How can we obtain high quality MODIS images by automatic selection?

Cristina Domingo, CREAF (ES)

A review of the ECOPOTENTIAL paper: L. Pesquer, C. Domingo-Marimon, X. Pons (2019), “Spatial and spectral pattern identification for the automatic selection of high quality MODIS images”, Journal of Applied Remote Sensing, 13(1), 014510. doi: 10.1117/1.JRS.13.014510.

One of the favourite instruments of the Remote Sensing (RS) research community is the long-life MODIS (Moderate-resolution Imaging Spectroradiometer) which is collecting data since 2000 on board of the Terra (EOS/AM-1) satellite. The versatility of the instrument, which achieves a fascinating balance between temporal, spectral, spatial and radiometric resolutions, allows interesting finds and analysis of land, ocean and atmosphere processes and results in a better understanding of global changes by monitoring environmental and ecological dynamics. Another reason of the success of MODIS is the large variety of processed products being distributed, from raw radiance and surface reflectance to derived vegetation indices or particulate organic/inorganic carbon. Finally, MODIS data has been always free and open accessible. MODIS is a passive sensor with 36 spectral bands and covers from visible to thermal infrared (0.45 to 14.385 μm 14.385  μm) wavelengths at medium spatial resolution (from 250 m to 1 km at nadir).

However, this dataset can contain images with specific issues that could reduce their quality. For instance, several atmospheric effects interfere with the remotely sensed signal. In particular, cloud coverage or aerosols (pollution, dust, sea salt, etc.) interfere on the measurement of land surface conditions. The Clouds, and their projected shadows over the Earth, represent indeed the major constraint on applications of optical remote sensing. Their automatic detection is still a challenge to overcome, especially when automatic methods for analysing large volumes of data are used. Moreover, MODIS has a singular field of view of 110 degrees that, combined with its distance to the ground, achieves a wide swath covering 2330 km (across track) in one scan, causing a panoramic distortion whereas partially overlapping scans produce data repetition. 

This feature, known as the bow-tie effect, is exacerbated by Earth’s curvature (bands with a spatial resolution of 250 m at nadir have 1207 m at the image edges, while 1 km resolution bands have 4816 m at the image edges). In addition, these data might also have some inaccuracies derived from the different levels of post-acquisition processing, such as detector noise or data faults. Although many of these issues can be filtered out using the MODIS ancillary data, the long-time series of daily images do not facilitate the work performed by technicians of selecting high quality images before its massive analysis. Therefore, advancing in more complete, yet automatic, methods for selecting high-quality images becomes necessary.

Indeed, a range of methods for selecting a subset of high-quality images while preserving their original values already exist, but the advancement achieved by our work consists in providing an automatic selection based on a geostatistical approach together with a tailor-made set of spatial pattern analysis tools. The idea behind this research is that the spatial and spectral patterns of images are key for filtering data. Certainly, among others, geostatistics can provide parameters for describing spatial patterns and modelling uncertainty in spatial analyses. For instance, MODIS images can be analysed using a variogram analysis in order to identify those presenting higher quality. This analysis can be performed using any of the bands of MODIS or some combination of them. However, the spatial pattern characterized by the variogram is dependent on the spectral band, so the band (or bands combination) to be used should not be randomly chosen, but it is recommended to use the one that best describes the structural parameters of the variogram. Therefore, it is important to previously analyse which band results in a subsequent better selection of high-quality images.

ECOPOTENTIAL uses long time series of MODIS data for monitoring the evolution of changes, so there is a clear need of achieving an automatic and massive geoprocessing environment implemented in a chain of GIS and RS algorithms to manage huge series of RS data. ECOPOTENTIAL has worked on the analysis of which band of MODIS provides better results in automatic selection of high quality images. With this aim, we have explored the influence of the spatial pattern of different MODIS spectral bands (bands 01, 02, 03, 04, 06, and 07) and the first component of the principal component analysis (PCA) of two subsets of spectral bands on the variability of the variogram.

A quality filtering by expert and comparison with present methodology have been carried on. The results show that the highest accuracy of automatic selection is achieved using band 01 (93.0%) and band 03 (91.6%), corresponding to the red and blue visible spectrum bands, respectively. On the other hand, the least appropriated bands are band 02 (64.8%) and band 07 (80.3%), corresponding to the NIR1 and SWIR2 regions of the electromagnetic spectrum, respectively. These bands are the most sensitive to phenology and vegetation water content, both of which show strong seasonal variability that decreases the effectivity of the variogram analysis. The combination of bands does not improve the best results of each individual one, the combination merges patterns and it difficult a bit more the automatic selection of high-quality images.

In addition, we propose a protocol adaptable to different study regions in order to select the optimal band to be used in the geostatistical approach. This is especially recommended in those regions with strong seasonal vegetation changes where the methodology will automatically discard the bands presenting large anomalies regarding the global spatial pattern, appointing the band that will obtain a time series of high-quality images. Finally, the analysis and results demonstrate that the approach is suitable for processing huge amounts of RS data, such as data provided by the MODIS Terra (or Aqua) full archive, with very low human dedication to the task and a modest computational effort.