Location: Aerial Application Technology Research
Title: Applying machine learning classifiers to Sentinel-2 imagery for early identification of cotton fields to advance boll weevil eradicationAuthor
Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal Publication Acceptance Date: 9/20/2023 Publication Date: 9/28/2023 Citation: Yang, C., Suh, C.P. 2023. Applying machine learning classifiers to Sentinel-2 imagery for early identification of cotton fields to advance boll weevil eradication. Computers and Electronics in Agriculture. https://doi.org/10.1016/j.compag.2023.108268. DOI: https://doi.org/10.1016/j.compag.2023.108268 Interpretive Summary: Early identification of cotton fields at risk of boll weevil infestation is crucial for effective control and eventual eradication of this pest. The use of Sentinel-2 satellite imagery and machine learning techniques were evaluated to identify cotton fields before plants started to bloom. Computer programs were developed to automate the classifier training and image classification process. The results showed that the random forest method had the best performance and it was found to be the most effective classifier in terms of accuracy and processing time. The results and methodologies from this study will provide boll weevil eradication program managers with accurate and rapid classification techniques to identify cotton fields over large geographic areas at relatively early growth stages. Technical Abstract: Early identification of cotton fields that may be at risk of boll weevil infestation is critical for effective control and eventual eradication of this pest. This study evaluated the use of no-cost Sentinel-2 satellite imagery and machine learning techniques to identify cotton fields before cotton plants start to bloom. Three Sentinel-2 images acquired in May and June 2020 over a 10 km by 11 km study area near Snook, Texas, were selected. Field boundaries were digitized for all 320 fields in the study area, and 62 were selected as training fields. To distinguish cotton from other crops, three machine learning classifiers, including random forest (RF), support vector machine (SVM), and k-nearest neighbor (KNN), were applied to the four-band images with 10-m resolution. Pixel- and object-based classification methods were applied to the original images and segmented images, respectively. To automate the classifier training and image classification process, Python programs were developed in the ArcGIS Pro environment to generate 246 classification models based on the three classifiers with different combinations of parameters. Accuracy assessment of the classification maps for the validation fields showed that RF had the best performance with the shortest processing time. Overall accuracy (OA) for RF ranged from 85.8% for May 1 to 90.2% for May 6 to 92.6% for June 10 for all crop areas, with respective F-scores of 0.915, 0.936, and 0.951 for cotton. SVM and KNN had lower accuracy for the two May dates, but they exhibited higher accuracy for June 10. The OA was 94.4% for SVM and 93.3% for KNN with respective cotton F-scores of 0.953 and 0.952 for June 10. However, both SVM and KNN with pixel-based classification took significantly longer processing times. Among the three classifiers, RF is the most effective classifier in terms of accuracy and processing time and therefore is recommended as the first choice for cotton identification. Furthermore, the integration of field boundaries and original classification maps allowed for field-level classification with increased accuracy. The results and methodologies from this study will provide boll weevil eradication program managers with accurate and rapid classification techniques to identify cotton fields over large geographic areas at relatively early growth stages. |