陈力'blogs

遥感和深度学习问题

遥感数据具有多种模式:
1.光学(多光谱,高光谱),雷达图像synthetic aperture radar (SAR),成像几何结构和内容完全不同;Remote sensing data are often multi-modal, e.g. from optical (multi- and hyperspectral) and synthetic aperture radar (SAR) sensors, where both the imaging geometries and the content are completely different. (从一种图像模式迁移到另一种图像模式)

2。遥感图像是具有地理信息的。 Remote sensing data are geo-located, such as GIS layers, geo-tagged images from social media, or simply other sensors (as above).

3.遥感图像结合大地测量;(DSM等?)Remote Sensing data are geodetic measurements with controlled quality. This enables us to retrieve geo-parameters with confidence estimates. However, differently from purely data-driven approaches, the role of prior knowledge about the sensors adequacy and data quality becomes even more crucial. For example, to retrieve topographic information, even at the same spatial resolution, interferograms acquired using single-pass SAR system are considered to be more important than the ones acquired in repeat-pass manner.

4.时间信息的维度;The Copernicus program (某计划?)guarantees continuous data acquisition for decades. For instances, Sentinel-1 images the entire Earth every six days.

5.遥感图像的大数据问题;整个地球数据是非常庞大的。

6.在许多情况下,遥感图像的目的是用于检索而不是检测和识别。

—— VGG (scale jittering data augmentation

RS and DL
Hyperspectral Image (高光谱图像)
高光谱传感器具有数百个窄谱带的特点。这种非常高的光谱分辨率使我们能够通过光谱分析识别像素中包含的材料。高光谱数据分析在土地覆盖/利用分类或变化、目标检测等许多实际应用中具有重要意义。
Hyperspectral sensors are characterized by hundreds of narrow spectral bands. This very high spectral resolution enables us to identify the materials contained in the pixel via spectroscopic analysis. Analysis of hyperspectral data is of high importance in many practical applications, such as land cover/use classification or change and object detection.
!!!!!!! e.g., via the launch of EnMAP, planned in 2020, and DESIS, planned in 2017,
List(20 - 38)

SAR 可解释性 (39 - 67)
automatic target recognition (ATR), 自动目标识别
terrain surface classification
parameter inversion

高分遥感图像
场景分类(68-92)BoW (78,79)

多模态数据融合 (DSM 什么等)

3D重构 (DSM , DTM)


遥感图像是遥感数据中的一种;(地球观测技术) 军事,文化等
高纬度,少量可用的标签

问题1
First, many RS data, especially hyperspectral images (HSIs), contain hundreds of bands that can cause a small patch to involve a really large amount of data, which would demand a large number of neurons in a DL network

Apart from the visual geometrical patterns within each band, the spectral curve vectors across bands may also provide important information.
(如何综合使用,也是问题)

问题2,
大样本的问题

问题3,
The high spatial resolution RS images may involve various types of objects, which are also different in size, color, location and rotation.

(不懂,需要复看)
Hyperspectral remote sensors capture digital images in hundreds of continuous narrow spectral bands, producing threedimensional
(3D) hyperspectral imagery (HSI) which simultaneously involves spectral and spatial information. Also, with
high-quality hyperspectral satellite data becoming available (e.g., via the launch of EnMAP, scheduled in 2020, and DESIS,
originally planned for 2017, Zhu et al., 2017) more HSI data will become available. Such very rich spectral information is
potentially very useful to help reveal any interesting unknown content contained within the images. In particular, HSI has been
widely used in a range of practical applications, such as land cover, change detection and object identification. Indeed, applications of HSI data have been one of the most active research directions, with classification of individual pixels in an HSI
image playing a crucial role in such applications.

AVIRIS is a airborne visible/infrared imaging spectrometer which belongs to
the Jet Propulsion Laboratory in the USA. ROSIS is a reflective optics system imaging spectrometer from the National Aeronautics
and Space Agency of Germany. Taking the data set of Indian Pines for instance, it was acquired by the AVIRIS sensor
in Indiana in June 1992. This data set covered 145 lines by 145 pixels, the geometric resolution of which is 20 m.
The original data have 224 spectral in the wavelength range of 400–2,500 nm. The 24 bands covering the region of water absorption were removed to noise. The Indian Pines ground truth contains 16 classes with 10,249 pixels labeled in total.

kappa coefficient(K)新的指标,

光谱特征分类 Spectral feature classification
空间特征分类 Spatial feature classification

  • [ ] NIPS论文修改
    实验的完整性上;

理论上缺乏证明,这是一个硬伤;

公式的定义问题;完善定义;

overall score!

低级错误:

  • What is f() in (3)? No definition is given.
  • l139: Is RS(l, j) a typo of RS(l,i)?
  • Model 1 & 2 seem to be swapped in the caption of Fig.1
  • What is f(x_l,i) in (4)?

  • the para at Line 128,

  • para at Line 233

SVHN, MNIST or Imagenet.

Overall, the problem addressed is significant. However, the significance of the method replies on more empirical
evaluation under different settings/datasets for a more holistic picture.

  • Please check Fig, 1, Model 2 has three low ARS kernels, whereas the caption Model 1 is claimed to have
    them. This is confusing, perhaps Model 1 and 2 are interchanged.
  • All figures are blurred and they need to be clearer. It is difficult to make out details with a blurred figure. One
    suggestion is to vectorize the graphics.
  • Make better use of space in the plots for Fig. 5
  • How is class selectivity formally defined and measured? Please add brief description in the paper.

标签:

专题:

发表于2019-07-24 11:51:24,最后修改于2019-09-10 09:21:47。

本站文章欢迎链接分享,禁止全文转载。


« 上一篇 Hello World

推荐阅读

Big Image