Building footprints are among the most predominant features in urban areas, and provide valuable information for urban planning, solar energy suitability analysis, etc. We aim to automatically and rapidly identify building footprints by leveraging deep learning techniques and the increased availability of remote sensing datasets at high spatial resolution. The task is computationally challenging due to the use of large training datasets and large number of parameters. In related work, You-Only-Look-Once (YOLO) is a state-of-the-art deep learning framework for object detection. However, YOLO is limited in its capacity to identify small objects that appear in groups, which is the case for building footprints. We propose a LOcally-COnstrained (LOCO) You-Only-Look-Once framework to detect small and densely-distributed building footprints. LOCO is a variant of YOLO. Its layer architecture is determined by the spatial characteristics of building footprints and it uses a constrained regression modeling to improve the robustness of building size predictions. We also present an invariant augmentation based voting scheme to further improve the precision in the prediction phase. Experiments show that LOCO can greatly improve the solution quality of building detection compared to related work.

Tree management is becoming a big issue in a variety of societal domains. In recent years, historic wildfires and blackouts caused by failures in tree management have increased in both quantity and severity, resulting in many deaths and financial loses in the tens of billions of dollars. Many communities are also suffering from massive tree loss (e.g., in the millions) that affects the health and well-being of citizens. These problems are likely to worsen due to climate change, aging infrastructure and population growth. Tree management needs a revolution to deal with these urgent problems. This opens up new challenges and opportunities for the spatial community. This paper presents some of the open research problems from the perspectives of individual tree mapping and characterization as well as decision making and in-field intervention.

Class ambiguity refers to the phenomenon whereby similar features correspond to different classes at different locations. Given heterogeneous geographic data with class ambiguity, the spatial ensemble learning (SEL) problem aims to find a decomposition of the geographic area into disjoint zones such that class ambiguity is minimized and a local classifier can be learned in each zone. The problem is important for applications such as land cover mapping from heterogeneous earth observation data with spectral confusion. However, the problem is challenging due to its high computational cost. Related work in ensemble learning either assumes an identical sample distribution (e.g., bagging, boosting, random forest) or decomposes multi-modular input data in the feature vector space (e.g., mixture of experts, multimodal ensemble) and thus cannot effectively minimize class ambiguity. In contrast, we propose a spatial ensemble framework that explicitly partitions input data in geographic space. Our approach first preprocesses data into homogeneous spatial patches and uses a greedy heuristic to allocate pairs of patches with high class ambiguity into different zones. We further extend our spatial ensemble learning framework with spatial dependency between nearby zones based on the spatial autocorrelation effect. Both theoretical analysis and experimental evaluations on two real world wetland mapping datasets show the feasibility of the proposed approach.

Given remote sensing datasets in a spatial domain, we aim to detect geospatial objects with minimum bounding rectangles (i.e., angle-aware) leveraging deep learning frameworks. Geospatial objects (e.g., buildings, vehicles, farms) provide meaningful information for a variety of societal applications, including urban planning, census, sustainable development, security surveillance, agricultural management, etc. The detection of these objects are challenging because their directions are often heavily mixed and not parallel to the orthogonal directions of an image frame due to topography, planning, etc. In addition, there is very limited training data with angle information for most types of objects. In related work, state-of-the-art deep learning frameworks detect objects using orthogonal bounding rectangles (i.e., sides are parallel to the sides of an input image), so they cannot identify the directions of objects and generate loose rectangular bounds on objects. We propose an Unsupervised Augmentation (UA) framework to detect geospatial objects with general minimum bounding rectangles (i.e., with angles). The UA framework contains two schemes, namely a ROtation-Vector (ROV) based scheme and a context-based scheme. The schemes completely avoid the need for: (1) additional ground-truth data with annotated angles; (2) restructuring of existing network architectures; and (3) re-training. Experimental results show that the UA framework can well approximate the angles of objects and generate much tighter bounding boxes on objects.

Tree inventories are important datasets for many societal applications (e.g., urban planning). However, tree inventories still remain unavailable in most urban areas. We aim to automate tree identification at individual levels in urban areas at a large scale using remote sensing datasets. The problem is challenging due to the complexity of the landscape in urban scenarios and the lack of ground truth data. In related work, tree identification algorithms have mainly focused on controlled forest regions where the landscape is mostly homogeneous with trees, making the methods difficult to generalize to urban environments. We propose a TIMBER framework to find individual trees in complex urban environments and a Core Object REduction (CORE) algorithm to improve the computational efficiency of TIMBER. Experiments show that TIMBER can efficiently detect urban trees with high accuracy.

Class ambiguity refers to the phenomenon whereby samples with similar features belong to different classes at different locations. Given heterogeneous geographic data with class ambiguity, the spatial ensemble learning (SEL) problem aims to find a decomposition of the geographic area into disjoint zones such that class ambiguity is minimized and a local classifier can be learned in each zone. SEL problem is important for applications such as land cover mapping from heterogeneous earth observation data with spectral confusion. However, the problem is challenging due to its high computational cost (finding an optimal zone partition is NP-hard). Related work in ensemble learning either assumes an identical sample distribution (e.g., bagging, boosting, random forest) or decomposes multi-modular input data in the feature vector space (e.g., mixture of experts, multimodal ensemble), and thus cannot effectively minimize class ambiguity. In contrast, our spatial ensemble framework explicitly partitions input data in geographic space. Our approach first preprocesses data into homogeneous spatial patches and uses a greedy heuristic to allocate pairs of patches with high class ambiguity into different zones. Both theoretical analysis and experimental evaluations on two real world wetland mapping datasets show the feasibility of the proposed approach.

