Soil saline patches, as a direct and early indicators of soil salinization degradation, pose a serious threat to agricultural productivity and land sustainability in coastal and arid areas. However, most existing studies focus on regional-scale classification using medium- and low-resolution imagery, lacking fine-grained, high-precision identification of soil salinity patches. In this study, we propose a semi-supervised knowledge integration (SSKI) framework for the accurate extraction of soil saline patches using unmanned aerial vehicle (UAV) RGB imagery. To address the time-consuming and labor-intensive problem of sample labeling, a dual-sample optimization strategy is designed. It simultaneously enhances the diversity of labeled samples via color and spatial augmentation, and selects reliable pseudo-labels for unlabeled samples through sample stability evaluation. Moreover, considering the unique characteristics of soil saline patches, a multimodal information learning network is proposed, which integrates multi-scale soil saline patch information, spectral-ecological features, and texture-semantic joint information, aiming to enhance boundary precision and inter-class discrimination. We validate the proposed method using a newly constructed UAV dataset (DongY) covering approximately 16 km2 of saline-affected farmland in Dongying, China in 2022 and 2024. Experimental results show that SSKI consistently outperforms state-of-the-art methods. In particular, the multi-year, UAV-based saline patch identification demonstrates the strong temporal generalization capability of the proposed method. Furthermore, by mapping the spatial distribution of soil saline patches, the study provides a basis for assessing the impact of salinization on crop growth and offers essential data support for the management and remediation of saline-alkali land.