Aiming at the problem that most of the land use monitoring in the past uses supervised classification algorithms, which have high costs, wrong leakage points, and greatly affected by human factors, a semi-supervised classification algorithm was proposed for particle swarm optimization probability neural networks, which improved the classification accuracy. The algorithm optimized the parameters of the classifier through the particle swarm optimization algorithm, improved the accuracy of the classifier, and Shannon entropy was used to select high-confidence samples to expand the initial training sample set, a large number of unlabeled samples were expanded to the training sample set, the number of initial label samples were reduced, costs were saved, and it was compared and analyzed with random forest, maximum likelihood method, and probabilistic neural network algorithm, the classification accuracy was improved by 1.25~6.57 percentage points compared with that of other algorithms, and the Kappa coefficient reached more than 0.8. Through the land classification of the remote sensing images of Xinxiang City in 1996, 2004, 2013 and 2020, the results showed that the construction land of Xinxiang City from 1996 to 2020 was continuously expanded in Xinxiang County in the central region, and the cultivated land area was also increased, and the area of other land used was decreased, and the area along the Yellow green area was increased; the land circulation was the most obvious for the conversion of cultivated land to construction land. The research results provided a certain reference for the further rational development of land resources in Xinxiang City.