Given learning samples from a raster data set, spatial decision tree learning aims to find a decision tree classifier that minimizes classification errors as well as salt-and-pepper noise. The problem has important societal applications such as land cover classification for natural resource management. However, the problem is challenging due to the fact that learning samples show spatial autocorrelation in class labels, instead of being independently identically distributed. Related work relies on local tests (i.e., testing feature information of a location) and cannot adequately model the spatial autocorrelation effect, resulting in salt-and-pepper noise. In contrast, we recently proposed a focal-test-based spatial decision tree (FTSDT), in which the tree traversal direction of a sample is based on both local and focal (neighborhood) information. Preliminary results showed that FTSDT reduces classification errors and salt-and-pepper noise. This paper extends our recent work by introducing a new focal test approach with adaptive neighborhoods that avoids over-smoothing in wedge-shaped areas. We also conduct computational refinement on the FTSDT training algorithm by reusing focal values across candidate thresholds. Theoretical analysis shows that the refined training algorithm is correct and more scalable. Experiment results on real world data sets show that new FTSDT with adaptive neighborhoods improves classification accuracy, and that our computational refinement significantly reduces training time.

A short-term precipitation event near Duluth, Minnesota, USA caused flooding, erosion, and deposition that impacted the natural and anthropogenic landscape. This study quantified these impacts with an object-based image analysis approach that integrated multi-temporal lidar and optical data. Flooding inundated 3% of the study area and impacted 28% of the buildings. Topographic change volumes from erosion and deposition ranged from 36.3 m3to -32.2 m3. Erosion occurred over 21% of the area while deposition occurred over 11% of the area. These results provide decision managers with a spatially-explicit framework for assessing and monitoring inundation and topographic change in response short-term precipitation events.

Combining historical white and black aerial photo with more recent LiDAR and high resolution imagery, this research mapped wetland with high accuracy from 1930s to 2010s in object-based image analysis approach (OBIA). This research shows good potential in combining grey level information with OBIA method to map accurate historical wetland. We found that there were more small wetlands in 1938, but more large wetlands in 1978. In 2011, there were similar amount of small wetlands with 1938. Though 2011 had fewer large wetlands than 1978, 2011 year's individual large wetland had much larger area than those in 1978 and thus contributed to much larger total wetland area in 2011 than in 1978. We found significantly increasing precipitation and decreasing temperature over the time series, also drought in 1938, and this may explain wetland change well.

Given a raster spatial framework, as well as training and test sets, the spatial decision tree learning (SDTL) problem aims to minimize classification errors as well as salt-and-pepper noise. The SDTL problem is important due to many societal applications such as land cover classification in remote sensing. However, the SDTL problem is challenging due to the spatial autocorrelation of class labels, and the potentially exponential number of candidate trees. Related work is limited due to the use of local-test-based decision nodes, which can not adequately model spatial autocorrelation during test phase, leading to high salt-and-pepper noise. In contrast, we propose a focal-test-based spatial decision tree (FTSDT) model, where the tree traversal direction for a location is based on not only local but also focal (i.e., neighborhood) properties of the location. Experimental results on real world remote sensing datasets show that the proposed approach reduces salt-and-pepper noise and improves classification accuracy.

Given learning samples from a spatial raster dataset, the geographical classification problem aims to learn a decision tree classifier that minimizes classification errors as well as salt-n-pepper noise. The problem is important in many applications, such as land cover classification in remote sensing and lesion classification in medical diagnosis. However, the problem is challenging due to spatial autocorrelation. Existing decision tree learning algorithms, i.e. ID3, C4.5, CART, produce a lot of salt-n-pepper noise in classification results, due to their assumption that data items are drawn independently from identical distributions. In contrast, we propose a spatial decision tree learning algorithm, which incorporates spatial autocorrelation effect by a new spatial information gain (SIG) measure. The proposed approach is evaluated in a case study on a remote sensing dataset from Chanhassen, MN. Case study results show that the proposed approach outperforms the traditional approach in not only reducing salt-n-pepper noise but also improving classification accuracy.

Land-cover (LC) maps derived from remotely sensed data are often presented using a minimum mapping unit (MMU) to characterize a particular landscape theme of interest. The choice of an MMU that is appropriate for the projected use of a classification is an important consideration. The objective of this experiment was to determine the effect of MMU on a LC classification of the Neuse River Basin (NRB) in North Carolina. The results of this work indicate that MMU size had a significant effect on accuracy estimates only when the MMU was changed by relatively large amounts. Typically, an MMU is selected as close as possible to the original data resolution so as to reduce the loss of specificity introduced in the resampling process. Since only large MMU changes resulted in significant differences in the accuracy estimates, an analyst may have the flexibility to select from a range of MMUs that are appropriate for a given application.