Takashi Yamaguchi

Artificial neural network ensemble-based land-cover classifiers using MODIS data

Terra and Aqua, two satellites launched by the NASA-centered International Earth Observing System project, house MODIS (moderate resolution imaging spectroradiometer) sensors. Moderate-resolution remote sensing allows the quantifying of land-surface type and extent, which can be used to monitor changes in land cover and land use for extended periods of time. In this article, we propose land-surface classification by applying an ensemble technique based on fault masking among individual classifiers in N-version programming. An N-version programming ensemble of artificial neural networks is created, in which the majority vote result is used to predict land-surface cover from MODIS data. It is shown by experiment that an N-version programming ensemble of neural networks greatly improves the classification error rate of land-cover type.

A proposal of interactive growing hierarchical SOM

Self Organizing Map is trained using unsupervised learning to produce a two-dimensional discretized representation of input space of the training cases. Growing Hierarchical SOM is an architecture which grows both in a hierarchical way representing the structure of data distribution and in a horizontal way representation the size of each individual maps. The control method of the growing degree of GHSOM by pruning off the redundant branch of hierarchy in SOM is proposed in this paper. Moreover, the interface tool for the proposed method called interactive GHSOM is developed. We discuss the computation results of Iris data by using the developed tool.

Artificial neural networks paddy-field classifier using spatiotemporal remote sensing data

Monitoring changes in a paddy-field area is important since rice is a staple food and paddy agriculture is a major cropping system in Asia. For monitoring changes in land surface, various applications using different satellites have been researched in the field of remote sensing. However, monitoring a paddy-field area with remote sensing is difficult owing to the temporal changes in the land surface, and the differences in the spatiotemporal characteristics in countries and regions. In this article, we used an artificial neural network to classify paddy-field areas using moderate resolution sensor data that includes spatiotemporal information. Our aim is to automatically generate a paddy-field classifier in order to create localized classifiers for each country and region.

Ensemble of artificial neural network based land cover classifiers using satellite data

Terra and Aqua, 2 satellites launched by the NASA-centered international Earth Observing System project, house MODIS (Moderate Resolution Imaging Spectroradiometer) sensors. Moderate resolution remote sensing allows the quantifying of land surface type and extent, which can be used to monitor changes in land cover and land use for extended periods of time. In this paper, we propose applying an ensemble technique, based on fault masking among individual classifier for N-version programming. We create an N-version programming ensemble of artificial neural networks and use the majority voting result to predict land surface cover from MODIS data. We show that an N-version programming ensemble of neural networks greatly improves the classification error rate of land cover type.

Cluster ensemble in adaptive tree structured clustering

Adaptive tree structured clustering (ATSC) is our proposed divisive hierarchical clustering method that recursively divides a data set into two subsets using self-organising feature map (SOM). In each partition, after the data set is quantised by SOM, the quantised data is divided using agglomerative hierarchical clustering. ATSC can divide the data sets regardless of data size in feasible time. On the other hand the number of cluster and the members of each cluster are not universal in each run. This non-universality is fundamental problem in the other divisive hierarchical clustering and partitioned clustering. In this paper, we apply cluster ensemble to each data partition of ATSC in order to improve universality. Cluster ensemble is a framework by using multiple learners for improving universality. From the computer simulation, we showed that the proposed method is effective for improving universality. Moreover, the accuracy was improved by solving the non-universality of each partition.

Detection limit of fishing boats by the day night band (DNB) on VIIRS

The detection limit of DNB was proposed as a function of the brightness temperature (BT) at 3.7 μm, where the transmittance of cloud could be observed as a change of surface temperature. The shortwave infrared band exhibited a wide distribution in BT more than the thermal infrared band for the same level of DNB radiance. The lights from surface were identified even under the full Moon condition with the proposed method, where clouds were reflecting the lunar lights. A different distribution of clouds for day to day and a change of the Moon phase with its elevation make this problem more complicated. But the approach of contrast based evaluation of surface lights and lunar reflected lights could be one solution to distinguish the lights from the surface. Currently, a validation is necessary in the future to confirm this algorithm and to validate the detected pixels to be fishing boats with the stable light sources. The time series data of fishing boats could be studied to analyze the region of fishing area relative to the distribution of sea surface temperature and/or chlorophyll-a.

Estimation of vessel traffic density from Suomi NPP VIIRS day/night band

Continuous observation of change in vessels activity is important to detect the events and understand the circumstances within the maritime environment. In this paper, we propose the estimation and visualization method of vessel traffic density from day/night band in Visible Infrared Imaging Radiometer Suite (VIIRS) sensor on the Suomi National Polar-orbiting Partnership (Suomi NPP) in order to detect the activity of small vessels that their location data is not reporting by Automatic Identification System (AIS) data. Moreover, we investigate the proposed visualization method in the application on the region of South China Sea.

Application of particle swarm optimization to similar image search on satellite sensor data

Remote sensing using satellite monitored sensor data is one of the most important methods for global environmental monitoring. Similar image search is an important problem in the satellite image data analysis. The similar image search extracts local area images from a given global image. The similar image search using shape features can be used to analyze the cloud type, the volcanic activity, the change in vegetation and etc. However, the similar image search in satellite image data requires the fast computation infrastructure and search method due to the huge image data. In previous research, we proposed a variant of particle swarm optimization that globally searches using particle groups in dynamically changed problem space. In this paper, our PSO was applied to the similar image search problem based on Transfer learning concept. The transfer learning is a meta learning methods that uses the knowledge and data in an domain in order to solve the problem in the another domain. In this experiment, we compared the accuracy and the calculation time among the different transfer learning conditions in order to investigate the possibility of knowledge transfer in similar image search.

Visualization using multi-layered U-Matrix in growing Tree-Structured self-organizing feature map

Self-organizing feature map (SOM) is well known artificial neural network using unsupervised learning for the data visualization and vector quantization. SOM has been used for cluster analysis. On the other hand, SOM cannot produce clarified clusters. And so SOM clustering capability is depends on visualization method. We proposed a variant of SOM that construct hierarchical neural network structure to clarify cluster boundaries in previous research. In this paper, we proposed a visualization method for this growing Tree-Structured SOM and discuss the computational result of Iris data.

SOM for classifying data sets with missing values: application to clinical data of bladder cancer patients

In this paper we investigate applying SOM (Self-Organizing Maps) for classification and rule extraction in data sets with missing values, in particular from real clinical data of bladder cancer patients. For this experiment, we used real data of bladder cancer patients provided by Kitasato University Hospital. When using input data with missing values for SOM, the missing value is either interpolated in the preprocessing stage, or the missing value is replaced with a specific value or property that marks it as a missing value. In either case, there is a possibility some rules can be extracted from data with missing values. On the other hand, these data can have a negative influence for the classification for data sets for which missing values should be neglected. In this research we propose a method where SOM is trained using an input vector in which the properties for the missing values are excluded. The influence of information on the missing values can be reduced by using the proposed method. Through computer simulation, we showed that the proposed method gave good results in classification and rule extraction from clinical data of bladder cancer patients.