Remote Sensing Classification Study of Grassland Degradation in Inner Mongolia Based on Machine Learning

In the face of the dual challenges posed by global climate change and human activities, nearly 90% of the natural grasslands in Inner Mongolia Autonomous Region have undergone significant degradation. This degradation has resulted in a decline in biodiversity, increased occurrences of dust storms, and various ecological challenges, severely impacting the carrying capacity of the grasslands. The escalating conflict between grassland resources and livestock further compounds these issues, posing threats to regional harmony, stability, and the prospects of sustainable economic and social development. Therefore, gaining a comprehensive understanding of the spatial dynamics of grassland degradation is imperative for effective ecological conservation. To address this need, our study harnesses the power of quantitative remote sensing and machine learning techniques, utilizing nearly two decades of meteorological and multisource remote sensing data spanning from 2000 to 2017 in Inner Mongolia. Through meticulous processes such as sample construction, model training, parameter optimization, and prediction, we have developed a robust grassland degradation classification model. Our evaluation of various machine learning classification algorithms has revealed that K-Nearest Neighbors (KNN) stands out as the most accurate and versatile, closely followed by Support Vector Machine (SVM) and Decision Tree algorithms. Conversely, Naive Bayes exhibits the lowest accuracy among the algorithms assessed. Despite the commendable accuracy achieved by weighted KNN, the classification of saline-alkali areas remains a formidable challenge due to low correlation and the absence of soil data for effective training. In response to these challenges, we have introduced a Soil Adjustment Index to provide insights into surface soil conditions. The implementation of this new classification system has notably improved accuracy, elevating it from 56.4% to an impressive 74.6%. As we look ahead, our future research endeavors will be directed toward addressing the persistent challenges in grassland degradation classification and further enhancing the effectiveness of our methodologies.