Applying machine learning classifiers to Sentinel-2 imagery for early identification of cotton fields to advance boll weevil eradication

Early identification of cotton fields is critical for boll weevil eradication programs to establish field traps for monitoring weevil populations. However, most studies on crop type mapping have focused on late- or postseason crop classification. This study proposed and evaluated a field-level classification method that integrated field boundary data and early-season Sentinel-2 imagery to identify cotton fields 20-60 days after planting. Three cloud-free images were selected from May and June 2020, covering a 110-square-km study area in Texas. To achieve field-level classification, the boundaries of all 320 fields in the study area were used, and 62 of them were selected as training fields. Three machine learning classifiers - random forest (RF), support vector machine (SVM), and k-nearest neighbor (KNN) - were applied to the four 10-m bands for pixel-based classification and to segmented images by mean shift segmentation for object-based classification to determine the optimal classifier. To assess the impact of additional bands and vegetation indices (VIs) on classification accuracy, the six 20-m bands were combined with the 10-m bands to create 10-band images with 10-m resolution, and six vegetation indices (VIs) were derived and combined with the images. Recursive feature selection was used to identify optimal band combinations and principal component (PC) analysis was applied to reduce the dimensionality of the images. Based on accuracy measures derived from 2160 error matrices for all classification maps, RF provided the highest accuracy with faster processing times for the two May dates, while SVM and KNN exhibited slightly higher accuracy for June 10. The integration of field boundary data significantly improved classification accuracy for all classification maps. The overall accuracy (OA) of the optimal RF classification maps for the 4-band images increased from 71.4 %, 76.3 %, and 82.6 % to 85.8 %, 90.2 %, and 92.6 % (a net increase of 14.4 %, 13.9 %, and 10.0 %) for May 1, May 6, and June 10, respectively. Moreover, by including the six 20-m bands, the OA further increased by 4.8 %, 2.9 %, and 3.0 % for the respective dates. Similarly, the F-score for cotton increased from 0.801, 0.845, and 0.882 to 0.936, 0.947, and 0.976 for the three respective dates. In contrast, the inclusion of VIs had minimal effect on classification accuracy. The optimal bands and the PCs provided better classification results and were used for image classification to reduce and optimize input feature to the classifiers. These results demonstrate that combining RF with 10-band Sentinel-2 imagery and field boundary data is an effective method for early identification of cotton fields. This study provides practical techniques for boll weevil eradication program managers to identify cotton fields over large geographic areas during early growth stages.

Automated summary

Early identification of cotton fields is crucial for boll weevil eradication programs. This study proposes a field-level classification method using field boundary data and early-season Sentinel-2 imagery. Machine learning classifiers and dimensionality reduction techniques are used to effectively identify cotton fields.