Optimization of Sample Construction Based on NDVI for Cultivated Land Quality Prediction

The integrated use of remote sensing technology and machine learning models to evaluate cultivated land quality (CLQ) quickly and efficiently is vital for protecting these lands. The effectiveness of machine-learning methods can be profoundly influenced by training samples. However, in the existing research, samples have mainly been constructed by random point (RPO). Little attention has been devoted to the optimization of sample construction, which may affect the accuracy of evaluation results. In this study, we present two optimization methods for sample construction of random patch (RPA) and area sequence patch (ASP). Differing from RPO samples, it aims to include cultivated land area and its size into sample construction. Based on landsat-8 Operational Land Manager images and agricultural land grading data, the proposed sample construction methods were applied to the machine learning model to predict the CLQ in Dongtai City, Jiangsu Province, China. Four machine learning models (the backpropagation neural network, decision tree, random forest (RF), and support vector machine) were compared based on RPO samples to determine the accurate evaluation model. The best machine learning model was selected to compare RPA and ASP samples with RPO samples. Results determined that the RF model generated the highest accuracy. Meanwhile, a high correlation was noted between the cultivated land area and CLQ. Thus, incorporating cultivated land area in the sample construction attributes can improve the prediction accuracy of the model. Among the three sample construction methods, the ASP yielded the highest prediction accuracy, indicating that the use of a large, cultivated land patch as the sample unit can further elevate the model performance. This study provides a new sample construction method for predicting CLQ using a machine learning model, as well as providing a reference for related research.