Object-oriented land use classification based on ultra-high resolution images taken by unmanned aerial vehicle

Unmanned aerial vehicle (UAV) has been increasingly used to aid agricultural production and land management, and this paper investigates the feasibility of using ultra-high resolution UAV images to differentiate land usage. We took a farmland at Dehui of Jilin province as an example and acquired its UAV images. The digital surface model (DSM) and the digital orthophoto map (DOM) of the region was generated using the digital photogrammetry software. We then calculated the regional terrain factors and the vegetation indices, and combined them with the original orthophoto model to construct the baseline images for land use classification. The tool of estimation of scale parameters (ESP) was used to extract the optimal scale for segmentation, and a level of objects was constructed after processing the multi-scale segmentation based on the optimal value of each parameter. Terrain and morphological features of each object was used for the classification. It included two steps. The first one was to perform an object-oriented classification of all the five features based on the random forest algorithm. We used a five-feature classifications to analyze the impact of different features. The first one only used the spectrum feature to classify and the index features, morphological features, terrain features and textural features were added consecutively for further classification. The overall accuracy, kappa coefficient, the omission errors and the commission errors for each feature were compared. In the second step, Boruta feature selection method was applied to the original feature space to obtain an optimal feature subset. Based on the optimal feature subset, land use classification was conducted using the random forest algorithm, naive Bayesian algorithm, logistic regression method and the support vector machine (SVM). Using the same optimal feature subset, the influence of each method on the classification was tested. The results showed that the accuracy of the five feature selection schemes was 93.72%, 97.35%, 96.93%, 97.77%, and 98.04% respectively. Adding morphological features reduced accuracy, while adding other features improved accuracy. The scheme with five features gave the best result. The commission was mainly between the bare land and residential land, and the omissions were mainly among grassland, water canals and roads. The confusion between them was likely to be caused by the similarity of spectral, morphology, textural properties and their similar positions. In this study, the important features for classification were spectrum features, textural features, index features and terrain features, and the least important features were morphological features. There were 72 features passing the Boruta test and forming the optimal feature subset. Based on this feature space, the overall accuracy with the above four different algorithms was 98.19%, 96.79%, 90.22% and 96.23% respectively. The classification using the random forest algorithm gave the best result. In conclusion, adding the terrain features can assist classification of land coverage and improve accuracy. Compared with other algorithms, the random forest algorithm is most robust in classification of land coverage in using high dimensional feature space. © 2020, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.