Study area The TRB is situated in the transboundary Teesta sub-catchment of the northern portion of Bangladesh, encompassing about 3684 km2), crossing over the five main districts (Largest administrative unit) such as Nilphamary, Lalmanirhat, Kurigram, Rangpur, and Gaibandha districts. In Bangladesh part, the TRB basin is positioned between the geographical locations of latitudes of 25030′ 02′′ N-157 26018′ 37′′ N, and 88052′ 58′′ E − 89045′ 34′′ E longitudes. From the geomorphologic perspective, the floodplain area is the largest geomorphic unit in Bangladesh, and the drainage pattern comprises of several small rivers, which are spread over elevation ranging from 05 to 110 m. Morphological point of view, the depression of the TRB is shallow and in the moribund river valley, which alters of long morphologically in the pathways of the river. Flash floods are a common event in each year. Subsequently, the flash flood occurred, which leads to destroying a tremendous amount of farmland and more than 20,000 houses (Akhter et al. 2019). The hydrological characteristics of the study area are relatively complex. Geological point of view, the TRB is located under the Bengal basin, on the Rangpur saddle of the Indian platform. The new floodplain deposits such as silt, clay, fine to medium sand, are related to this basin (Islam et al. 2014). The subtropical monsoonal climate with two distinct dominant seasons, namely, monsoon (June to September) and dry season (October to May), are the major features of this basin. The mean annual precipitation is greater than in 1900 mm (Islam et al., 2020b), and responsible for more than 75% of the total annual precipitation mostly occurred in the monsoon rainy season. The mean annual temperature in the TRB during monsoon and dry seasons is nearly 35 ?C and 15 0C, respectively. 2.2. Data sources and preprocessing The different Landsat sensors acquired from the Earth Explorer of the United States of Geological Survey (USGS) (https://earthexplorer.usgs. gov/), including Landsat 4–5, the thematic mapper (TM) for the years of 2000, enhance thematic mapper plus (ETM + ) for the year of 2010 and operational land imager (OLI) for the year of 2019. It is also important to use a cloud-free scene. The images obtained almost or in the same dates is the basic factor for spatiotemporal analyses of LULC change. This eradicates the impacts of seasonal variations when assessing year-to-year variation. The same dates of images are often utilized because it minimizes the inconsistencies in reflectance triggered by the seasonal vegetation changes, the climatic changes, and the sun angle differences (Singh, 1989). 2.3. Method for image classification We classified into six LULC classes for this study, such as the water body, agricultural land, vegetation, sand bar, bare land and built-up area. In this study, we used the artificial neural network algorithm (ANN) for LULC mapping for the years 2000, 2010, and 2019. The ANN is an information-based model and can simply define as the large number of a simple, interrelated processor (neurons), which are managed in several layers working in a parallel within a network. This automatic programming or “learning” is accomplished through the dynamic adjustment of the network interconnection strengths, which associated with each neuron. 2.7. Methods for spatial trend analysis For assessing the pixel-wise trend of fragmentation parameters, we applied the least square regression (y = a + bx) model in spatial scale by incorporating all the individual fragmentation parameters for whole study periods. This method was applied for detecting trends at pixel scale by Paul and Pal (2019), Debanshi and Pal (2020) to explore the trends of surface water depth in wetlands. we estimated the trend of fragmentation status of different fragmentation indices based on 19 years of data by the following equation (2). Y = α + βX (2) where, Y represents the trend value of the time series data, a indicates intercept, β represents slope, and X means time series VHI images. The detailed calculation of intercept and slope can be obtained from Debanshi and Pal (2020). Decision tree-based sensitivity analysis: Decision trees (DT) are considered as a non-parametric supervised machine learning method, which has been employed for classification and prediction (Nefeslioglu et al., 2010). In general, two types of DT have been used for modelings, such as classification trees and regression trees. For predicting a discrete variable, the classification trees have been employed; on the other hand, for predicting a continuous variable, regression trees have been utilized. The basic idea behind this is to construct a model that predicts the value of a dependent factor by learning several decision rules derived from the whole data (Yeon et al., 2010).