Multiple machine learning algorithms were compared to determine the best model of predicting Amur tiger distribution in northeast China, including logistic regression, neural network, maximum entropy, genetic algorithm and ensemble learning (median, principal component analysis (PCA)). Model validation results show that neural network, logistic regression and ensemble median have the highest area under the ROC (receiver operating characteristic) function (AUC) values (0.674, 0.652 and 0.664 respectively). Visual interpretations also show that the three algorithms predict tiger habitat suitability accurately in the study area.
An analytical framework was developed to monitor grassland degradation in Mongolia, Afghanistan, Argentina and Tanzania using Mann Kendall statistic, linear modeling and Fourier seasonal trend analysis.
Based on the decreasing trends in greenness and moisture that are unexplained by rainfall, the median slopes of the greenness and moisture decreasing trends were classified into three categories (low, moderate and high increase) using quantiles. Each category accounts for 33.3% of the pixels that have decreasing trends. And the pixels that do not have decreasing trends were classified as the category of none. Finally the maximum operation was used to combine the greenness trend category and moisture trend category maps, to create the final degradation map
Parallel algorithms were implemented to extract environmental information of river, lake, oasis and desert in a cluster environment consisting of 10 computers. Firstly the images were divided into tiles and sent to the nodes in the cluster environment. After each node finished its processing task, their initial results were sent to the master node. Finally the master node collected and merged all the results from other nodes. Such tiled-based parallel strategy proves to be efficient for many image processing in the remote sensing applications.
Road and building recognition algorithms were developed using Canny edge detector, mean shift image segmentation and nearest neighbor classifier. Road and building are the two most common human-made objects in aerial and satellite images, but due to the complexities of urban environment, it is still a very challenging task. The proposed algorithms utilized both spatial and spectral features extracted from segmented images, and applied the nearest neighbor classifier to recognize roads and buildings. Experiments show the proposed methods are applicable in aerial images, but still needs further improvement.
An automatic drone image stitching algorithm was developed based on Scale Invariant Feature Transform (SIFT) descriptors and drone orientation and attitude parameters. Experiments showed that the proposed algorithm can successfully stitch 70% of the input images in average, which is equivalent to the performance of Microsoft Image Composite Editor. However, the efficiency of the proposed algorithm needs to be further improved.
In order to solve the problems of large memory consumption and low computation speed of SIFT matching algorithm, the proposed algorithm adopts the strategy of pyramid and partitioning to register original images coarsely and matches partitioned images to realize accurate registration. During the process of matching, the number of Gaussian image octave is restrained according to image resolution and the feature points are filtered. Meanwhile, the matching process is paralleled to improve the efficiency. Experiments show that the proposed algorithm ran over 20 times faster than the original SIFT algorithm in a computer with an 8-core Intel CPU.