Classification of main irrigated crop types towards sustainable development using Landsat and Sentinel data by GIS and remote sensing techniques in a semi-arid area in Tashkent Province, Uzbekistan
von Elbek Erdanaev
Datum der mündl. Prüfung:2023-02-07
Erschienen:2024-02-02
Betreuer:Prof. Dr. Martin Kappas
Gutachter:Prof. Dr. Martin Kappas
Gutachter:Prof. Dr. Gerhard Gerold
Dateien
Name:PhD_Thesis_EE_MK_Final_21122022.pdf
Size:6.02Mb
Format:PDF
Zusammenfassung
Englisch
As the world population increases and cropland expansion occurs, there will be a high need in the food supply soon which will require higher agricultural yields. Crop yield estimation, management, and production assessments at the regional and country-level are very important in Uzbekistan which requires supplemental spatial data that provides timely information on crop type's spatial distribution, condition, and potential yields. Crop-type identification at the local and regional level is very important in agricultural regions in developing countries where it contributes the main share of the country's GDP. Nowadays the number of satellites and free availability of these data with the integration of multi-sensor images offers coherent time series which gives new opportunities for land cover and crop type classification. Poor or developing countries compile their agricultural statistics in tabular form by their provincial administrative areas, which gives no information about the exact locations where specific crops are cultivated. Such data is poorly suited for early warning and assessment of crop production. 5-Daily Sentinel and 16-Daily Landsat satellites image time series of Tashkent Province, Uzbekistan, acquired in 2018 in combination with reported crop area statistics were used to produce the required crop types map. Three well-known machine learning algorithms Support Vector-Machine, Random Forest, and Maximum Likelihood classifications were used to derive crop types maps and compared for recommended suitable methods. Four indices NDVI, EVI, NDWI1, and NDWI2 were calculated using blue, green, near-infrared, SWIR 1, and SWIR 2 bands and used as input data. Firstly, based on the literature review it was found that only limited research was carried out to identify irrigated croplands by crop types at the provincial level. Most of the available land use land cover maps have low resolution and classified crop types as croplands or agriculture. Besides, we have not found any research which compares derived crop types area with official state statistics at the provincial level in the study area. Thus, it is very important to recommend an accurate and timely crop types 2 mapping method for the local land control and management authorities and policy makers. Preliminary climate change analysis over the 35 years of 1979 through 2013 demonstrated the increasing trend of temperature and decreasing trend of precipitation over the croplands, pasturelands, and grasslands of the study area. Precipitation decrease in the study area may reduce plant productivity and temperature increase may have either positive or negative influence on plant production due to more evaporation than precipitation. Besides, expansion of agricultural irrigated cropland and population increase has a significant influence on land-use intensity. Secondly, a comparison of three classifiers algorithms SVM, RF, and MLC performance was studied and the result showed SVM and RF classifiers produced a visually pleasant and realistic irrigated cropland map in the research area. Accuracy assessment results showed that SVM yielded the highest OA and KA. KA of classified images for SVM were 0.90 and 0.89 for the RF algorithm. Both performed well and achieved identical close values. But MLC showed a lower result of KA 0.60. Thirdly, further analysis of testing different indices (NDVI, EVI, NDWI1, and NDWI2) with recommended SVM and RF classifiers using Sentinel-2 and Landsat-8 sensors data were carried out. The results of OA, KA, UA, and PA have shown that RS imagery from both sensors is of comparable quality. But the differences in accuracy results vary higher based on the vegetation indices used than on sensor data. KA values vary between 75% to 88 % in all indices. The lowest KA values were achieved in all indices with the SVM classifier of L8 sensor data. The highest KA values 88% and 87% were achieved with the RF classifier of L8 data when EVI and EVI-NDVI were used respectively. Using NDWI 1 and NDWI 2 which uses SWIR 1 and SWIR 2 bands is not achieved good results in both accuracies point and area comparison and it is not recommended for irrigated crop types mapping. Fourthly, the difference between remote sensing derived classified maps area and officially recorded statistic crops area was compared. It was found that the smallest absolute weighted average value difference of 0.2 thousand ha was obtained using EVI-NDVI with RF method and NDVI with SVM method of Landsat 8 sensor data. For Sentinel 2 sensor data, the smallest absolute value difference result of 0.1 thousand 3 ha was obtained using EVI with RF method and 0.4 thousand ha using NDVI with SVM method. Finally, it can be recommended that using medium 30 m resolution Landsat sensor data is sufficient for mapping irrigated crop types over the study area, and since the launch of this satellite in 1972, historical irrigated cropland mapping is possible for the period up to today. Besides, the recent successful launch of Landsat-9 will successfully continue the Landsat data suite and enable new opportunities in the joint use of Landsat and Sentinel data to capture high temporal resolution during the vegetation growth period which helps to distinguish other minor crop types as well as increase classification accuracy. Classified irrigated crop types maps can benefit regional land management administration offices to monitor the spatial extent of crops location and its monitoring as well as modeling and predicting crop yields and production by different agroecological models. And also, the use of remote sensing-based irrigated crop types data periodically to monitor and evaluate agricultural land uses which can save time, effort, and capital that are needed for traditional human-based ground surveys.
Keywords: Land use land cover classification; Landsat; Sentinel; machine learning algorithms; crop types classification; support-vector machine; random forest; maximum likelihood classification