Siam Lawawirojwong

An Enhanced Deep Convolutional Encoder-Decoder Network for Road Segmentation on Aerial Imagery

Object classification from images is among the many practical examples where deep learning algorithms have successfully been applied. In this paper, we present an improved deep convolutional encoder-decoder network (DCED) for segmenting road objects from aerial images. Several aspects of the proposed method are enhanced, incl. incorporation of ELU (exponential linear unit)—as opposed to ReLU (rectified linear unit) that typically outperforms ELU in most object classification cases; amplification of datasets by adding incrementally-rotated images with eight different angles in the training corpus (this eliminates the limitation that the number of training aerial images is usually limited), thus the number of training datasets is increased by eight times; and lastly, adoption of landscape metrics to further improve the overall quality of results by removing false road objects. The most recent DCED approach for object segmentation, namely SegNet, is used as one of the benchmarks in evaluating our method. The experiments were conducted on a well-known aerial imagery, Massachusetts roads dataset (Mass. Roads), which is publicly available. The results showed that our method outperforms all of the baselines in terms of precision, recall, and F1 scores.

Moving Horizon Estimator with Pre-Estimation for crop start date estimation in tropical area

Accurate crop start date estimation is crucial for crop yield forecasting which is important not only for a government but also for agriculture-based or trading companies. The estimation can be done using the Normalized Difference Vegetation Index (NDVI) computed from radiant energy from the crops of interest. The NDVI collected from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra satellite is chosen in this study thanks to its free availability which is suitable for a developing country. In a tropical country as Thailand, the NDVI data is very noisy due to high density of clouds. An appropriate estimation technique must therefore be implemented. In this paper, the NDVI is modelled by a triply modulated cosine function with the mean, the amplitude and the initial phase as state variables. The state and the NDVI of single rice crop in the northeast Thailand are estimated using the Extended Kalman Filter (EKF), the Unscented Kalman Filter (UKF), the Moving Horizon Estimator (MHE) and the Moving Horizon Estimator with Pre-Estimation (MHE-PE). The MHE-PE, recently proposed in the literature, is an optimization-based estimator using an auxiliary estimator to describe the dynamics of the state over the horizon which has been shown to overcome the classical MHE strategy in terms of accuracy and computation time. The EKF and the MHE-PE provide the smallest start date estimation error compared to the others, which is 0 day in mean and 18 days in standard deviation. However, the EKF fail to detect the NDVI of preplant crops and parasite weeds while the MHE-PE does not.

Temporal kNN for short-term ocean current prediction based on HF radar observations

Ocean surface current prediction is at the core of various marine operational routines, including disaster monitoring, oil-spill backtracking, sea navigation and search-and-rescue operations. More accurate prediction can yield significant improvement to the overall system. Most existing short-term prediction methods applied numerical models based on physical processes. In this paper, we propose an alternative approach in predicting the surface current by utilizing temporal k-nearest-neighbor technique, which can predict the future surface current up to 24 hours in advance. Our model incorporates several pre-processing methods, e.g. feature extraction and data transformation, in order to capture the seasonal and temporal characteristics of the HF (high frequency) radar observation data. The developed model was implemented, validated and compared with the existing models using the same historical datasets collected from the HF coastal radar stations located along the Gulf of Thailand. Our experimental results indicate that the proposed model can achieve the highest accuracy among all methods, including ARIMA, exponential smoothing, and LSTM; and satisfy the oil-spill backtracking application requirements. In addition, we found that our system requires little to none maintenance and can easily be adapted to other coastal radar locations where the amount of historical HF radar observations is limited.

Comparative Results of Phenology Obtained from Satellite and Ground Observation Images on Paddy Field

Rice is the most significant economic crops in Thailand, which is required to monitor and estimate cultivated area in a wide region. Satellite images are often used to analyze for classifying the paddy area. To validate the results, it is necessary to have ground collection data. For the ground data, instead of field staffs, an equipment called Field Server (FS) has been installed and used to obtain daily images on the paddy field. In this paper, the comparative results of satellite and FS images are experimented in order to understand the correlation between two platforms. Based on the vegetation indices, satellite and FS images are computed for two phenological curves in an observation period. Two curves are compared by using a simple linear regression method (R squared), which can be described the correlation between each other. The phenological curve obtained from FS images is used as a reference. To determine the crop cycle on the satellite images (MODIS), we found that NDVI is more efficient than Excessive Green (ExG) index.