Deep learning based models for yield estimation using drone imagery

Abstract

Agriculture plays a crucial role in the economies of developing countries, contributing 37% to gross domestic product (GDP), nearly 60% to export earnings, and providing employment for over 76% of the population. Accurate crop yield estimation is essential for optimizing agricultural practices, ensuring food security, and enhancing economic stability. Traditional yield estimation methods, such as actual fruit counts, can exceed optimal numbers by 10 to 20 fold. Modern agricultural techniques, like deep learning, are required to manage yield components more effectively. This study developed deep learning models for yield estimation using the collected highresolution drone imagery. The research involved systematically collecting and annotating the drone image dataset for coffee, cashew and cocoa, developing advanced image classification and object detection models, and integrating these models into a robust yield estimation pipeline. The dataset collection was a collaborative effort involving the Makerere Artificial Intelligence Lab, Marconi Research and Innovation Lab, National Coffee Research Institute (NaCORI), and KaraAgroAI. We used different machine learning approaches, including traditional methods, deep learning, transfer learning, and foundation models, to find the best image classification model. The custom deep learning model emerged as the best, with an overall accuracy of 99% in classifying various crop types. For the object detection model, state-of-the-art YOLO (You Only Look Once) models and transformer models were customer-trained on our dataset, and YOLOv9 had better precision in identifying and locating the crops, with mean Average Precision (mAP) scores of 0.832 for cocoa, 0.546 for coffee, and 0.543 for cashew. We then used a Python package called Supervision to count the detected crops as the yield estimation system in the image. The image classification and object detection models were then used to develop a yield estimation pipeline in this study, which has proven to be accurate, providing a practical solution for real-time yield prediction. However, the study also encountered limitations, including constraints in data collection and a focus on a limited number of crops (only three were considered for yield estimation), high-resolution requirements, and significant computational resource needs. To address these limitations, we recommend expanding data collection efforts, utilizing more accessible technology, improving computational efficiency, and creating automated annotation tools. Furthermore, future research should explore multispectral and hyperspectral imaging, real-time monitoring systems (including using these models on video, not just images), and longitudinal studies to enhance yield estimation accuracy and applicability.