An automated method for estimating fractional vegetation cover from camera-based field measurements: Saturation-adaptive threshold for ExG (SATE)

Fractional vegetation cover (FVC) is a crucial metric for assessing vegetation cover on the Earth’s surface. The excess green index (ExG), derived from visible true-color RGB images, is widely recognized as a reliable metric for identifying green vegetation. However, the threshold to distinguish vegetation from background via ExG is highly sensitive to variations in illumination, limiting its robustness in real-world applications. Traditional thresholding methods, such as the bimodal thresholding method, maximum entropy thresholding method, and Otsu’s method, perform well under uniform illumination conditions but often fail to achieve high vegetation identification accuracy in scenarios with uneven illumination. Previous studies have shown that saturation (S) is strongly correlated with illumination intensity and can serve as an effective indicator of illumination variations. Specifically, under strong illumination conditions, both vegetation and non-vegetation appear more vivid, resulting in higher S values. For vegetation, the green-band digital number (DN) value increases more sharply than that of the red and blue bands, resulting in a notable rise in ExG. In comparison, non-vegetation like soil shows only a slight green-band increase, producing a smaller ExG gain. This contrast in ExG between the two surfaces becomes more distinct, so a higher segmentation threshold is required. Conversely, weak illumination conditions lead to lower S values and more uniform DN reductions across surface types, which diminishes ExG contrast and necessitates a lower threshold. Building upon this insight, this study introduced a novel method for automated vegetation coverage extraction: the saturation-adaptive threshold for ExG (SATE). SATE dynamically determines the optimal segmentation threshold for ExG on a pixel-by-pixel basis on the S value, then identifies vegetation pixels by comparing the ExG value of each pixel with its corresponding threshold, and finally calculates the FVC, thereby enhancing the adaptability to diverse illumination conditions. To validate its effectiveness, SATE was tested using 100 high-resolution unmanned aerial vehicle (UAV) red‒green-blue (RGB) images collected from five diverse regions across China, covering a range of illumination conditions, vegetation types, and complex land cover scenarios. The experimental results demonstrated that SATE can effectively address the challenges posed by uneven illumination, achieving an average vegetation recognition accuracy of 91–94 %. For vegetation identification, the performance of SATE combined with ExG surpassed that of traditional thresholding methods, including the bimodal thresholding method (86.4 %), maximum entropy thresholding method (67.0 %), and Otsu’s method (66.5 %). Moreover, SATE combined with ExG achieved an accuracy comparable to that of the manual thresholding method (95 %) while eliminating the need for subjective intervention, thus enhancing the automation and practical applicability. In addition, SATE outperformed other local adaptive thresholding methods, including Sauvola (65.3 %), Niblack (58.7 %), OpenCV mean (56.7 %), and OpenCV Gaussian (55.6 %), achieving a higher and more stable accuracy (94.1 %) under varying conditions. Under changing illumination conditions, SATE exhibited superior performance in terms of automation, stability, and applicability, making it a highly efficient and easily deployable solution for automated FVC estimation in camera-based field measurements, particularly in complex land-cover scenarios.