Optimizing Harvest Piles Detection using CNNs, Vision Transformers, and Remote Sensing
Open Access DepositedAccurate agricultural monitoring through satellite imagery plays an important role in curbing global challenges such as food insecurity, drought response, and climate change, particularly in developing countries where smallholder farming is predominant. Due to lack of infrastructure and resource limitations, the work of gathering statistical data to determine the productivity of remote areas is very difficult. In this research, we investigate the use of deep learning techniques to detect harvest piles with improved performance from previous studies. Using high performance machine learning techniques coupled with satellite imagery, it's possible to determine the presence or absence of agricultural production in remote areas.In this research, a robust dataset—HarvestNet_B— was constructed by integrating satellite imagery from Planet Scope, SkySat, and EuroSat, resulting in over 442,000 labeled training images and 16,821 test images across ten land use and land cover (LULC) classes. Multiple state-of-the-art convolutional and transformer-based models, including ResNet50, EfficientNetB2, ViT-B/16, and ViT-Max, were finetuned and evaluated on the new dataset. The results show significant improvements in classification performance, achieving over 90% precision, exceeding the performance obtained from using ResNet baseline models in prior studies. Furthermore, a novel ensemble architecture was developed by combining EfficientNetB2 and ViT-B/16 backbone architectures by adding feature fusion and dense layers. This hybrid model achieved 99.73% classification accuracy, 99.90% precision, and a 99.8% F1-score in detecting harvest piles, demonstrating its effectiveness for high-stakes geospatial classification tasks. Key preprocessing methods, such as image patching, were also introduced, significantly enhancing the learning capabilities of both CNN and Transformer models by improving their ability to focus on relevant local and global features. The proposed methods are well-suited for real-world deployment and decision support in agricultural policy, resource planning, and humanitarian logistics. This work lays a foundation for using satellite-based machine learning to inform agricultural productivity assessment. Future research is recommended to extend the current binary classification framework into a quantitative yield estimation model using object detection or instance segmentation, enabling scalable, fine-grained monitoring of agricultural production in data-scarce environments.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.