Crop residue management is a critical aspect of sustainable agriculture, affecting soil health, nutrient cycling, and environmental quality. Accurately measuring the crop residue present in the fields before and after tillage operation will help farmers plan their activity accordingly. Various methods for measuring and sensing crop residue, including traditional techniques like the line-transect method, photographic approaches, thresholding-based machine learning models, and more advanced deep learning methods, are available. However, each of these approaches has limitations in terms of accuracy, efficiency, and resource requirements. In this study, we present a unique approach of residue image data collection with a range of crop residue levels using on-machine sensors to capture RGB images from mono cameras. The multi-season and multi-year dataset of more than 25,000 images was utilized for supervised machine learning using pretrained ResNet18 model. Using both in-field line transect method and photographic method as ground truth, model accuracy was tested in different scenarios for predicting the crop residue cover (CRC). The model predicted CRC into 5 broader classes of 0-20%, 21-35%, 36-55%, 56-75%, and 76-100% CRC. A test accuracy of 85.8% and 61.8% was achieved for same-field images and new-field images, respectively. A broader evaluation metric, the ‘Delta±1 accuracy’ criteria was introduced to describe model performance with wider class tolerances to account for cases where the model’s prediction is close to the ground truth. An accuracy of 99.1% and 93.5% was achieved for same-field images, and new-field images respectively using the Delta±1 criteria. This study presents a novel approach to crop residue estimation using the crop residue imagery dataset, which could enable real-time sensing of crop residue via on-machine sensors.
Regmi et al. (2026) studied this question.