As a spatial scientist with a background in cognitive psychology and AI, I have applied pattern recognition algorithms to classify landcover using remote sensing data at a variety of spatial scales. I witnessed the early AI in the eighties as a graduate student. Classical AI was better at building cognitive process models. I used classical AI techniques that were useful for building a cognitive model of human navigation. My model called the Navigator simulated landmark, route, and survey learning. Navigator’s landmark recognition is still relevant in today’s driverless car technologies that use qualitative reasoning to navigate without collisions.

AI includes machine learning where the system learn a spatial pattern such as landcover classes based on error- or match-based supervised learning. I have used these two dominant paradigms as well as semi-supervised architectures for landcover classification and change detection using remote sensing data and collaborated with leaders in the AI community Gail Carpenter and Steve Grossberg. We published a series of papers introducing ART (Adaptive Resonance Theory) networks (Carpenter et al., 1997, 1999). Several doctoral candidates investigated learning in backpropagation networks in remote sensing (Moody et al., Liu et al., 2004, Abuelgasim et al., 1999; Gopal et al., 1999; Snell et al., 2000). Funding in AI research was reduced substantially during its long winter that lasted over a decade that impacted my research funding. There is now a resurgence of interest in AI centered around deep learning, spurred by companies such as Google, Facebook, and Amazon. I have received some funding from Rockefeller foundation for knowledge graph representation through text mining (2012) and from Microsoft to investigate energy transition (2018). I have investigated IoT and other spatial applications (Gopal 2017; Gopal 2018).

My team is now applying deep learning for spatial panel analysis over a longer period. We want to train the system to recognize dynamic macrophytes or floating aquatic vegetation in Lake Kyoga based on Landsat data. The difficulty in this dataset of 400 Landsat images dating from 1986-2018 is to delineate the macrophytes correctly. (Ma, Gopal, Kaufman, AGU presentation 2019). The second problem is to identify five different types of macrophytes using spectral signatures. We are asking the following research questions. How general is the result? How robust is the system in learning? Could it identify a macrophyte that it has not seen before? Can it recognize a pattern in a different scene if it is not part of its training data? How many labeled samples (training data) do we need to do better than experts in the labeling task? Is our labeled data biased, and how does bias manifest in resulting identification? There is much work to be done in applying deep learning to traditional spatial classification problems.