Maize is a key global food crop, with China being a major producer vital for global maize supply and food security. Accurately analysing the relationships between grain yields of maize and influencing factors is crucial for enhancing crop production, evaluating arable land quality, and optimizing planting structure. However, when modelling those relationships, the coefficients of each influencing factor vary spatially and have different spatial scales, suggesting that the importance of the influencing factors is multiscale. Traditional global and local modelling methods such as Ordinary Least Squares (OLS) and Geographically Weighted Regression (GWR) models cannot accurately explain those multiscale spatial relationships. The Multiscale GWR (MGWR) model, an extension of GWR, addresses these limitations by allowing each explanatory variable to have a unique spatial scale. By aligning the neighbourhood structure of each variable with its corresponding spatial scale, MGWR improves the accuracy of local regression coefficient estimations, providing a more refined analysis of spatial heterogeneity. In this paper, the multiscale importance of influencing factors on maize grain yield of the Chinese mainland was apportioned via MGWR model. Our findings verified that the relationships between maize grain yield and influencing factors differ at multiple spatial scales. MGWR model can comprehensively apportion the importance of influencing factors at multiple scale, while global and local modelling methods provide biased estimations, with OLS method leaving large residues and GWR model attributing part contribution to spatially varying intercept terms. With the MGWR model, organic fertilizer and terrain aspect are globally important and their relationships with yields keep stationary; the relationships between yields and soil pH value, GDP, DEM and slope vary on a medium-scale, presenting obvious regional differences; cultivation convenience and hydrothermal conditions affect yields at small scales. The comprehensive apportionment of multiscale relationships is an important guideline for the scientific management of agriculture and arable land resources.

Accurately predicting spatially continuous daily air temperature (Ta) is critical for agriculture, environmental management, and ecology. While meteorological stations provide precise Ta data, their spatial coverage is limited. Remotely sensed land surface temperature (LST), often fused with meteorological data, offers broader spatial coverage but struggles due to complex relationships between Ta and LST, influenced by factors like topography and human activities. Traditional supervised learning methods often fail to capture the spatial autocorrelation and heterogeneity inherent in the relationships, indicating the need for a more robust approach that integrates geographic knowledge. This study proposes the spatially varying coefficients random forest (SVCRF) model, to integrate exploratory spatial data analysis (ESDAs) into random forest (RF) to capture spatially nonstationary relationships. It first stratifies the study area based on bivariate Local Indicators of Spatial Association and geographical detector, then builds several spatial RFs with specific spatial positions and extent. In each spatial RF, the distance from observation/prediction sites to its position is added as a key predictor variable to model the local spatial variations of the relationships within the spatial extent. Applied to daily Ta mapping at 1 km resolution across China using data from 5425 meteorological stations, the SVCRF model demonstrated superior accuracy, achieving root-mean-squared error (RMSE) of$1.315~^{\circ }$C and mean absolute error (MAE) of$1.014~^{\circ }$C. Compared to RF, regression kriging (RK), and geographically weighted regression (GWR), it reduced MAE by$0.351~^{\circ }$C,$0.786~^{\circ }$C, and$0.831~^{\circ }$C, respectively. The model also offers high interpretability, with uncertainty estimates aligning with actual errors and spatially resolved variable importance highlighting spatial patterns.

Change detection, a critical and flourishing Earth observation technology, aims to identify changes through cross-temporal remote sensing images acquired over the same geographical area. With the widespread use in various change scenarios, it becomes essential to utilize heterogeneous images due to the high challenge of accessing the ideal homogeneous images. Nevertheless, domain shift, generated by different imaging factors (e.g., sensors, seasons, atmosphere, illumination), makes it unable to compare the heterogeneous images directly. To address this problem, we propose a deep domain adaptation and disentangled representation network for unsupervised heterogeneous change detection (DADR-HCD), which bridges the domain gap from the perspective of causal mechanisms and compares the differences in the content feature space. In the training stage, the deep features of the input bitemporal images are further disentangled into the domain-invariant (content) features and domain-specific (style) features through an explicit image translation network. Furthermore, unlike comparing the differences at the image level or deep feature space, the change probability maps are directly calculated based on the content feature similarity in the prediction stage, which minimizes the style noise and avoids the asymmetry of image-level translation. Finally, the binary change maps are obtained using threshold segmentation and morphological post-processing strategies. The comprehensive experimental results and detailed analysis on five typical datasets demonstrate the effectiveness and superiority of the proposed DADR-HCD network in the unsupervised heterogeneous change detection task.

Accurate information on the spatial and temporal distribution of abandoned cropland (AC) is crucial for protecting arable land, and maintaining regional food security and ecological stability. Nevertheless, the unavailability of dedicated monitoring for AC, along with the extended time frame required for remote sensing surveillance and the intricate transformation of land cover types following abandonment, poses considerable challenges in producing accurate AC maps for scientific purposes. To address these challenges, a new framework for AC identification was proposed. Specifically, the change detection was performed using the phenology-enhanced change vector analysis (PECVA) method, followed by a co-training semi-supervised classification method based on historical samples and sparse samples (HS-SSC) to classify the land cover type in change areas and obtain long-term land cover mapping results. Next, a land cover type change detector method was applied to identify the location and time of AC occurrences. Sufficient comparative experiments and accuracy validation were conducted to confirm the effectiveness of the PECVA and HS-SSC. The proposed method was verified in areas with single and double cropping per year, with average precision rates of 85.40% and 83.57% respectively for multi-year AC identification. This study offers a promising tool for identifying AC, which can aid in AC recultivation and serve for arable land conservation and land resource management. Our code is available at https://github.com/zhangtingting114/ACI.

Agricultural land abandonment and retirement are important and lead to different types of land use and cover change. Generally, abandonment and retirement are caused by different social and environmental factors and result in different ecological and economic benefits and costs. Faced with the complexity of agricultural land change over time, this study aims to develop a new framework to distinguish between agricultural land abandonment and retirement and detect the extent and exact year of abandonment and retirement using Google Earth Engine (GEE). We tested our approach for three typical regions in the northern China crop-pasture band, where agricultural land abandonment and retirement are widespread. First, based on the spectral features obtained from Landsat images, annual land-cover maps were obtained with sample migration and random forest from 1998 to 2019 (0.87 overall accuracy). Second, a temporal consistency check method was proposed to further improve the classification performance (0.92 overall accuracy). Third, a trajectory-based change detection approach was developed to identify abandonment and retirement (F1 score for abandonment: 0.74, retirement: 0.83). Our results indicate that the spatiotemporal patterns of abandonment and retirement in the study area greatly differed. Overlapping of the topography and climate data showed that agricultural land with steep slopes ($> 10^{\circ }$) was more likely to be retired and that abandonment was more likely to occur in areas with less precipitation. Overall, the methods used herein are robust for agricultural land abandonment and retirement monitoring and may be extended to other land-cover change studies.