摘要
遥感指数的形式通常基于经验设定,无论是通过选择特定的反射率波段、方程形式,还是其系数。这些光谱指数常作为目标检测与分类任务前的预处理步骤。然而,目前尚无研究尝试通过函数逼近的方法系统性地探索最优指数形式,以实现分类和/或分割性能的优化。本研究旨在提出一种方法,通过在多种通用方程形式上采用梯度下降的统计优化策略,自动寻找最优遥感指数。基于六波段影像数据,共测试了五种方程形式:线性、线性比值、多项式、通用函数逼近器以及密集形态学模型。同时,还将若干信号处理与图像分析技术整合进深度学习框架中,以增强模型表达能力。为评估标准指数与所提出的DeepIndices的性能,采用两个评价指标:Dice系数(与F1分数类似)和平均交并比(mIoU)。研究聚焦于一种用于近场获取土壤与植被表面的多光谱相机。所构建的DeepIndices在相同植被数据集和评价指标下,与89种常用植被指数进行了对比。以最广泛使用的植被指数NDVI(归一化差异植被指数)为例,其mIoU得分为63.98%;而本研究提出的最优模型通过解析解重构的指数,mIoU提升至82.19%。该性能提升具有显著意义,不仅增强了分割结果的准确性与鲁棒性,使其对各类外部干扰因素更具适应性,同时也有助于更精确地刻画检测目标的几何形态。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 2d-semantic-segmentation-on-deep-indices | Soil Adjusted Vegetation Index | mIoU: 67.28 |
| 2d-semantic-segmentation-on-deep-indices | Modified Chlorophyll Absorption In Reflectance Index 1 | mIoU: 73.68 |
| 2d-semantic-segmentation-on-deep-indices | 5x5 universal-function + ibf + sprb | mIoU: 80.63 |
| 2d-semantic-segmentation-on-deep-indices | 7x7 linear + ibf+sprb | mIoU: 81.49 |
| 2d-semantic-segmentation-on-deep-indices | 5x5 linear-ratio + ibf + sprb | mIoU: 80.08 |
| 2d-semantic-segmentation-on-deep-indices | 7x7 dense-morphological + ibf + sprb | mIoU: 82.19 |
| 2d-semantic-segmentation-on-deep-indices | Global Environment Monitoring Index | mIoU: 65.04 |
| 2d-semantic-segmentation-on-deep-indices | 3x3 dense-morphological + ibf + sprb | mIoU: 80.29 |
| 2d-semantic-segmentation-on-deep-indices | Adjusted Transformed Soil Adjusted VI | mIoU: 64.96 |
| 2d-semantic-segmentation-on-deep-indices | 5x5 polynomial + ibf + sprb | mIoU: 80.67 |
| 2d-semantic-segmentation-on-deep-indices | 1x1 polynomial + ibf | mIoU: 80.03 |
| 2d-semantic-segmentation-on-deep-indices | 7x7 polynomial + ibf + sprb | mIoU: 81.21 |
| 2d-semantic-segmentation-on-deep-indices | NDVI | mIoU: 63.98 |
| 2d-semantic-segmentation-on-deep-indices | Enhanced Vegetation Index 3 | mIoU: 65.05 |
| 2d-semantic-segmentation-on-deep-indices | 3x3 universal-function + ibf + sprb | mIoU: 81.08 |
| 2d-semantic-segmentation-on-deep-indices | 7x7 universal-function + ibf + sprb | mIoU: 80.36 |
| 2d-semantic-segmentation-on-deep-indices | Soil And Atmospherically Resistant VI 3 | mIoU: 65.86 |
| 2d-semantic-segmentation-on-deep-indices | Modified Triangular Vegetation Index 1 | mIoU: 73.71 |
| 2d-semantic-segmentation-on-deep-indices | 1x1 universal-function + ibf + sprb | mIoU: 80.15 |
| 2d-semantic-segmentation-on-deep-indices | Enhanced Vegetation Index 2 | mIoU: 67.94 |
| 2d-semantic-segmentation-on-deep-indices | 7x7 linear-ratio + ibf + sprb | mIoU: 81.35 |
| 2d-semantic-segmentation-on-deep-indices | 5x5 dense-morphological + ibf + sprb | mIoU: 81.92 |
| 2d-semantic-segmentation-on-deep-indices | 1x1 dense-morphological + ibf + sprb | mIoU: 80.00 |