Detecting surface residual film coverage rate in pre-sowing cotton fields using pixel block and machine learning

Abstract: Residual film pollution has posed a great threat to the agricultural environment in recent years. It is a high demand for residual film identification and coverage rate detection in cotton fields before sowing. In this study, an evaluation method of residual film coverage rate was proposed using pixel block and machine learning model, in order to effectively recognize the residual film in the cotton field before sowing by the pixel block classification. Fifty images of 1 m×1 m cotton field surface with the residual film were collected by random sampling in the cotton growing area of Kuitun, Xinjiang, China. The image with a resolution of 4 608×3 456 (Pixel) was cropped along the boundary of 1 m×1 m sampling area. After cropping, the image was resized to 1 000×1 000 (Pixel), and the brightness was corrected using normalization. The image pixel was then manually labelled, in which the residual film was labelled as 1, and the soil background was labelled as 0. Then, 45 images were randomly selected to train the models, and the rest 5 images were used to verify the final model for the evaluation of residual film coverage rate. The recognizing pixel blocks and machine learning were selected to make better use of the color and texture features of images. Each image was cut into the 10 000, 2 500 and 625 pixel blocks, according to the sizes of 10×10, 20×20 and 40×40 (Pixel), respectively. The extraction was performed on the first, second, and third order color moments of R, G and B channels, and Gray-level Co-occurrence Matrix (GLCM) of each pixel block. Meanwhile, the intensities of R, G and B channels of each pixel were extracted for comparison. The stochastic down-sampling and Synthetic Minority Oversampling Technique (SMOTE) were used to equalize the pixel block data. Principal Component Analysis was employed to extract the top 10 principal components of pixel block, in order to prevent the over-fitting for the high training speed. Consequently, 70% of the data was used for the training, and 30% was for the testing. Random Forests (RF), Xtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Artificial Neural Network (ANN) were used to optimize the parameters via the style search and cross validation. The residual film coverage rate was calculated to evaluate the segmentation of different sizes of pixel blocks and the machine learning models. The ANN model combined with 20×20 (pixel) blocks performed the best, with the Mean Intersection Over Union (MIOU) of 71.25%. The relative error was 0.51% in the residual film coverage rate between the prediction and actual value, and the detection time was 0.29 s. Therefore, the improved model is feasible for the accurate identification of residual film on the surface of the cotton field before sowing. This finding can provide theoretical support for the rapid detection system of residual film pollution using UAV imaging.