Evaluation of different spectral indices for wheat lodging assessment using machine learning algorithms

Wheat lodging is a recurrent phenomenon that significantly affects grain yield and impedes the harvesting efficiency. Therefore, the precise and rapid assessment of wheat lodging is crucial in minimizing its impact on grain yield and quality. Recently few studies related to machine learning based wheat lodging have been reported; however, the literature still lacks comprehensive assessments of machine learning algorithms for wheat lodging over Indian agricultural fields. This study presented a systematic approach for detecting the wheat lodging occurred during the end of March and April 2023 in the Ludhiana district of Punjab (India) from multi-temporal Sentinel-2 data using the machine learning algorithms. The ground control points for healthy and lodged areas were collected during March and April 2023. The temporal characteristics of crop phenology from November 2022 to April 2023 were analyzed for wheat classification. The normalized difference vegetation index (NDVI) was computed during this period followed by implementation of random forest (RF), decision tree (DT), and support vector machine (SVM) algorithms to evaluate their performance for wheat classification. It was found that RF outperformed the other models in terms of prediction accuracy and wheat area extraction. To distinguish between lodged and non-lodged wheat, eight spectral indices were computed from the visible and infrared bands of Sentinel-2. These indices were used as inputs to RF, DT, and SVM models. The optimal set of features were identified using random forest feature importance selection approach. Among the spectral indices, spectral sum index (SSI) derived from blue, green, red, and near-infrared bands followed by generalized difference vegetation index (GDVI) accurately separated lodged wheat from non-lodged wheat. Among the three algorithms, the RF model combined with SSI and GDVI achieved the highest overall accuracy of 89.2%. These results suggested that SSI and GDVI derived from Sentinel-2 data coupled with random forest model is effective for assessing the wheat lodging on spatio-temporal scale which may be helpful for developing the decision support system to assess the loss of crop yield loss.