way for further developments in precision farming practices and applying emerging technologies in agriculture.