Determination of spatial distributions of greenhouses using satellite images and object-based image analysis approach

Abstract

In the face of the expected pressure on agricultural production systems with the increasing world population, one of the most suitable options for sustainable intensification of agricultural production is greenhouse activities that allow an increase in production on existing agricultural lands. Greenhouse activities can cause environmental problems at the local and regional scales. Since the primary material used in the covering of greenhouses is plastic, ecological problems are expected in the near future due to the excessive use of plastic. Besides, they may affect the integrity of ecosystems by changing land use and land cover (LULC) into extensive agricultural areas. On the other hand, the economy of many rural regions is supported by greenhouse activities, especially in Mediterranean countries. Moreover, due to the exposure of these structures to floods, especially with climate change effects, producers face economic and social problems. While all these situations make the production system unsustainable, they also endanger the ecology and economy of the region. Thanks to synoptic data acquisition and high temporal resolution, remote sensing images allow periodic agricultural sector monitoring. Considering the positive outcomes and adverse effects of greenhouses, determining greenhouse areas using remote sensing images is essential in providing better management strategies. In that case, monitoring through remote sensing images is the most suitable approach to obtain information about the effects of greenhouses on climate and environment and improve their economic output. Within the scope of this thesis, answers to different questions were sought by using the object-based image analysis (OBIA) approach, which is stated to give better results in the literature to determine greenhouses. OBIA approach consists of mainly three stages which are image segmentation, feature extraction, and image or object classification, and these sections formed the structure of this thesis In the image segmentation step, which is the first step of the OBIA, answers were sought for two crucial questions for the segmentation of plastic-covered greenhouses (PCG). The first of these questions is which of the supervised segmentation quality assessment metrics performs better in evaluating PCG segmentation. An experimental design was formed in which segmentation metrics were evaluated together with interpreter evaluations. At this stage, sixteen different datasets consisting of different spatial resolutions (medium and high spatial resolution), seasons (summer and winter), study areas (Almería (Spain) and Antalya (Turkey)), and reflection storage scales (RSS) (16Bit and Percent) were used. Various segmentation outputs were created using the Multiresolution segmentation (MRS) algorithm. Six different interpreters evaluated these outputs and compared them with the eight segmentation quality metrics. As a result of the evaluations, it was concluded that Modified Euclidean Distance 2 (MED2) was the most successful metric in the evaluation of PCG segmentation. On the other hand, Fitness and F-metric failed to identify the best segmentation output compared to other metrics investigated. In addition, the effects of different factors on the visual interpretation results were analyzed statistically. It was revealed that the RSS is an essential factor in visual interpretation. In detail, it was concluded that when evaluating the segmentation outputs created by using the Percent format, the interpreters were more in agreement and interpreted this data type more efficiently. In the second part of the segmentation phase, how much factors or their interactions affect the greenhouse segmentation was investigated. Approximately 4,000 segmentation outputs were produced from sixteen data sets, and MED2 values were calculated. For each shape parameter in each data set, the values reaching the best MED2 value were determined and statistically tested by analysis of variance (ANOVA). The segmentation outputs calculated from the datasets showed that the optimal scale parameters clustered by taking values close to each other in Percent format and took values in a broader range in 16Bit format. This showed that it would be effortless to determine the most appropriate segmentation outputs obtained from the Percent format. In addition, statistical tests have shown that the segmentation accuracy calculated from different RSS formats is directly dependent on the shape parameter. While segmentation accuracy increases with decreasing shape parameters in Percent format, this is the opposite in 16Bit format. This situation revealed that the shape parameter selection is critical depending on the RSS. In summary, it has been revealed that the Percent format is the appropriate data format for PCG segmentation with the MRS algorithm, and in addition, low-shape parameters should be preferred in the Percent format. In the second stage of the thesis, it was hypothesized that different feature space evaluation methods and feature space dimensions affect the classification in terms of accuracy and time. Based on this hypothesis, 128 features were obtained from Sentinel-2 images of the Almería and Antalya study areas, and classification performance was evaluated by random forest (RF) algorithm by applying different feature space evaluation methods. As a result of this evaluation, it was seen that the reduction of the feature space has a direct effect on the accuracy. But moreover, it has been determined that reducing the size of the feature space significantly reduces the time required to run the classification algorithm. Therefore, among the examined feature space evaluation algorithms, it has been concluded that RF and Recursive Feature Elimination (RFE)-RF (RFE-RF) algorithms are more suitable for classification accuracy and the time required to run the algorithm. Moreover, it has been found that these algorithms are less dependent on feature space variation in terms of classification accuracy, but reducing the feature space significantly reduces the computation time. In addition, among a total of 128 features obtained from the segments, including spectral, textural, geometric features and spectral indices, Plastic GreenHouse Index (PGHI) and Normalized Difference Vegetation Index (NDVI) were the most relevant features for PCG mapping according to RF and RFE-RF methods. As a result, the necessity of including indices such as PGHI and NDVI in the feature space and the application of one of the feature space evaluation methods such as RF or RFE-RF in terms of reducing the calculation time are the main outputs of this stage. In the third and final stage of the thesis, the effectiveness of ensemble learning algorithms for the PCG classification has been tested. According to the experimental results, Categorical boosting (Catboost), RF, and support vector machines (SVM) algorithms performed well in both studied areas (Almería and Antalya), but the implementation time required for CatBoost and SVM is higher than all other algorithms studied. K-nearest neighbor (KNN) and AdaBoost algorithms achieved lower classification performance in both study areas. In addition to these algorithms, the light gradient boosting machines (LightGBM) algorithm achieved an F1 score of over 90% in both study areas in a short time. In summary, considering the computation time and classification accuracy, RF and LightGBM are the two up-front algorithms. In general, within the scope of this thesis, answers to the questions encountered in the three steps of OBIA were sought to reach the best PCG determination approach. The determination of greenhouses from satellite images was carried out in two essential study areas in the Mediterranean Basin, where greenhouse activities are intensively carried out. Although these outputs belong to selected test sites, they provide important outputs for generalizing the findings on a large scale. Determining the spatial distribution of PCG to minimize the negative effects on the environment and increase their economic returns will make an important contribution to planners and decision-makers in achieving sustainable agriculture goals.