Artificial intelligence based detection of insect damage in forests using remote sensing data

This thesis aims to predict the damage areas caused by two different insect species on forest trees using Remote Sensing (RS) data and Artificial Intelligence (AI) algorithms. Oak lace bug, Corythucha arcuata (Say 1832) and pine processionary moth, Thaumetopoea wilkinsoni) were identified as the target species. Forest areas damaged by these two insect species in Düzce province were selected as study areas and RS data of these areas were collected. Sentinel-2 satellite images and Unmanned Aerial Vehicle (UAV) data were used as RS data. The thesis aims to detect insect damage areas and forecasting future spread. For the detection and forecast of future spread of the oak lace bug, an ARIMA model was used using 48 months of Sentinel-2 satellite imagery for time series analysis. The validity of the forecasting models was verified with true data and future distribution predictions were made in the next step. For the detection of pine processionary moth damage areas, feature extraction was first performed using UAV photos and e-Cognition software. At this stage, the photos were divided into damaged and healthy regions using multi-resolution segmentation. Texture parameters such as entropy, homogeneity and contrast were calculated using the software's Gray Level Co-occurrence Matrix (GLCM) algorithms based on the reflectance values of the obtained segments. Using the texture parameters entropy and homogeneity, the damaged areas were successfully separated. In the next stage, Convolutional Neural Networks (CNN) based on artificial intelligence in Python programming language were used to detect pine processionary caterpillar damage areas. At this stage, DeepLabV3++ and Unet++ architectures were used as segmentation models and SE-NET, Efficientnet-B6 and Efficientnet-B7 architectures were used as coding networks. For the training of the architectures, datasets were created using semi-automatic segmentation method with e-Cognition software. The classes in the dataset are labelled as damaged, healthy and other. The datasets were designed as two different data inputs, photographs and orthophotos. Each dataset is divided into training and test datasets. To evaluate the accuracy of the segmentation models, confusion matrix, Jaccard index and Intersection over Union (IoU) metrics were used. Among the CNN models used to detect pine processionary moth damage areas on UAV photos, the model with the best overall accuracy (0,90) was Unet++ / SE-NET, while the model with the best overall accuracy (0,81) on UAV orthophotos was DeepLabV3+ / SE-NET. Keywords: ARIMA, Texture parameter, Convolutional neural networks, UAV